✨ New from the same author: SpindleStudio — a Threads-style parallel image generation studio. Its images carry SD WebUI–compatible metadata and are indexed by IIB out of the box.

中文文档 Change log Installation / Running

Infinite Image Browsing (IIB)

Software Support and Development Progress Overview

If you would like to support more software, please refer to: parsers or pixiv_iib_plugin

Key Features

🔥 Excellent Performance

• Once caching is generated, images can be displayed in just a few milliseconds.
• Images are displayed with thumbnails by default, with a default size of 512 pixels. You can adjust the thumbnail resolution on the global settings page.
• You can also control the width of the grid images, allowing them to be displayed in widths ranging from 64px to 1024px.
• Supports pre-generating thumbnails and video covers to improve performance using --generate_video_cover and --generate_image_cache.
• Supports specifying the cache directory through the IIB_CACHE_DIR environment variable.

🔍 Image Search & Favorite

• The prompt, model, Lora, and other information will be converted into tags and sorted by frequency of use for precise searching.
• Supports tag autocomplete, auto-translation, and customization.
• Image favorite can be achieved by toggling custom tags for images in the right-click menu.
• Support for advanced search similar to Google
• Also supports fuzzy search, you can search by a part of the filename or generated information.
• Support adding custom search paths for easy management of folders created by the user.
• Media type filtering, video tag search, and random sort.
• Auto-tagging with custom rules.

🖼️ View Images/Videos & Send To

• Supports viewing image generation information. Also supported in full-screen preview mode.
• EXIF/metadata is integrated in full-screen preview with nested JSON navigation and highlighting.
• Supports sending images to other tabs and third-party extensions such as ControlNet , openOutpaint.
• Support full-screen preview and enable custom shortcut key operations while in full-screen preview mode.
• Support navigating to the previous or next image in full-screen preview mode by pressing arrow keys or clicking buttons.
• Support playing video files from a remote server.

💻 Multiple Usage Methods

• You can install it as an extension on SD-webui.
• You can run it independently using Python.
• The desktop app version is also available.
• Supports multiple popular AI software.
• NEW: Use with AI agents (Claude Code, Cursor, OpenClaw, etc.)

📊 Trend & Statistics

• GitHub-style contribution heatmap showing daily image generation activity.
• Monthly trend bar chart, top models, and top source distribution.
• Total image count and disk usage at a glance.
• Click the 📊 icon on the startup page or find "Trend" under Launch.

🎵 TikTok-Style View

• TikTok-style vertical browsing for images and videos.
• Polished info panel with backdrop/preview return improvements.
• Delete events stay in sync across the TikTok view.

🚶‍♀️ Walk Mode

• Automatically load the next folder (similar to os.walk), allowing you to browse all images without paging.
• Tested to work properly with over 27,000 files.
• When there are folders, you can switch to walk mode from other modes by clicking the walk button in the upper right corner. It will flatten