Data Curation Tool Modern

v5.8.48 Update

• Adds app-wide active assistant overlays so live action notes are visible from any tab while a top-level assistant, Tag Editor assistant, Graph assistant, or Code assistant is running.
• Adds a separate live chain-of-thought-style reasoning trace overlay that is enabled by default alongside action notes. This is the model/app-visible reasoning trace; provider/private hidden reasoning is not extracted.
• Adds default settings and per-surface controls for the live reasoning overlay, plus Settings persistence for assistant_show_live_chain_of_thought and assistant_show_live_reasoning_trace.
• Extends assistant planning responses with visible_reasoning_trace / visible_chain_of_thought metadata so completed turns preserve the visible trace next to the visible plan/action notes.
• Keeps automatic token/context condensation enabled by default while showing live context-budget pressure in the global overlays and conversation panel.

See docs/V5_8_48_GLOBAL_ASSISTANT_ACTION_AND_REASONING_OVERLAYS.md and docs/wiki/78-Global-Assistant-Action-and-Reasoning-Overlays.md.

• Adds system-RAM guardrails for long-running LLM/VLM sessions and disables automatic CPU offload by default so a sharded/offloaded model cannot silently consume all system memory unless explicitly requested.
• Bounds stored chat context/response payloads and loads conversation JSON with SQL-level truncation guards to reduce overnight assistant memory growth.
• Expands the condensed conversation memory section with copy/download controls so long summaries are readable.
• Wires Tag Editor assistant tag pruning to active Dataset Pipeline / LoRA rules and applies assistant JSON keep/remove/add/final_tags directives to the real media tag list.
• Adds approved assistant/orchestrator model queue tools for model load, inference, unload, and wait-for-job workflows, surfaced through the same visible queue system as manual Quick Tag operations.

See docs/V5_8_47_MEMORY_GUARD_ASSISTANT_MODEL_TOOLS_TAG_PRUNING.md and docs/wiki/77-Memory-Guard-Assistant-Model-Tools-Tag-Pruning.md.

v5.8.46 Update

• Fixes the remaining Quick Tag multi-model unload path so Queue Unload Selected captures every highlighted model and sends every selected unload request through a shared batch endpoint.
• Adds live token/context-budget progress while the Tag Editor Assistant is generating: the circular context meter updates during the request instead of appearing only after the response returns.
• Keeps visible planning/action-note trace controls open and on by default for assistant conversations. This is user-facing visible reasoning/action trace, not provider/private hidden chain-of-thought.
• Confirms and hardens automatic context condensation before and after assistant chat requests so near-limit conversations are summarized into compact memory and continued from the condensed state.
• Adds regression coverage for multi-unload batching, live context-budget ticking, visible trace defaults, and auto-condense metadata.

See docs/V5_8_46_MULTI_UNLOAD_LIVE_TOKEN_CONTEXT.md and docs/wiki/76-Multi-Unload-Live-Token-Context.md.

v5.8.45 Update

• Improves attention heatmap overlays in Tag Editor, Compare, and Attention Visualizer with Hydra-demo-style signed CAM rendering: green indicates positive tag evidence and red indicates negative evidence.
• Adds optional native Hydra CAM extraction when a compatible local RedRocket Hydra 3.5 model/repo is available; otherwise the attention service uses a deterministic signed fallback heatmap instead of the old single-color blob overlay.
• Adds CAM depth and heat-strength controls to inline heatmap cards while keeping clustering/projection methods in the dedicated Attention Visualizer tab.
• Hardens attention overlay dropdowns against background refresh so native dropdowns stay open while scrolling/selecting.
• Reduces Quick Tag queue completed-row linger time and keeps the queue panel patching live so completed model rows clear without switching tabs.

See docs/V5_8_45_ATTENTION_HEATMAP_DROPDOWN_QUEUE_REFRESH.md and docs/wiki/75-Attention-Heatmap-Dropdown-Queue-Refresh.md.

v5.8.44 Update

• Keeps every selected Quick Tag model visible in the model queue while download, load, unload, and inference jobs are being submitted or running.
• Clarifies Submit queue requests in parallel semantics: disabled means requests are submitted one by one, but all selected rows still appear immediately as queued placeholders.
• Preserves client-side placeholder queue rows during live /api/jobs polling so slow backend responses cannot make multi-model Quick Tag queues collapse to a single row.
• Restores hover prediction bars by patching per-media/model score data directly from completed inference job results before the slower full media refresh.
• Reattaches recoverable model probabilities to normalized bare-string tags so each queued model applies only tags at or above its own threshold.
• Extends the Nightshade/Glaze integrity classifier workflow to video by sampling frames with user-selectable FPS, max frames, PNG/JPEG/WebP output, compression/quality, and highest-quality presets.
• Speeds Gallery thumbnail generation with a non-blocking thumbnail path plus OpenCV CPU/CUDA resize when available.

See docs/V5_8_44_QUICK_QUEUE_LIVE_ROWS_VIDEO_INTEGRITY_SCORES.md and docs/wiki/74-Quick-Queue-Live-Rows-Video-Integrity-Scores.md.

v5.8.41 Update

Improves Quick Tag queue selection and live model-state updates. The multi-model selector in the Tag Editor now explicitly supports Ctrl/Cmd-click toggling, Shift-click range selection, and Ctrl/Cmd+A select-all while preserving selected options during background status refreshes.

The Quick Tag queue panel now shows download, load, unload, and inference jobs for the selected quick-tag-capable models. Model option rows are patched in place as each individual job changes state, so the dropdown no longer waits for the entire queue to finish before showing that a specific model has downloaded, loaded, unloaded, failed, or started inference.

Improves Agentic Graph Editor node selection. Canvas nodes now support Ctrl/Cmd-click toggling and Shift-click range selection in addition to the existing Ctrl/Cmd drag selection box, copy, cut, paste, delete, and grouped dragging behavior.

See docs/V5_8_41_QUICK_TAG_MULTISELECT_LIVE_QUEUE_GRAPH_SELECTION.md and docs/wiki/71-Quick-Tag-Multiselect-Live-Queue-Graph-Selection.md.

v5.8.39 Update

Fixes the remaining live Models-tab refresh issue. Loaded model cards now patch their lifecycle circles, loaded/downloaded state, highlighted background, and active-model ordering in place as soon as the load/unload lifecycle reaches a terminal state; the update path preserves the Models tab scroll position and does not require a tab switch.

Improves GPU/RAM resource visibility. The resource panel now shows actual driver/torch-used VRAM, actual free VRAM, torch allocated/reserved memory, app reservation budget, loaded model count, and system RAM in the live poller so the user and assistant/orchestrator can make placement decisions from current memory state rather than stale catalog flags.

Hardens Gallery/media refresh behavior after model failures. Media loading now preserves the existing Gallery state and shows a non-blocking warning if a refresh fails, while background polling is isolated so failed model jobs cannot break Gallery rendering or general UI performance.

Adds a model runtime planning context endpoint for assistant/orchestrator workflows. /api/models/runtime-planning-context exposes live resources, compact model metadata, strict-GPU placement policy, and sharding guidance so LLM/VLM/GLM/supervisor models can propose GPU IDs, sharding strategy, tensor-parallel settings, and queueable model jobs using the same resource data shown to the user.

See docs/V5_8_39_LIVE_MODEL_CARD_VRAM_GALLERY_RESILIENCE.md and docs/wiki/69-Live-Model-Card-VRAM-Gallery-Resilience.md.

v5.8.38 Update

Adds live model-runtime polling for model lifecycle, GPU VRAM reservations, driver-reported CUDA memory, and regular system RAM. Models, Tag Editor, Agent Tools, Jobs, and assistant/orchestration surfaces now patch status circles, model dropdown styling, and resource panels in place without requiring a tab switch and without forcing a full page render that would reset scroll position.

Adds a queued Quick Tag inference workflow in the Tag Editor. The quick-tag card now supports a multi-select model queue, parallel enqueueing, active per-model progress rows, per-job ETAs, and an overall circular queue progress indicator. Completed jobs disappear from the active quick queue while remaining visible in the Jobs tab.

Adds a global model-job queue panel to Agent Tools so assistant/orchestrator-driven model handoffs can be monitored separately from the quick-tag subset. The new generic /api/models/queue-runs backend endpoint queues multiple model inference jobs through the existing model-inference lane and exposes the same live progress/ETA UI used by Quick Tag.

See docs/V5_8_38_LIVE_MODEL_RUNTIME_QUEUES.md and docs/wiki/68-Live-Model-Runtime-Queues.md.

v5.8.37 Update

Fixes the Tag Editor Quick Tag / Rating Model feedback loop. Selecting a model in the dropdown now immediately swaps the lifecycle/progress circles to that exact model and refreshes status in place, so the user no longer has to switch tabs to see whether the selected model is loaded, downloading, loading, or ready.

Fixes the post-inference refresh path for quick tag/rating runs. Completed model inference jobs now refresh the affected media rows, clear stale tag drafts, request prediction scores, and hard-refresh the active media review tab while preserving scroll position. The completed job payload now also includes full candidate/applied tag maps and candidate score rows so the frontend can optimistically patch the current image if the media fetch is slow.

Tightens model-tag normalization for applied tags. Model-emitted tags/classes now continue through the selected tag profile with alias resolution and implication expansion enabled by default, then save with the selected tag text mode and ordering strategy. This keeps PixAI/WD/Thouph/Hydra/JTP outputs aligned with the user's active e621/e926/Danbooru/etc. dictionary preset instead of silently creating disconnected new tags when a known alias/implication path exists.

See docs/V5_8_37_QUICK_TAG_REFRESH_AND_ALIAS_SCORE_SYNC.md and docs/wiki/67-Quick-Tag-Refresh-and-Alias-Score-Sync.md.

v5.8.33 Update

Fixes model-download progress so Hugging Face snapshot downloads, small ONNX tagger repos, and local payload finalization now update the model download lifecycle circle and the Models tab without requiring a tab switch. The loader also reconciles a completed local payload before blocking a load on an apparently active download, so a model that finished downloading should not remain stuck behind a stale running circle.

Adds three known-good Agentic Graph templates plus a new Agentic Workflow READMEs tab with rendered instructions, create/run buttons, expected results, and manual next steps. The baseline templates are intentionally local-only and dry-run-safe so the user can verify the graph runtime before expanding into model calls, media probing, downloader steps, MCP tools, shell commands, or trainer handoffs.

See docs/V5_8_33_DOWNLOAD_PROGRESS_AND_GUARANTEED_GRAPHS.md and docs/wiki/63-Download-Progress-and-Guaranteed-Agentic-Graphs.md.

v5.8.32 Update

Fixes migrated model detection so model folders copied or moved from an older install are loaded from local disk instead of falling back to a remote repo id and redownloading. The Models tab now has a Rescan / Reconcile Migrated Models button, load/chat paths use local-files-only resolution whenever a downloaded local payload is detected, and local model loads temporarily force Hugging Face/Transformers offline mode so helper libraries cannot silently fetch missing snapshots.

See docs/V5_8_32_MIGRATED_MODEL_DETECTION_NO_REDOWNLOAD.md and docs/wiki/62-Migrated-Model-Detection-No-Redownload.md.

v5.8.31 Update

Adds graph chunk-selection and grouped dragging with Ctrl+drag, graph-node copy/cut/paste shortcuts, searchable/filterable node palette menus with recent-node history, multi-input/multi-output port rendering, cursor-accurate right-click node menus, Attention Visualizer tag suggestions, an Augment-tab render fix, a closed-loop model-training improvement graph template, and isolated ONNX adapters for the PixAI and WD v3 tagger rows that previously failed with “Model adapter is not available.”

See docs/V5_8_31_GRAPH_SELECTION_MENU_WD_PIXAI_FIXES.md and docs/wiki/61-Graph-Selection-Menu-WD-PixAI-Fixes.md.

v5.8.30 Update

Fixes the Agentic Graph Editor refresh/input regression by separating graph clicks, right-clicks, node drags, port gestures, and graph zoom from actual page scrolling. Graph-local canvas and node-inspector updates now repaint immediately while preserving the tab scroll position. Also fixes the Agentic Graph Chat render failure in inline selected-model runtime controls.

See docs/V5_8_30_GRAPH_IMMEDIATE_INTERACTION_AND_CHAT_FIX.md and docs/wiki/60-Graph-Immediate-Interaction-and-Chat-Fix.md.

v5.8.29 Update

Moves heatmap-style attention review into the Tag Editor and Compare tabs as toggleable semi-transparent overlays, sets the default classifier/tagger threshold to 0.70, fixes several Agentic Graph Editor canvas/menu/inspector interactions, and adds a graph-linked chat tab with visible plan/action-trace support.

See docs/V5_8_29_ATTENTION_OVERLAY_GRAPH_CHAT_THRESHOLD.md and docs/wiki/59-Attention-Overlay-Graph-Chat-Threshold.md.

v5.8.28 Update

Adds model-specific Thouph preprocessing, soft automatic refreshes that preserve Models-tab scroll/dropdowns, Tag Editor prediction sort modes, Agentic Graph Editor preset-render fixes, audio/voice/video+audio dataset scope additions, and LingBot-Video world-model catalog/runtime rows.

See docs/V5_8_28_PREPROCESS_SCROLL_GRAPH_WORLD_MODELS.md and docs/wiki/58-Preprocessing-Scroll-Graph-LingBot-World-Models.md.

v5.8.27 Update

Adds a new Multimodal Dataset Builder tab for image/video/audio dataset preparation, structured captions, LTX-2.3 exports, Wan 2.2 Musubi/DiffSynth/SimpleTuner/AI Toolkit exports, training MCP handoffs, explicit Tripo P1.0/Rodin/Hunyuan3D 3.1 3D provider rows, and Tag Editor model-score highlighting by one/multiple/all model prediction rows.

See docs/V5_8_27_MULTIMODAL_DATASET_BUILDER.md and docs/wiki/57-Multimodal-Dataset-Builder-LTX-Wan-MCP.md.

v5.8.25 Update

Fixes the remaining legacy EVA attn_pool inference compatibility error and corrects the frontend refresh/defer behavior that made Models, Gallery, Jobs, and log/error views look stale until switching tabs.

See docs/V5_8_25_EVA_ATTN_POOL_AND_LIVE_TAB_REFRESH_FIXES.md and docs/wiki/55-EVA-Attention-Pool-and-Live-Tab-Refresh-Fixes.md.

v5.8.23 Update

Migration tag-database import no longer stalls on stale tag_export_files metadata from older installs. Export-file metadata is rebuilt from migrated local cache files, the redundant legacy dictionary mirror is skipped when normalized dictionaries are present, and tag-table import now commits before progress callbacks so the Dashboard/Jobs progress indicator can update live instead of appearing stuck.

See docs/V5_8_23_MIGRATION_TAG_DB_STALL_FIX.md and docs/wiki/53-Migration-Tag-Database-Stall-Fix.md.

Repository: https://github.com/x-CK-x/Dataset-Curation-Tool

Data Curation Tool Modern is a local-first dataset curation, tagging, model-management, downloader, 3D generation, media metadata, and tool-orchestration application. The current documentation set is stored in docs/wiki/ and is designed to be pushed directly into the repository or copied into the GitHub Wiki.

v5.8.22 Update

• Install Migration now treats complete model folders as atomic model groups and can move them as one fast same-drive directory operation instead of thousands of individual file transfers.
• Existing complete target model groups are detected and skipped/deleted from the source when duplicate-source deletion is enabled.
• Runtime duplicate checks prevent rewriting large model shards that are already present in the target install.
• Migration progress no longer compresses the file-transfer stage into only 75% of the job, so long migrations should not appear stuck around 70–71%.

See docs/V5_8_22_MIGRATION_FAST_MODEL_GROUPS_AND_PROGRESS_FIX.md and docs/wiki/52-Migration-Fast-Model-Groups-and-Progress-Fix.md.

v5.8.21 Update

• Install Migration now uses fast same-drive moves in Move mode, enabled by default.
• Large model files on the same SSD/volume are renamed into the new install instead of copied byte-for-byte and then deleted.
• Cross-drive moves automatically fall back to the prior chunked copy-and-delete path.

See docs/V5_8_21_FAST_MODEL_MOVE_MIGRATION.md and docs/wiki/51-Fast-Same-Drive-Model-Migration.md.

New in v5.8.19 — migration local-cache reuse and weekly startup sync

• Migration now defaults to local-only cache reuse, so older-install tag exports/database rows are used instead of forcing a post-migration internet refresh.
• Startup tag-dictionary network checks are now gated to once per week by default.
• Cached runtime/tag_exports/<profile>/ files can be imported directly into the current dictionary without downloading fresh copies.
• Settings and Install Migration now expose controls for local-only migration and weekly tag-sync intervals.

New in v5.8.18 — EVA norm-pre and graph editor event-handler fixes

• Fixed the remaining legacy EVA/timm inference failure where older pickled EVA models could miss norm_pre.
• Added identity fallbacks for newer optional EVA fields used by current timm forward paths.
• Restored missing Agentic Graph Editor frontend handlers for node update, node-kind change, and config JSON editing.
• Graph canvas right-click, wheel zoom, pointer-drag pan, node creation, node movement, and node-port connection now render immediately without being blocked by page-scroll debounce protection.

New in v5.8.15 — dashboard migration progress refresh

• Added a Refresh Dashboard Now control so the Dashboard can explicitly reload startup/migration maintenance state.
• Manual migration now publishes an immediate startup-maintenance state when the migration job is queued.
• Migration now emits live progress during previous-install scanning, large file copy/move operations, and migrated tag database imports.
• Startup status requests are no-cache/cache-busted so the Dashboard does not display stale migration progress.

New in v5.8.14 — legacy tagger inference and scroll stability fixes

• Fixed legacy EfficientNetV2-M inference so every image is converted to the static 1 x 3 x 448 x 448 tensor expected by its ONNX/PyTorch classifiers.
• Fixed legacy EVA/timm pickle inference by patching nullable EVA attributes such as reg_token and mask_token when older serialized model objects do not contain them.
• Reduced scrollbar jitter by deferring polling-driven renders while the user is actively scrolling and by replacing long delayed scroll-restore loops with short-lived tokenized restore passes.

New in v5.8.13 — startup migration progress, attention heatmaps, and functional graph runtime

• Dashboard Startup Maintenance now mirrors manual migration jobs immediately, including queued/running progress, elapsed time, ETA, job id/type, and post-migration tag dictionary reconciliation.
• The tab scroll restoration path was tightened again so switching tabs preserves the main scroll container instead of restoring the previous shell scroll position over the target tab.
• Attention Visualizer now exposes immediate Grad-CAM/CAM, Hydra CAM/PCA, diffusion U-Net cross-attention, and t-SNE/embedding visualization artifacts with local overlay preview links.
• Agentic Graph Editor now has local graph presets, export snapshots, edge metadata editing, graph-runtime session execution, selected-node execution, runtime result inspection, executable bundle-limit policies, supervisor fanout previews, model-call preview packets, and approval-gated browser/MCP/tool call previews.

New in v5.8.12 — model prediction normalization and unload UI fixes

• Model-generated tags now pass through one canonical postprocessor before being stored, scored, shown, or applied.
• Active tag text mode is respected, so underscore-emitting models no longer create duplicate “new” tags when the tool is configured for spaces.
• Alias resolution and implication expansion now preserve prediction scores.
• Tag-hover score panels show per-model colored rows and an average score when multiple model predictions exist for the same tag.
• Model unload actions now surface immediately on the load/unload lifecycle circle across tabs.

New in v5.8.11 — Graph editor port, browser MCPs, scroll persistence, and startup resume

• Expanded the integrated Agentic Graph Editor with the standalone graph-editor node concepts while keeping the existing dark/neon canvas style.
• Added browser MCP entries for user-approved visible browsing/search handoff through default browser, Edge, Chrome, Firefox, Chromium, and Tor Browser.
• Fixed tab switching so large tabs preserve scroll position instead of jumping back to the top.
• Install Migration now mirrors progress into the Dashboard startup-maintenance card and resumes post-migration cache/dictionary reconciliation.
• Agent tool smoke tests use cached results on future launches, making repeated startup maintenance faster.

New in v5.8.9 — Startup progress crash fix

• Fixed the startup initialization job failure caused by the missing time import used by the live startup progress/ETA tracker.
• Startup tag-dictionary sync now uses the package __version__ for its default user-agent string instead of a hardcoded previous release value.

New in v5.8.8 — Hydra UTF-8 inference + startup progress fixes

• Fixed the next local RedRocket Hydra 3.5 inference failure on Windows where Hydra wrote Unicode tag labels to a CP1252 stdout stream and crashed with UnicodeEncodeError.
• Hydra local inference now writes repo-native CSV output to a UTF-8 temporary file instead of stdout, while also forcing UTF-8 child-process environment variables.
• Added a live Dashboard startup-maintenance indicator with a white circular progress ring, elapsed time, ETA, current phase, and recent startup steps.
• Long startup maintenance now reports progress through /api/system/startup-status instead of leaving the user without feedback.
• Kept the v5.8.7 workflow preset, scroll preservation, Gallery refresh, and page-limit fixes.
• Bumped the release version to 5.8.8.

See docs/V5_8_8_HYDRA_UTF8_STARTUP_PROGRESS_GALLERY_FIXES.md and docs/wiki/39-Hydra-UTF8-Startup-Progress-and-Gallery-Fixes.md.

New in v5.8.7 — Hydra loader compatibility + workflow/gallery UI fixes

• Fixed another local RedRocket Hydra 3.5 inference failure where upstream inference.py called Loader.heuristic_max_workers(...) while the downloaded utils/loader.py only exposed heuristic_workers(...).
• Extended the Hydra local-repo compatibility patch so it also accepts Loader(..., max_workers=...) in older loader snapshots.
• Kept the unavailable Z3D/Zack3D model removed from the local tagger catalog.
• Expanded Automation Workflow templates and goal dropdown coverage for style, concept, character, and character+style / OC+style prep.
• Fixed Gallery/page navigation refresh behavior so explicit reload/page buttons force a render instead of being deferred by dense-tab interaction protection.
• Added backend and frontend page clamping so the Gallery cannot browse past its last available page.
• Bumped the release version to 5.8.7.

See docs/V5_8_7_HYDRA_LOADER_WORKFLOW_GALLERY_FIXES.md and docs/wiki/38-Hydra-Loader-Workflow-and-Gallery-Fixes.md.

New in v5.8.6 — Hydra Python 3.11 source patch + unavailable Z3D removal

• Removed the unavailable Z3D/Zack3D legacy tagger entry from the Models catalog and legacy config registry.
• Added an automatic Hydra local-repo compatibility patch for Python versions where multiprocessing.Queue cannot be used as Queue[str] in runtime annotations.
• The Hydra adapter now patches the downloaded utils/loader.py before load/inference and writes a small patch marker so the same repo is not repeatedly modified.
• Bumped the release version to 5.8.6.

See docs/V5_8_6_HYDRA_PY311_QUEUE_PATCH_AND_Z3D_REMOVAL.md and docs/wiki/37-Hydra-Python-Queue-Patch-and-Z3D-Removal.md.

New in v5.8.5 — Hydra Windows libvips DLL loader fix

• Fixed a second local RedRocket Hydra 3.5 Windows runtime failure where pyvips was present but libvips-42.dll could not be loaded.
• Kept Windows DLL-directory handles alive for the process instead of adding and immediately losing them.
• Added broader Conda/PATH/manual-libvips probing and clearer diagnostics for missing libvips files.
• Added model-load auto-repair for the pyvips/libvips chain using pyvips[binary]>=3.0.0, pyvips-binary>=8.16.0, and Conda fallback repair when Conda is visible.
• Added scripts/repair_hydra_runtime_dependencies.py and updated install.bat, update.bat, install.sh, update.sh, and Hydra repair scripts to use it.
• Bumped the release version to 5.8.5.

See docs/V5_8_5_HYDRA_WINDOWS_LIBVIPS_DLL_FIX.md and docs/wiki/36-Hydra-Windows-libvips-DLL-Loader-Fix.md.

New in v5.8.4 — Hydra runtime dependency fix

• Fixed local RedRocket Hydra 3.5 inference dependency handling so missing pyvips/libvips is detected before the native Hydra subprocess runs.
• Added pyvips and libvips to the Conda environment so fresh installs and updates include Hydra's image-loading runtime.
• Added install_hydra_runtime_deps.bat and install_hydra_runtime_deps.sh for repairing existing environments without reinstalling the whole tool.
• Added scripts/check_hydra_runtime_dependencies.py to verify local Hydra inference prerequisites.
• Bumped the release version to 5.8.4.

See docs/V5_8_4_HYDRA_RUNTIME_DEPENDENCY_FIX.md and docs/wiki/35-Hydra-Runtime-Dependency-Fix.md.

New in v5.8.3 — Agentic Graph Editor

• Added an Agentic Graph Editor tab for visual node/edge orchestration workflows.
• Graphs can be created from workflow templates, edited manually, drafted/refined by the selected assistant/orchestrator model, validated, dry-run, converted to Automation Workflows, saved as reusable workflows, or queued through Jobs.
• Graph execution reuses the Automation Workflow backend so approval gates, branch-safe global dataset behavior, and run manifests stay consistent.
• Added persistent graph storage under runtime/agentic_graphs/, API routes under /api/graph-editor/*, and compatibility aliases under /api/graphs/* for older graph-editor style integrations.
• Bumped the release version to 5.8.3.

See docs/V5_8_3_AGENTIC_GRAPH_EDITOR.md and docs/wiki/34-Agentic-Graph-Editor.md.

New in v5.8.1 — Character Reference and LoRA Augmentation Automation

• Added a Character Reference tab for zero/one/few-shot character search, ranking, and pruning without training a new model.
• Added reusable character profiles with positive and negative reference memory, branch-aware pruning plans, and optional DINOv2/CLIP/SigLIP/OWLv2/Grounding-DINO/SAM/SAM2 backend contracts.
• Added LoRA augmentation and regularization automation in Dataset Pipeline / Pipeline Prep for character, character+style, style, concept, IC-LoRA, ControlNet, and embedding/textual-inversion datasets.
• Branch-local variants can be generated for headshot/detail crops, upper-body crops, style/texture crops, concept/context crops, square/bucket-safe copies, light denoise, and conservative upscaling while keeping global originals read-only.
• Bumped the release version to 5.8.1.

See docs/V5_8_1_CHARACTER_REFERENCE_LORA_AUGMENTATION.md.

New in v5.8.0 — Legacy local taggers and tag/caption translation

• Added the legacy/local image taggers from the original model configuration to the Models catalog: Thouph EVA02-CLIP 7704, Thouph EVA02 ViT-Large 448 8046, and Thouph Experimental EfficientNetV2-M 8035.
• Added a shared legacy tagger adapter that preserves model-specific preprocessing, thresholds, ONNX/PyTorch runtime selection, tag metadata ordering, placeholder output handling, and e621 alias/implication cleanup.
• Added a profile-aware tag/caption translation tool in Tag Dictionaries for cross-booru tag conversion and caption-format conversion, with optional local/cloud model assistance for uncertain mappings.
• Set the application release version to 5.8.0.

See docs/V5_8_0_LEGACY_TAGGERS_TAG_TRANSLATION.md.

New in v5.78.13 — Dataset Pipeline / Training Prep / 3D-Print Handoff

• Added a Dataset Pipeline tab for repeatable pre-training dataset preparation on top of the Global Dataset branch layer.
• Added rule presets and prompt packets for LoRA, IC-LoRA, ControlNet, embedding/textual-inversion, style, character, character+style, and concept datasets.
• Added deterministic dry-run/apply rule cleanup for branch-local tag/caption sidecars while keeping global originals untouched.
• Added external trainer handoff rows/MCP interfaces for Kohya SS, OneTrainer, Diffusers scripts, LTX Trainer, ComfyUI training nodes, cloud training providers, and future source-manifest/webscraper bridges.
• Added 3D-print/slicer handoff support for Blender, ZBrush, MeshLab, PrusaSlicer, OrcaSlicer, CuraEngine/Cura, Bambu Studio, and Slic3r.

See docs/V5_78_13_DATASET_PIPELINE_TRAINING_PREP_3D_PRINT.md.

New in v5.78.12 — Global Dataset + 3D/ZBrush Expansion

• Added a Global Dataset tab that stores original media once by SHA-256 and tracks source/post mappings to reduce duplicate downloads.
• Added branch/model-dataset configs that point at global originals while keeping editable tag/caption sidecar copies separate from the original data layer.
• Added augmented-media variant tracking so branch-generated variants remain linked to their source original.
• Added Dream Textures, QuickMaker, Meshy official API, Blender MCP add-on/server, and ZBrush Python/MCP refinement rows to the 3D/MCP catalog.
• Added /api/global-dataset/* endpoints for status, asset search, manual ingest, branch linking, and variant registration.

See docs/V5_78_12_GLOBAL_DATASET_3D_MCP_ZBRUSH.md.

Current highlights

• Modern dataset import, gallery review, tag editing, comparison, metadata extraction, and media tooling.
• Model catalog with local/cloud model settings, lifecycle progress indicators, GPU placement, loading, inference, and training status surfaces.
• 3D generation catalog covering text-to-3D, image-to-3D, multi-image-to-3D, and video-to-3D provider workflows.
• MCP tool integration surfaces for Blender, Krita, Audacity, OBS Studio, and ComfyUI.
• Downloader updates with e621/booru Boolean logic gates, presets, queues, deduplication, and tag dictionaries.
• GitHub-ready visual documentation assets under docs/wiki/assets/images/.
• Automation Workflows tab for user/model/co-edited curation pipelines, dry-runs, approval gates, branch-safe execution, and run manifests.
• Agentic Graph Editor tab for visual user/model/co-edited orchestration graphs that compile into Automation Workflows.

Visual documentation map

All visuals are stored in docs/wiki/assets/images/ and are referenced with repo-local paths so they work after the files are pushed to GitHub.

v5.8.2 automation workflows

• Added an Automation Workflows tab for template workflows, assistant/model workflow drafting, direct workflow JSON editing, validation, dry-run, queued execution, and run manifests.
• Workflows can wrap global dataset branches, dataset pipeline rules, character-reference pruning, augmentation planning, regularization planning, export, and trainer/tool handoff.
• Unsafe or expensive workflow steps keep approval gates by default.

See docs/V5_8_2_AUTOMATION_WORKFLOWS.md and docs/wiki/33-Automation-Workflows-and-Cooperative-Curation.md.

Historical patch notes

v5.45 configurable Assistant / Orchestrator model

• Added Assistant-tab controls to choose which downloaded/local/API LLM or VLM acts as the app assistant.
• Added persistent assistant_model_name and orchestrator_model_name settings.
• Added GET/PUT /api/models/assistant-config for saving and reading Assistant/Orchestrator defaults.
• Added load/unload buttons in the Assistant tab so the selected assistant model uses the same lifecycle circles, GPU placement controls, and RAM/VRAM unload path as other models.
• Orchestration templates now use the configured orchestrator model instead of hardwiring the built-in dataset-assistant.
• Explicit dataset-assistant chat remains available as the no-model fallback; orchestrator placeholders resolve to the configured orchestrator default.

See docs/V5_45_CONFIGURABLE_ASSISTANT_ORCHESTRATOR.md.

v5.44 stoppable downloads and self-activating Conda scripts

• Added /api/jobs/cancel plus Jobs-tab buttons to stop queued/running download jobs. Queued downloads are cancelled immediately; running downloads stop cooperatively.
• Model downloads run in a dedicated model_download worker lane so large downloads do not consume the whole general job pool.
• run, install, and update scripts now locate a user-level Conda install and activate data-curation-tool automatically.

See docs/V5_44_STOPPABLE_DOWNLOADS_CONDA_SCRIPTS.md.

v5.43 download queue, asset rescan, and VLM defaults

• Added Hugging Face model-download progress heartbeats.
• Refreshing model status rescans local model folders and marks migrated models as downloaded/completed.
• Tag dictionary status reconciles migrated runtime/tag_exports files and reports cached export row counts.
• VLM/LLM Tag Editor selection defaults to validating existing image tags and can consider all categories.
• Hugging Face text/VLM load failures now surface stronger Gemma-style diagnostics.

See docs/V5_43_DOWNLOAD_QUEUE_RESCAN_VLM_DEFAULTS.md.

v5.42 GPU controls, dropdown stability, and manual/VLM tag selection

This build fixes global dropdown auto-closing during polling, extends model GPU residency controls, adds manual Tag Editor highlight/deselect controls, and routes chat/VLM models such as Gemma 4 E4B IT through real image-aware tag-selection inference instead of the old heuristic-only path.

See docs/V5_42_GPU_TAG_SELECTION_DROPDOWN_FIXES.md.

v5.41 GPU placement, residency, and unload controls

• Added per-model GPU placement controls so you can choose exact CUDA GPU ids before loading a model.
• Added sharding, dtype, quantization, runtime-engine, tensor-parallel, and per-GPU max-memory fields to the Models tab load panel.
• Added a GPU residency/status panel showing detected GPUs, usable VRAM, app-reserved VRAM, driver-free VRAM, loaded models, and models currently loading.
• Load requests now run a placement/VRAM plan first and block overcommit before the adapter starts allocating memory.
• Added queued unload jobs with lifecycle progress so unloading from RAM/VRAM is reflected in the same status circles.
• Existing loaded models reject incompatible placement changes until you unload and reload with the new GPU selection.

See docs/V5_41_GPU_PLACEMENT_UNLOAD_CONTROLS.md.

v5.40 dropdown stability and previous-install asset migration

• Prevented automatic polling rerenders from closing model/category/profile dropdown menus while the user is trying to select an option.
• Added an Install Migration tab to scan one or more previous app installs and move/copy reusable assets into the current install.
• Migrates model folders, cached tag DB-export files, imported tag dictionary rows, custom tags, and optionally presets/download cache/outputs.
• Processes newest installs first, then older installs for unique assets, so the most recent setup wins while older one-off model downloads can still be recovered.
• Added startup migration settings and environment overrides so prior assets can be imported before startup tag DB sync attempts another network download.

See docs/V5_40_DROPDOWN_MIGRATION_FIXES.md.

v5.39 model tag refresh, duplicate-load, and typing-focus fixes

• Prevented duplicate explicit load jobs when a model is already loaded in memory; the API now returns a completed no-op instead of queueing another load.
• Kept the load lifecycle circle tied to the original load job when inference uses an already-loaded model, so running a model no longer looks like a second reload.
• Refreshed active media rows, tag drafts, and prediction-score caches when completed model inference applies tags/captions, so Tag Editor and Compare update without a full browser reload.
• Refreshed synchronous LLM/VLM/Assistant tag-selection apply operations immediately, including the active image in Tag Editor.
• Preserved text input/textarea focus and caret position during automatic polling renders; active typing now defers nonessential rerenders instead of bumping the cursor out of the prompt field.

See docs/V5_39_MODEL_TAG_REFRESH_FOCUS_FIXES.md.

v5.38 status, gallery, scroll, and downloader stability fixes

• Fixed model load/run visibility after queueing: Jobs and lifecycle circles refresh immediately without a browser reload.
• Preserved selected model dropdowns across status polling so buttons act on the model the user selected.
• Stopped Gallery model-score request recursion by caching score requests per media/tag signature.
• Preserved global and nested tab scroll positions across automatic renders.
• Hardened downloader parallel preset dedupe using source post/file identity keys, not just URL strings.
• Changed downloader defaults: Download All Posts enabled, API/page delay 7s, file delay 7s, timeout 60s, retries 3, backoff 2s.

See docs/V5_38_STATUS_GALLERY_DOWNLOAD_FIXES.md.

v5.37 model lifecycle status circles

• Added a shared model lifecycle status system for four long-running phases: download, load into memory, inference, and training.
• Added circular progress indicators in the Models tab, per-model cards, quick tag/rating cards, tag-selection cards, and annotation model panels.
• Added /api/models/status and /api/models/load so model downloads, explicit memory loads, lazy loads before inference/chat/tag-selection, annotation model downloads/loads, and training scaffolds all report consistent status.
• Running inference/chat/tag selection now refuses to start while the same model is still downloading or loading, preventing use before a compatible adapter has finished initializing.
• The Jobs table remains the durable log; the new lifecycle status layer is a lightweight live UI surface and does not remove the existing download/run/unload behavior.

See docs/V5_37_MODEL_LIFECYCLE_STATUS.md.

v5.31 FlexAvatar complete 3D head avatar integration

• Added a dedicated FlexAvatar tab for single-image, few-shot, and monocular-video complete 3D Gaussian head avatars.
• Added an isolated dct-flexavatar Conda runtime so the upstream Python 3.9 / CUDA 11.8 stack does not overwrite the main tool environment.
• Added official FLEX-1 checkpoint installation, pretracked example seeding, Pixel3DMM custom-input tracking, latent fitting, custom/bundled drivers, 360° rendering, avatar-code reuse, and identity interpolation.
• Added output previews, the official interactive viewer launcher, workspace asset management, and queued Jobs integration.
• Added mixed-supervision research bundles with 2D/3D bias-sink labels and paper-baseline settings, plus an honest external-trainer validation/launch path. The supplied upstream release does not contain a complete official base-model training entrypoint.
• Preserved the upstream CC BY-NC 4.0 component as a separately executed optional integration; its model checkpoint is not redistributed.

See docs/V5_31_FLEXAVATAR_INTEGRATION.md. Run python scripts/verify_v531_flexavatar.py for an offline integration check that does not require the large checkpoint or CUDA inference.

v5.30 editable 2D/3D pose, 3D generation/rigging, Blender, and cumulative SAM prompting

• Pose bones are now visibly rendered and remain connected while joints are dragged on the image or in the interactive 3D viewer.
• Pose inference includes YOLO, MediaPipe Pose Landmarker, MMPose RTMPose/ViTPose/WholeBody/Animal, MotionBERT human 3D, InterNet hand 3D, and custom MMPose paths.
• The new 3D Studio provides queued TripoSR, Stable Fast 3D, TRELLIS, Hunyuan3D, Meshy, generic REST, UniRig, and Blender pose-rig adapters plus a managed GLB/FBX/OBJ asset library.
• Blender bridge v0.3 exchanges pose/asset data and can queue generation or rigging from Blender.
• The cumulative SAM workflow includes one-click family setup, positive/negative point prompts, box-plus-point prompting, instance candidates, and semantic union masks.
• A top-level Help & Workflows tab documents SAM, pose, 3D generation, rigging, Blender, and troubleshooting.

See docs/V5_30_POSE_3D_GENERATION_RIGGING.md and docs/V5_30_SAM_SETUP_POINT_PROMPTS.md.

v5.28 persistent spatial layers and mask compositor

Detection boxes and segmentation masks now behave as persistent, editable layer stacks. Users can select and combine any number of model-generated, manual, Krita-imported, or previously composed layers; reorder, duplicate, hide, lock, rename, recolor, edit, and delete them; and retain the layers for later model runs. Detection supports union, intersection, average, and confidence-weighted composites. Segmentation supports soft-mask-preserving union, intersection, subtract, and XOR plus threshold, feather, and grow/shrink processing. Selected unsaved model previews can be promoted automatically when they are composed with saved layers.

The mask editor now includes brush, eraser, variable size/opacity/hardness, lasso, ellipse, rectangle, magic selection, add/subtract/replace modes, overlay transparency, and undo/redo. Saved edits keep revision/provenance data, and model-only cleanup no longer removes user-edited or composite layers. See docs/V5_28_SPATIAL_LAYER_COMPOSITOR.md.

v5.27 class-aware detection and segmentation

• Fixed closed-set YOLO inference so the requested class/name is resolved to a real trained class ID and passed to the model instead of merely renaming identical outputs after inference.
• Added safe custom-model class discovery from model.names, YAML/JSON/TXT/CSV sidecars, ONNX metadata, and safetensors metadata, with searchable class lists in Detection and Segmentation.
• Added generated-preview clear controls and separate deletion of saved model-generated boxes/masks while preserving user-drawn annotations.
• Added detector-guided semantic segmentation: a class-aware detector finds every matching instance and sends each bbox to SAM/SAM-HQ/SAM2 as an independent prompt.
• SAM-family models now explicitly reject a semantic class token without a spatial prompt or detector guide, preventing misleading identical class-labeled masks.
• Wired multiple-output controls into the actual runtimes: YOLO max_det, class filtering, NMS IoU, class-agnostic NMS, TTA, retina masks, and configurable SAM automatic-mask thresholds/crop settings.

See docs/V5_27_CLASS_AWARE_SPATIAL_INFERENCE.md for model semantics and workflow details.

v5.26 local-app handoff and separated spatial annotation workflows

• Gallery external-app actions now launch immediately and report errors directly instead of only changing to the Augment tab or hiding failures in a background job.
• Added bounded discovery for portable/local Topaz, Krita, and ComfyUI installations under the user home directory, common Windows application folders, and project-adjacent folders.
• Interactive handoffs use timestamped safe copies and a manifest so dataset originals are not overwritten. ComfyUI handoffs copy selected images into its input directory.
• Split the former combined annotation screen into Detection & Boxes, Segmentation & Masks, and Pose & 3D tabs.
• Added explicit /api/spatial/detection/* and /api/spatial/segmentation/* contracts for future orchestration nodes.
• Detection rejects mask-only models. Segmentation rejects bbox-only output and previews real mask PNGs as translucent overlays.
• Manual masks now use unique filenames, preserving multiple masks on the same image instead of overwriting earlier masks.

v5.24 model/source/geckodriver audit pass

• Added first-class RedRocket JTP-3 and RedRocket e6 Visual Ratings model rows for tag/rating workflows in Models, Tag Editor quick-run, Batch Tags, Compare, Annotation context, and orchestration. JTP-3 uses a native downloaded-repo inference path so its threshold, category exclusion, and implication options are not hidden behind a generic placeholder.
• Added offline validation for all bundled booru/generic downloader source definitions: e621, e926, Danbooru, Gelbooru, Safebooru, Rule34, Konachan, Yande.re, and Generic JSON. The Downloads API now exposes source validation for parser/page/tag/file-url configuration sanity checks.
• Improved the Firefox/geckodriver Source Browser with a private-mode smoke-test endpoint/button, stronger Firefox profile/download/log handling, fallback geckodriver asset resolution, and clearer status output.
• Added a system feature-audit endpoint to confirm critical feature groups and model/source coverage without needing to inspect code manually.

v5.19 reference/annotation/download stability patch

• Fixed Reference Finder rendering by wiring the missing query/training dataset selectors.
• Added an Annotation Editor tab with visual bbox drawing, polygon/mask drawing, mask rasterization, annotation listing, and model/assistant proposal hooks.
• Added Krita annotation-package export and mask import endpoints, plus an updated optional Krita bridge plugin action for sending masks back to the app.
• Hardened downloader job progress so parallel downloads report one monotonic total instead of jumping backward between presets/pages.
• Added max-pages/start-page downloader controls and safer pagination/duplicate-url guards to prevent stuck download loops.

v5.18 dependency, CUDA, tag-count, jobs, and downloader-control repair

This build hardens the large booru tag-export path and the install/update scripts.

• pyarrow is a required core dependency and is installed through both Conda and pip fallback paths.
• run.bat / run.sh check required dependencies and call the updater when the existing environment is missing core packages.
• update.bat / update.sh now install requirements.txt as well as the editable package.
• CUDA torch repair now replaces a CPU-only torch wheel with CUDA 12.8 wheels when NVIDIA GPUs are detected and DCT_INSTALL_TORCH=auto or cu128.
• e621/e926 official tag exports now require the canonical tags.csv.gz import to exceed one million rows, preventing partial exports such as ~119k rows from being accepted as complete.
• scripts/verify_tag_loader_full_count.py verifies a 1,668,155-row tags.csv.gz import end-to-end.
• Jobs can be cleared by checked rows, completed/failed status, or all non-running jobs.
• Downloader UI now exposes API/page delay, file delay, request timeout, retries, backoff, and parallel workers.
• Reference Finder and Orchestrate tab render errors caused by undefined frontend variables were fixed.

v5.16 PyArrow tag DB-export loader and alias/implication support

This build focuses on the booru tag dictionary import path. Official tags.csv.gz files with id,name,category,post_count now use a PyArrow-first streaming loader and keep all valid rows, including zero-count rows, so the active dictionary/search tables can exceed 100,000 tags instead of stopping at a small partial count.

The loader also parses tag_aliases.csv.gz and tag_implications.csv.gz using antecedent_name and consequent_name, and parses artists.csv.gz using name plus braced other_names values such as {alias1,alias2}. Alias and implication terms are inserted into the searchable dictionary with alias_target, alias_targets, or implications metadata so the suggestion menu can show corrections and related terms while users type.

The e621/e926 numeric category mapping is explicit: 0=general, 1=artist, 2=rating, 3=copyright, 4=character, 5=species, 6=invalid, 7=meta, 8=lore. The active dictionary, autocomplete search table, and legacy dictionary mirror are rebuilt after import so gallery/editor/comparer coloring and suggestions use the same loaded data.

v5.15 reference-finder, annotation, training-set, and source-browser integration

• Added a Reference Finder workflow ported from the uploaded character-reference tool as native FastAPI/HUD features rather than Gradio tabs.
• Added reference target/concept memory, saved reference images, reference runs, candidate detections, user verification labels, and reusable verification memory in SQLite.
• Added an always-available CPU demo backend for one/few-reference search using deterministic image similarity, with optional pipeline contracts for SigLIP2, OWLv2, Grounding DINO, YOLOE, and future SAM-style segmentation adapters.
• Added BBCode/booru-style tag query parsing and evaluation against verified positives/negatives, including [tag:name], quoted tags, bare tags, implicit AND, explicit AND/OR/NOT, parentheses, and wildcard tag terms.
• Added annotation helpers for bbox/polygon/mask records and YOLO/caption training-set export scaffolds so curated reference results can become training data.
• Added a Source Browser tool that installs/manages local geckodriver, launches Firefox in private mode by default, and exposes status/stop controls for authorized source-review workflows.
• Added model catalog rows for reference-finder detectors/verifiers and source-browser/Krita/media integration documentation while preserving the existing tag dictionary, metadata, downloader, model, orchestration, and media tooling.

v5.15 reference search, source browser, annotations, and training-set bridge

This build integrates the reusable dataset-curation features from the reference-image finder prototype into the modern FastAPI/HUD application. It adds a Reference Finder tab for one/few-reference character or concept search, duplicate-aware memory, user verification feedback, BBCode/tag-query optimization, manual bbox/polygon/mask annotation helpers, training-set creation, YOLO/caption exports, and training-job script scaffolding.

It also adds a Source Browser tab that can verify/install a local geckodriver, launch Firefox in private mode through Selenium when optional browser dependencies are installed, and stop the controlled browser session.

The always-available backend is a deterministic CPU ColorHash sanity workflow so users can test the full reference-search/verification/query loop without model downloads. OWLv2, SigLIP2, GroundingDINO/SAM, and YOLOE-related rows are exposed as optional model-backed contracts/download targets so support can be expanded model-by-model without changing the UI or database layout.

v5.14 original-style DB-export downloader hardening

This build fixes the observed e621 startup-sync failure where a canonical tags export around 69k rows was rejected before alias and implication exports could expand the searchable dictionary. The loader now imports canonical tags first, imports aliases and implications, expands the autocomplete/category dictionary from those relation files, and only then enforces the >100k searchable dictionary guardrail.

v5.12 verified tag DB-export ingestion pass

• refuses to treat alias-sized/partial tag exports as complete dictionaries
• pre-counts candidate tag CSV/GZ files before replacing the active dictionary
• keeps alias and implication exports in normalized tables without inflating tag counts
• migrates legacy alias-placeholder rows out of the dictionary/search tables
• adds large synthetic import tests proving more than 100,000 tag rows are retained

v5.10 complete tag dictionary import and metadata JSON field selection pass

• Tag dictionary import no longer discards zero-count rows from booru tag CSV exports; all valid tag names are preserved so autocomplete, category coloring, comparison, pruning, and assistant criteria can see the full dictionary.
• e621/e926 default export sync now always adds dated tags-YYYY-MM-DD.csv.gz, tag_aliases-YYYY-MM-DD.csv.gz, and tag_implications-YYYY-MM-DD.csv.gz candidates, even when generic fixed filenames exist.
• Startup sync now treats very small e621/e926 dictionaries as incomplete and retries the default export import instead of keeping a partial tag table.
• Metadata extraction now exposes a JSON schema/field picker for extracted image/video/LoRA metadata, including nested JSON strings. Users can select any paths, concatenate them in arbitrary order, choose input and output delimiters, and decide whether to keep or strip parentheses, curly braces, square brackets, and weight syntax.

v5.10 tag-dictionary completeness and metadata schema composition pass

• Fixed booru tag DB-export imports so zero-count tags are retained instead of skipped. The previous fast importer only kept rows with post_count > 0, which could reduce a several-hundred-thousand-row tag CSV to a much smaller active-tag subset. Zero-count tags now remain available for category color lookup, autocomplete, implications/aliases, comparison, batch edits, and custom dataset cleanup while still sorting below high-count tags.
• Metadata extraction now exposes a full flattened JSON/schema path browser. JSON-looking strings inside PNG/WebP/video/LoRA metadata are parsed recursively and shown with selectable paths using a ::<json> path segment.
• Added metadata field composition: users can select multiple metadata paths, choose the original delimiter, choose the final output delimiter, reorder paths manually, split to individual tags, preserve/remove parentheses/curly braces/square brackets, preserve/remove weighted prompt syntax, preview token analysis, and apply the result as tags or a caption.
• Added tests proving zero-count tags import and nested metadata JSON composition.

v5.9 parallel downloader and multi-GPU model runtime pass

• Downloader UI now exposes category/tag expansion, date range filters, newest/oldest ordering, per-category limits, and parallel preset/file download controls.
• Job runtime supports more concurrent jobs, while downloader/model jobs expose their own safe per-job parallel worker controls.
• Model registry now includes additional current local Hugging Face LLM/VLM rows, OpenAI cloud rows, OpenRouter cloud rows, and explicit GPU placement/sharding metadata.
• Model run settings support default no-sharding one-GPU placement, optional Accelerate-style sharding across selected GPU IDs, dtype, quantization, max-memory maps, and parallel prediction workers.
• Optional heavy serving dependencies for vLLM/SGLang/bitsandbytes are listed separately in requirements-models-serving.txt so Conda installs stay stable by default.

A local-first dataset curation application for building, auditing, tagging, captioning, pruning, augmenting, downloading, and exporting image/video datasets. This modernization replaces the old single-file UI approach with a Conda-first FastAPI backend, a browser-based single-page HUD, a SQLite project database, background jobs, and optional model adapters for auto-tagging, captioning, VLM/ViT inference, embeddings, deduplication, segmentation-assisted workflows, and dataset chat.

What changed

• Conda-first installation with additional pip dependencies for model ecosystems.
• Automatic CUDA PyTorch installation on NVIDIA systems when nvidia-smi is detected.
• No Gradio dependency.
• FastAPI backend with typed API contracts and background jobs.
• Local static single-page HUD served from the same Python app.
• SQLite database for datasets, media, tags, captions, groups, duplicates, predictions, presets, settings, download runs, and job history.
• Native folder picker endpoint for local dataset/import/output folder selection.
• Optional model layer that can launch without PyTorch, Transformers, CUDA, or model packages installed.
• Model registry supports built-in rule-based tagging plus optional Hugging Face, ViT, VLM, CLIP/DINO-style embedding, segmentation, and captioning adapters.
• Assistant tab for talking with selected dataset media through a built-in helper, a local text LLM, or a local VLM adapter.
• Gallery-first workflow with visible ordered tag strings, pointer-based drag/reorder tag chips, per-tag delete buttons, batch tag editing, category colors, side-by-side comparison, duplicate review, and export tools.
• Downloader plugins for multiple compatible booru-style JSON/DAPI sources plus a generic JSON source, gated behind user authorization confirmation.

v5.8 fast tag/database import and fast dataset indexing pass

• Tag DB-export imports now stream tags.csv.gz through one long SQLite transaction, insert only canonical profile rows while parsing, and rebuild autocomplete/legacy mirrors with set-based SQL instead of per-row Python inserts.
• Tag category normalization during import is now precomputed per profile instead of querying the profile/category table for every CSV row.
• Alias and implication exports are imported into normalized tag_aliases and tag_implications tables with bulk inserts instead of repeatedly updating JSON blobs one row at a time.
• Startup tag sync job progress now reports row-count messages while importing large tag exports.
• Dataset import now commits media, tag rows, captions, and metadata in batches instead of one SQLite transaction per file or per image tag list.
• Import category lookups are batched across the whole commit batch, which removes repeated per-tag/per-image dictionary queries.
• Embedded metadata extraction is no longer enabled by default during import because it is intentionally much heavier than sidecar parsing. It remains available as an explicit checkbox and in Settings.
• Perceptual hashes and near-duplicate scanning are explicit import options. Near-duplicate scanning is disabled by default because it can be quadratic on large datasets; run it only when you need that audit step.
• Import now exposes compute SHA-256, compute perceptual hash, probe dimensions, near-duplicate scan, worker count, and SQLite commit batch size controls.

v5.7 metadata/media/Krita/audio pass

• Added a Media Tools tab for generation-metadata extraction, video frame extraction, audio extraction/recording, and Krita edit packages.
• Vendored the reusable parser modules from the uploaded metadata toolkit as standalone backend integration code; no ComfyUI runtime or node registration is required.
• Image metadata extraction supports A1111-compatible PNG parameters, ComfyUI prompt/workflow metadata, NovelAI-style JSON/stealth metadata, Fooocus/Civitai/Invoke-style JSON comments, generic PNG/EXIF/WebP metadata exposed by Pillow, and LoRA references found in prompts.
• Video metadata extraction uses ffprobe when available and parses generation metadata stored in container/stream tags.
• LoRA metadata extraction supports safetensors headers, trigger words, training tags, base model, and architecture fields without tensor deserialization.
• Import can optionally read embedded generation metadata and use it as a fallback tag/caption source when .txt, .caption, or JSON sidecars are missing.
• Metadata extraction can preview tags/captions or apply them to selected media with the active tag profile/order strategy.
• Video tools can extract PNG frames at a chosen FPS or every-N-frames interval without adding JPEG-style recompression loss after decoding.
• Audio tools can extract audio streams from videos and save browser-recorded custom audio clips for future audio dataset workflows.
• Krita tools can create edit handoff packages with sidecars and import edited images back into the dataset while preserving tags/captions.

v5.6 DB-export refresh, category override, selected-editor navigation, and model catalog pass

• Selected/default tag profile DB exports are now considered stale after 336 hours / 14 days and are refreshed at startup when sync is enabled.
• e621/e926 sync prefers /db_exports/, keeps legacy /db_export/ as fallback, discovers dated export links, and imports the newest successful export for tags, tag aliases, and tag implications.
• Tag export metadata is tracked in SQLite so the HUD can show latest import time, missing roles, stale status, and per-role import results.
• The tag CSV importer follows the original e621 export mapping: name, category, and post_count, while also tolerating equivalent/custom column names. Zero-count tags are skipped for autocomplete.
• TXT-only custom datasets use the active profile dictionary for tag categories, ordering, autocomplete, implication pruning, and assistant criteria when JSON sidecars are absent.
• Global custom tag/category overrides now take priority over every selected booru profile and are persisted in runtime/custom_tags.json plus SQLite autocomplete tables.
• Tag Dictionaries now includes a global custom category/color editor and a global tag→category override editor.
• Gallery now has Reapply Profile/Custom Categories to recalculate visible/selected media tag colors from custom overrides and the active profile.
• Tag Editor can cycle through multiple selected gallery images with previous/next controls and a selected-image dropdown.
• Model catalog was expanded with newer/staged Hugging Face model rows for VLMs, OCR, image classification, watermark/safety filtering, PixAI-style tagging, JoyCaption-style captioning, model-builder classifier contracts, and clean-tags/LLM pruning workflows.

v5.5.1 DB-exports sync path fix, model downloads, and parallel import

• Startup can now auto-sync the selected/default booru tag database when the local dictionary is empty or stale.
• Supported default DB-export sync pulls compatible tag, alias, and implication exports when available, caches them under runtime/tag_exports/, and imports them into the local SQLite search tables.
• Custom datasets with only .txt tag sidecars can still use the selected booru profile for category colors, autocomplete, implication pruning, and tag ordering after dictionary sync.
• Import now parallelizes SHA-256 hashing, perceptual hashing, image-size probing, and JSON/TXT sidecar parsing across configurable CPU worker threads.
• DB writes remain serialized to protect SQLite integrity. GPU acceleration is intentionally reserved for optional post-import model inference/classification/tagging jobs.
• The Import tab exposes Parallel import workers and Auto-sync selected booru tag DB if empty/stale ≥2 weeks controls.
• Settings now include startup DB-export sync options, cache age, and default import worker count.
• The Models tab now has a scrollable model catalog with approximate model size, approximate VRAM need, precision, provider, repo id, download status, dry-run size check, download, and unload actions.
• The model registry includes runnable built-ins plus staged Hugging Face classifier/caption/LLM/VLM entries and placeholder records for larger tagger, embedding, and segmentation families that can be implemented one adapter at a time.

v5.4 gallery refresh and JSON-sidecar category fix

This build fixes the gallery/category issues reported after v5.3:

• Gallery toolbar now has Search / Refresh, Reload Page, and Refresh JSON/Sidecars + Reload buttons.
• The sidecar refresh action re-reads adjacent .txt, .caption, .json, image.ext.json, and .metadata.json files for the selected media, or the visible page if nothing is selected.
• JSON sidecars are parsed for booru-style grouped tags such as {"tags": {"character": [...]}}, explicit {"categories": {"tag": "category"}} maps, flat tag lists, tag dictionaries, captions, descriptions, and rating fields.
• Category maps from JSON sidecars are preserved in the media tag rows instead of being overwritten by dictionary guesses.
• Numeric booru category IDs are normalized before rendering, so values like 4 render as character, 3 as copyright, and 1 as artist.
• Gallery thumbnails request metadata for visible tags so dictionary-backed colors can render even when sidecar category data is incomplete.

v5.4 editor/compare assistant pass

• Added LLM/VLM/Assistant Tag Selection directly to the Tag Editor for the open image. Preview highlights matching chips; apply operations update only that image.
• Added LLM/VLM/Assistant Tag Selection directly to Compare for the current left/right pair. Preview selects matching chips on each side; apply operations update only those two displayed images.
• Multi-image comparison now defaults to the first two selected images in gallery order, with independent left/right cycling and dropdown selection so users can compare one-by-one without changing the full selection.
• Category criteria used by the assistant selector now consult saved per-image JSON-sidecar categories first, then dictionary metadata.

v5.3 concrete tag workflow pass

This build replaces the earlier placeholder/scaffolded tag-HUD work with visible, tested controls:

• Dedicated Tag Dictionaries tab with profile status, import-from-file, import-from-URL, and Load Default DB Export actions.
• Profile-specific autocomplete dictionaries for booru/custom workflows.
• Indexed SQLite search mirror for responsive top-k suggestions while typing.
• Suggestion dropdowns show category-colored tag chips and post counts.
• Unknown tags can be remembered into the active profile and are persisted in runtime/custom_tags.json.
• The tag editor uses one prompt-style left-to-right strip with category colors, pointer drag/drop reordering, insertion marker, raw text sync, and per-chip × deletion.
• Added a dedicated Batch Tags tab so quick select and add/remove/set/replace/copy/prune/order operations are always visible.
• Compare view exposes selectable tag chips, add/move/copy buttons, first-two defaults for multi-selection, left/right image cycling, and assistant tag selection for the current pair.
• Profile precedence can be edited from the HUD so booru, custom dataset, and LoRA/style/character/concept orderings can be maintained by the user.

Install

Windows

Run from Anaconda Prompt or Miniconda Prompt:

install.bat
run.bat

The default installer uses this behavior:

• DCT_INSTALL_TORCH=auto by default.
• If nvidia-smi is found, CUDA 12.8 PyTorch wheels are installed.
• If no NVIDIA GPU is detected, PyTorch is skipped by default.
• Set DCT_INSTALL_TORCH=cpu to force CPU PyTorch.
• Set DCT_INSTALL_TORCH=skip to skip PyTorch.

For a CUDA reinstall inside the Conda environment:

install_torch_cuda128.bat
verify_gpu.bat

Linux/macOS

chmod +x install.sh run.sh
./install.sh
./run.sh

CUDA helper:

./install_torch_cuda128.sh
./verify_gpu.sh

Manual Conda

conda env update -f environment.yml --prune
conda activate data-curation-tool
python -m pip install -e .
python run.py --host 127.0.0.1 --port 7865 --open-browser

Open the local URL printed in the terminal. By default, the app uses:

http://127.0.0.1:7865

Core feature coverage

Dataset and media management

• Load local datasets recursively from folders.
• Use HUD buttons to select dataset folders through the native operating-system folder picker.
• Import one folder or a queue of multiple folders.
• Track image, animated image, video, and metadata sidecar files.
• Validate common image formats: JPG/JPEG, PNG, WEBP, BMP, GIF, TIFF, AVIF when Pillow supports it.
• Track video-like files for later frame/audio extraction: MP4, WEBM, MOV, MKV, AVI.
• Generate thumbnails without changing original files.
• Compute SHA-256 and perceptual average hashes.
• Skip duplicates by exact hash and flag near duplicates by Hamming distance.
• Store metadata in SQLite for fast filtering and resume-safe workflows.

Ordered category-colored tag workflow

• Tags remain in one left-to-right prompt-style order.
• Tag category is shown with color only; categories are not forced into separate sections.
• A legend shows the category/color mapping.
• Drag tags with the mouse to reorder them.
• Click the × inside a tag chip to delete it.
• Save the exact ordered tag string back to the sidecar.
• Raw tag-string editing remains available for power users.
• Bulk add, remove, replace, set, copy, and prune tags.
• Tag autocomplete with trie-safe sanitization for malformed CSV rows.
• Tag implication pruning so broader tags can be removed when more specific tags exist.
• Caption splitting into valid tags, inferred tags, and filler text.
• Side-by-side image comparison with tag transfer/removal.
• Persistent image groups with save/load/delete/subtract operations.

Model support

• Rule-based offline tagger that works immediately.
• Built-in caption splitter adapter.
• Built-in dataset assistant for tag/caption planning before a model is installed.
• Optional Hugging Face image classifier adapter.
• Optional Hugging Face image captioning adapter.
• Optional Hugging Face text LLM chat adapter using options.model_id or a local model path.
• Optional Hugging Face VLM image chat adapter using selected dataset images or external image paths.
• Optional ViT/embedding adapter contract for similarity, clustering, and dedupe-assisted review.
• Optional segmentation adapter contract for mask-assisted curation.
• GPU/CPU capability detection that reports NVIDIA GPUs through nvidia-smi even when torch is not importable yet.
• Torch CUDA readiness check through torch.cuda.is_available() when torch is installed.
• Multi-device scheduling hooks for CPU/GPU routing and future distributed inference.

Assistant workflow

The Assistant tab lets the user talk to the selected dataset/media context.

Supported modes:

• Built-in no-model assistant for immediate tag/caption strategy.
• Text LLM chat through the local Hugging Face text-generation adapter.
• VLM image chat through the local Hugging Face image-text-to-text adapter.
• Optional voice input from browsers that support Web Speech Recognition.
• Optional application of suggested tags/captions to the selected media.

For model-backed chat, install torch and provide a local path or Hugging Face model id in the HUD model-id field.

Downloading, presets, and archives

Supported source adapters:

• e621 JSON API
• e926 JSON API
• Danbooru JSON API
• Gelbooru DAPI JSON
• Safebooru DAPI JSON
• Rule34 DAPI JSON
• Konachan JSON API
• Yande.re JSON API
• Generic JSON source

Downloader behavior:

• Runs only after the user confirms they are authorized to download from the configured source.
• Saves downloaded files to a selected output folder.
• Writes tag sidecars when the source response includes tags.
• Writes .download.json metadata sidecars for traceability.
• Presets are stored separately from runtime settings.
• Preset import supports positive/negative sections separated by ;;;.
• Batch preset actions include archive, remove, and delete with normalized single/multi-select behavior.

Augmentation and post-processing

• Include augmented data with originals or generate augmented-only outputs.
• Image augmentations include horizontal flip, rotation, brightness, contrast, sharpen, denoise, crop square, and long-side resize.
• The preprocessing resize used for model inference is not saved back over the curated image unless the user explicitly exports an augmentation.
• Session output options allow generated frames or extracted artifacts to be attached to the current dataset.

v5.7 metadata, video, audio, and Krita bridge additions

• Embedded generation metadata extraction for PNG/JPEG/WebP/video files using the standalone parser code from the media metadata toolkit.
• Supported metadata sources include Automatic1111-style parameters, ComfyUI prompt/workflow JSON, NovelAI-style metadata/stealth PNG payloads, Fooocus/Civitai JSON comments, video container tags exposed through ffprobe, and safetensors LoRA headers.
• Metadata extraction can preview, persist, or apply derived tag strings/captions to selected media while preserving the selected tag profile and category/color resolution.
• Video tools can decode high-quality PNG frames at a target FPS, every-N-frame interval, and optional start/end timestamps.
• Video audio extraction can save WAV/FLAC/M4A/MP3/Opus/copy outputs when ffmpeg is installed.
• Browser audio recording can save custom clips for future audio dataset workflows.
• Optional Krita bridge workflow can create editable image packages, import edited images back into a dataset, and download a minimal Krita Python plugin for sending the active document back to the local app.

Export and interoperability

• Export tag/caption sidecars.
• Export JSONL manifests.
• Export CSV metadata.
• Export YOLO classification text files.
• Export COCO-like metadata skeleton for downstream annotation pipelines.
• SQL query console for advanced users.

Runtime folders

runtime/
  app.db
  settings.json
  thumbnails/
  presets/
  downloads/
  exports/
models/
  hf/
  local/
outputs/

The runtime folders are created on first launch and are ignored by git.

Launch flags

python run.py --host 127.0.0.1 --port 7865
python run.py --runtime ./runtime --models ./models --open-browser
python run.py --reload

API overview

• GET /api/health — health check.

• POST /api/system/pick-folder — open native local folder picker.

• GET /api/system/devices — detect CPU/GPU/torch devices.

• POST /api/datasets/import — import a local folder as a dataset.

• POST /api/datasets/import-many — import multiple local folders as one background job.

• GET /api/datasets — list datasets.

• GET /api/media — gallery/filter page.

• GET /api/media/{id} — media metadata.

• GET /api/media/{id}/file — raw media file.

• GET /api/media/{id}/thumbnail — thumbnail.

• PUT /api/media/{id}/tags — replace visible ordered tag string.

• POST /api/tags/bulk — bulk tag operations.

• POST /api/tags/prune — tag implication pruning.

• POST /api/models/run — run model inference as a background job.

• POST /api/models/chat — chat with the dataset assistant, local LLM, or local VLM.

• GET /api/downloads/sources — list compatible downloader adapters.

• POST /api/downloads/run — run direct or preset downloads.

• GET /api/jobs — list jobs.

• GET /api/jobs/{id} — inspect job status.

• POST /api/augment/run — run augmentation as a background job.

• POST /api/export/run — export dataset assets/manifests.

• GET /api/metadata/sources — list embedded metadata parser capabilities.

• POST /api/metadata/extract-now — extract/apply metadata immediately.

• POST /api/media-tools/video/extract-frames — queue frame extraction from selected videos.

• POST /api/media-tools/video/extract-audio — queue audio extraction from selected videos.

• POST /api/media-tools/audio/recording — save browser-recorded audio.

• GET /api/krita/plugin — download the optional Krita bridge plugin zip.

• POST /api/database/query — execute read-only SQL.

Development

conda activate data-curation-tool
python -m pip install -e .
pytest
python -m compileall data_curation_tool

The frontend is plain modern browser JavaScript so it can run offline without Node, npm, or a bundler. A future Tauri/Electron shell can wrap the same backend and static HUD without changing the API.

License

GPL-3.0-or-later.

v5.7 media metadata, video/audio extraction, and Krita bridge pass

This pass adds a Media Tools workflow without turning the application into a ComfyUI node pack. The reusable parser logic is integrated as normal backend services so datasets can be curated directly inside the local web application.

Generation metadata extraction

The Media Tools tab can extract tags, tag strings, captions, LoRA references, and prompt/settings text from supported image/video metadata sources, including:

• Automatic1111-style PNG parameters
• ComfyUI prompt/workflow JSON stored in image metadata
• StableSwarm/SwarmUI-style JSON metadata
• NovelAI metadata and stealth PNG payloads where available
• Fooocus/Civitai/InvokeAI/generic JSON comment blocks
• XMP/EXIF/comment metadata exposed by Pillow
• video container tags discovered through ffprobe
• LoRA safetensors headers and training tag-frequency metadata through the safe metadata parser

Metadata extraction has non-destructive preview behavior by default. When requested, extracted tags/captions can be applied to selected media while preserving the existing tag-order/category system.

Video frame extraction

Selected videos, or external video paths, can be decoded to high-quality PNG frames at a user-selected FPS or every-N-frames interval. PNG is the default output format to avoid adding another JPEG-style lossy encode after video decoding. The extraction job can optionally attach generated frames back into the current dataset.

Audio extraction and recording

Videos can have audio streams extracted to WAV/FLAC/M4A/MP3 for future audio-dataset workflows. Browsers that support MediaRecorder can also record custom audio clips from the Media Tools tab and save them to the local runtime recording folder.

Krita bridge

The Krita bridge exports selected images plus sidecars into an edit package with a manifest. Users can open the exported image in Krita, save an edited copy, and import that edited image back into the source dataset while preserving tags and captions when requested.

v5.15 reference finder, query optimizer, annotation, and browser pass

This pass integrates reference-image dataset curation workflows into the modern FastAPI/HUD application.

Reference Finder

The Reference Finder tab can register a target character/object, attach one or more reference images, and scan selected media, an active dataset, or an external folder. A CPU-only ColorHash verifier is always available for smoke tests and light local matching. Optional OWLv2/SigLIP2-style detector/verifier pipelines are represented in the model catalog and service contracts so model-specific adapters can be enabled as dependencies and weights are installed.

Verification memory and query optimization

Search results can be marked correct, incorrect, or uncertain. The app stores verification memory and can evaluate BBCode/tag-style queries such as [tag:blue_hair] AND NOT [tag:group] against known positives/negatives. This helps users discover better tag queries before downloading, filtering, or building a training set.

Annotation and training exports

The tool now stores bbox, polygon, and mask-style annotation records. It can create query-derived training sets and export YOLO detection/segmentation folder structures or caption JSONL files for downstream training workflows.

Source Browser

A local Firefox/geckodriver source browser tab was added for authorized browsing/review workflows. It can install/verify geckodriver, launch Firefox in private mode by default, and stop the local browser session from the HUD.

v5.15 reference-finder and annotation/training additions

The current build includes a new Reference Finder tab for one/few-reference image search, verification feedback, BBCode-style tag-query evaluation, annotation primitives, training-set creation, YOLO export, caption JSONL export, and training script scaffolding. It also includes a Source Browser tab for Firefox/geckodriver status, local geckodriver installation, private-mode Firefox launching, and browser shutdown.

The built-in demo_colorhash reference pipeline is available immediately as a no-model sanity backend. OWLv2, SigLIP2, Florence-2, Grounding DINO, SAM2, and YOLO/YOLOE are represented in the model catalog as downloadable/staged integration points so model-specific adapters can be completed incrementally without breaking startup on machines that do not have those optional dependencies installed.

v5.16 tag DB-export loader fix

The tag dictionary importer now uses PyArrow first for booru-style tags.csv.gz files with id,name,category,post_count, then falls back to the built-in Python CSV path for custom files. It also imports tag_aliases.csv.gz, tag_implications.csv.gz, and artists.csv.gz as separate export roles so tag suggestions can surface alias targets, implication targets, and artist alias corrections while the user types.

The e621/e926 numeric category mapping is now explicit:

{"0":"general","1":"artist","2":"rating","3":"copyright","4":"character","5":"species","6":"invalid","7":"meta","8":"lore"}

The active profile dictionary, autocomplete/search table, and legacy mirror are rebuilt from the same loaded data so category coloring and suggestions stay consistent across the gallery, tag editor, comparer, batch tools, and assistant tag-selection workflows.

Recovering from a malformed local SQLite database

If the computer/app is stopped while a very large DB-export import is running and startup reports database disk image is malformed, run:

Windows:

repair_database.bat

Linux/macOS:

./repair_database.sh

The repair script does not delete files permanently. It moves runtime/app.db, runtime/app.db-wal, and runtime/app.db-shm into runtime/corrupt_databases/<timestamp>/ and lets the app create a clean database at next startup.

v5.18 database safety and large tag import recovery

This build adds a startup SQLite quick-check. If runtime/app.db is malformed, the app automatically moves app.db, app.db-wal, and app.db-shm into runtime/corrupt_databases/<timestamp>/ and creates a fresh database instead of crashing.

The tag DB-export importer now stages large tag dictionaries under a temporary source first, validates the row count, and only then atomically promotes the staged rows to the selected profile. It also removes unsafe PRAGMA synchronous=OFF usage from tag imports.

For an existing malformed database, use repair_database.bat or ./repair_database.sh to quarantine the database files manually before startup.

v5.20 annotation model tooling

The Annotation Editor includes first-class bbox, mask, 2D pose, 3D pose, and animation-pose dataset controls. SAM/SAM-HQ/SAM2, YOLO detection/segmentation/pose, custom local YOLO/SAM-style checkpoints, VLM/API proposal routes, and pose dataset contract rows are exposed in the model catalog. Optional heavy annotation dependencies can be installed with install_annotation_models.bat or install_annotation_models.sh.

v5.21 annotation/model/media updates

This build improves the annotation workflow with first-class dependency installation, on-page annotation model downloads, load/validate/unload controls, additional YOLO11 detection/segmentation/pose rows, a lightweight 3D pose/bone viewer, a Blender bridge plugin, annotation deletion/loading into draft, and richer gallery playback for image, animated, video, and audio media.

For annotation model runtimes, use the Annotation Editor button Install Annotation Dependencies or run:

install_annotation_models.bat

Then use Download Weights, Load / Validate, and Preview Proposal or Generate + Save from the Annotation Editor.

v5.22 annotation model preview/generation hardening

This build removes the no-model deterministic preview fallback from the annotation workflow. Preview and Generate + Save now only create proposals when a selected runnable model returns actual model output, or when the user manually draws/saves an annotation. Missing optional dependencies, missing checkpoints, unsupported SAM2/SAM-HQ adapter rows, and zero-detection model runs return structured errors with zero proposals instead of fake center boxes/masks.

The Annotation Editor now displays this behavior in the HUD and includes tests proving that no-model preview/save cannot create fake annotations.

v5.23 SAM/SAM-HQ/SAM2 annotation dispatch fix

• Restricts the Annotation Editor model dropdown to real spatial annotation models so generic assistants cannot be accidentally used as SAM checkpoints.
• Fixes checkpoint resolution so SAM/SAM-HQ/SAM2 use actual downloaded files or explicit custom local paths, not model keys such as dataset-assistant.
• Adds SAM-HQ runtime support through segment-anything-hq and exposes it in the annotation dependency installer.
• Adds a SAM2.1 image proposal adapter path for bbox-prompt and automatic masks when the optional sam2 runtime is installed.
• Keeps no-fake-fallback behavior: failed/missing models return structured errors and do not create artificial boxes or masks.

v5.35 Metadata, Model Runtime, and Conversation-State Update

• Added model runtime contract audit with a dedicated JTP-3 parser regression.
• Added inline metadata extraction/compose controls to Tag Editor, Compare, and Batch Tags.
• Added persistent assistant conversations with resume, archive, and fork-from-message support.
• Added metadata-aware assistant context for selected media and external paths.
• Preserved neutral metadata/ComfyUI bridge naming without legacy source-node prefixes.

See docs/V5_35_METADATA_MODEL_CHAT_AUDIT.md.

v5.40 Dropdown Stability and Previous-Install Asset Migration

• Fixed polling-driven rerenders that could close open model/category/profile dropdowns before the user could choose an option. The frontend now defers automatic renders while any input, textarea, or select control is actively being used.
• Added an Install Migration tab for reusing assets from one or more older app installs. It scans newest installs first, then older installs for unique files.
• Migration can move or copy model folders, cached tag DB-export files, imported tag dictionary database rows, custom tags, presets, downloaded media cache, and outputs into the current install layout.
• Added startup migration settings and environment overrides so cached tag exports/model assets can be brought in before startup tag DB sync decides whether a network download is needed.

v5.78 additions

• Expanded the 3D generation catalog and 3D Studio provider contracts for text-to-3D, image-to-3D, multi-image-to-3D, video-to-3D, ComfyUI workflow handoff, and Blender refinement overlap.
• Added MCP tool discovery/status/config endpoints and installers for Blender, Krita, Audacity, OBS Studio, and ComfyUI bridge usage.
• Added provider-level cloud model runtime defaults, including an OpenRouter DeepSeek V4 Pro template with context-shrinker metadata.
• Added booru/e621 Boolean logic gates for Downloads and Presets, with preview expansion and raw append mode.
• Updated the GitHub wiki docs under docs/wiki/27-3D-Generation-MCP-Cloud-and-Booru-Logic.md.

Dataset Pipeline / Training Prep / 3D Print Handoff

The Dataset Pipeline tab prepares branch datasets for external training tools without mutating global originals. It can generate caption/tag rule packets and model/VLM prompts for LoRA, IC-LoRA, ControlNet, and embedding workflows; evaluate branch readiness; dry-run/apply deterministic cleanup to branch-local sidecars; export trainer manifests/config stubs for tools such as Kohya SS, OneTrainer, Diffusers scripts, LTX Trainer, ComfyUI training nodes, cloud trainers, or future source-manifest/webscraper bridges; and package 3D assets for Blender/ZBrush/MeshLab/slicer handoff. See docs/wiki/29-Dataset-Pipeline-Training-Prep-and-3D-Print-Handoff.md.

v5.78.14 — RedRocket Hydra 3.5, FLUX/Chroma prep, and remote tagger offload

This build adds RedRocket Hydra 3.5 as a first-class tagger/rating classifier across Models, Tag Editor, Compare, Batch Tags, annotation quick-run cards, and prediction/tag-selection surfaces. Hydra can run locally through the downloaded repo-native inference.py path or remotely through service.py by setting options.hydra_service_url. Remote Devices can now start Hydra service processes, plan model-run shards, and dispatch tagger/model inference to configured worker APIs. Dataset Pipeline and Pipeline Prep also include FLUX.1 Dev/Schnell/Kontext/Fill/Depth/Canny/Redux plus Chroma/FLUX-tuned caption-rule targets. See docs/wiki/30-Hydra-3-5-FLUX-Chroma-and-Remote-Tagger-Offload.md.

v5.8.0 — Legacy local taggers and tag/caption translation

This release adds the legacy/local image taggers from the original model configuration into the Models catalog, including Thouph EVA02-CLIP 7704, Thouph EVA02 ViT-Large 448 8046, and Thouph Experimental EfficientNetV2-M 8035. These models use a shared legacy adapter that preserves their original preprocessing, thresholds, ONNX/PyTorch runtime choices, tag metadata ordering, placeholder output handling, and e621 alias/implication cleanup. The Tag Dictionaries tab also now includes a profile-aware tag/caption translator for converting between booru profiles or caption formats, with optional local/cloud LLM/VLM handoff for uncertain mappings. See docs/wiki/31-Legacy-Taggers-and-Tag-Caption-Translation.md.

• v5.8.15 Dashboard Migration Progress Refresh
• v5.8.14 Legacy Tagger Runtime and Scroll Stability Fixes
• v5.8.13 Startup Migration Progress, Attention Heatmaps, and Graph Runtime
• v5.8.12 Model Prediction Normalization and Unload UI Fixes
• v5.8.11 Graph Editor, Browser MCP, Scroll, and Startup Resume Fixes
• v5.8.9 Startup Progress Time Import Fix
• v5.8.8 Hydra UTF-8, Startup Progress, and Gallery Fixes

v5.8.10 — Agentic Graph Editor and Browser MCPs

• Ports standalone graph-editor concepts into the integrated Agentic Graph Editor while keeping the existing dark/neon canvas style.

• Adds multimodal input nodes, bundle/context nodes, model-call nodes, supervisor nodes, external-tool nodes, event-console support, browser search/open nodes, port connections, pan/zoom, edge deletion, and flow animation.

• Adds browser MCP entries for default browser, Edge, Chrome, Firefox, Chromium, and Tor Browser. Browser MCP actions are visible/user-approved handoffs, not hidden scraping.

• See docs/V5_8_10_AGENTIC_GRAPH_BROWSER_MCP.md and docs/wiki/40-Agentic-Graph-Editor-Standalone-Port-and-Browser-MCPs.md.

• v5.8.11: fixes tab scroll preservation and makes migration jobs resume the Dashboard startup-maintenance progress indicator after a cancelled first-run tag sync.

v5.8.13 Startup / Attention / Agentic Graph Functionality

Adds live startup progress during migration, an Attention Visualizer tab for Grad-CAM/CAM/U-Net/t-SNE/cross-attention contracts, and deeper standalone graph-editor compatibility.

v5.8.15

• Dashboard manual refresh and live manual-migration startup-maintenance progress.
• Chunked migration copy/move progress and chunked migrated tag database import progress.
• No-cache startup status polling.

v5.8.14

• Fixed legacy EfficientNetV2-M preprocessing so ONNX/PyTorch inference always receives the required 448x448 tensor.
• Patched legacy EVA/timm pickle compatibility for missing reg_token.
• Reduced scroll jitter by replacing long multi-second restoration loops with short-lived tokenized scroll restores.

v5.8.16 Update

• Agentic Graph Editor palette cards now expose standalone-style node categories and palette-driven customization fields.
• Quick model inference completion now refreshes prediction scores and affected media more directly so Tag Editor hover cards update sooner.
• Prediction hover rows use stable, unique per-model color variables with multi-model average rows.
• Legacy EVA/TIMM taggers receive additional neutral compatibility defaults for newer runtime environments.

See docs/V5_8_16_GRAPH_PALETTE_PREDICTION_REFRESH_EVA_FIXES.md.

v5.8.17 Update

This update fixes the Agentic Graph Editor interaction layer, migration-finalization progress, and the post-100% frontend hydration stall. The graph canvas now supports background drag panning, cursor-centered wheel zoom, right-click node selection menus, port-based node connection, node context menus, and smoother direct-manipulation updates without replacing the existing dark/neon visual theme. Manual migration progress is also remapped so the raw file-copy phase no longer shows as 100% while reconciliation and post-migration tag/model checks are still running. The browser now renders dashboard essentials first and hydrates optional catalogs in the background while the Dashboard startup-maintenance circle displays a frontend_hydration phase.

See docs/V5_8_17_GRAPH_CANVAS_AND_MIGRATION_FINALIZATION.md.

v5.8.18 Update

• v5.8.18 EVA Norm-Pre and Graph Editor Event Handler Fixes

v5.8.19 Update

• v5.8.19 Migration Local Cache Reuse and Weekly Startup Sync

v5.8.20 Update

• Install Migration now supports parallel file transfers, enabled by default with four workers for SSD migrations.
• New installs with no migrated/local tag dictionary rows bootstrap the active/default tag dictionary immediately.
• Startup tag dictionary network checks remain gated to once per week by default, can be disabled, and can be forced manually from Settings or Tag Dictionaries.

See docs/V5_8_20_PARALLEL_MIGRATION_AND_FIRST_RUN_SYNC.md and docs/wiki/50-Parallel-Migration-and-First-Run-Tag-Sync.md.

v5.8.24 runtime and UI fixes

• Hydra 3.5 is treated as a local tagger by default rather than an MCP/API-oriented catalog row.
• Legacy EVA taggers include additional newer-timm compatibility shims.
• Explicit GPU assignments are stricter for legacy PyTorch/ONNX taggers, including ONNX CUDA device IDs.
• Model-page interactions use short catalog caching, and gallery/tag tabs refresh more reliably after completed jobs.

v5.8.25 EVA compatibility and live refresh fixes

• Adds the remaining neutral EVA/timm compatibility defaults for the Thouph EVA02 legacy tagger, including attn_pool, head_drop, and pos_drop.
• Explicit model, gallery, and job refresh actions now force state-preserving UI repaint instead of waiting for tab-switch refresh.
• Completed inference jobs trigger a faster visible refresh of model/tag/gallery state.

See docs/V5_8_25_EVA_ATTENTION_POOL_AND_LIVE_REFRESH_FIXES.md and docs/wiki/55-EVA-Attention-Pool-and-Live-Refresh-Fixes.md.

v5.8.26 Update

• Fixes the app-wide scroll jump caused by automatic polling refreshes restoring stale shell/window scroll state.
• Restores the legacy EfficientNetV2-M tagger to the original PyTorch-first runtime preference.
• Updates ONNX model dependencies to use onnxruntime-gpu so ONNX taggers can use CUDA providers after running the updater.
• Allows ONNX-only legacy rows to remain loadable with an explicit CPU fallback warning if the CUDA provider is still unavailable.

See docs/V5_8_26_SCROLL_AND_LEGACY_ONNX_GPU_FIXES.md.

v5.8.47 memory guard, assistant model tools, and tag pruning

• Added system-RAM guardrails for long-running LLM/VLM sessions. Automatic CPU offload is disabled by default and skipped under RAM pressure.
• Bounded stored chat context/response payloads to reduce overnight memory growth.
• Expanded the condensed conversation memory panel so it can be read, copied, or downloaded without truncation.
• Wired Tag Editor assistant pruning to active Dataset Pipeline / LoRA rules and real tag-list updates.
• Added approved assistant/orchestrator model queue tools for model load, inference, unload, and wait-for-job operations.

v5.8.47 Update

Memory guard, assistant model tools, and Tag Editor tag pruning fixes.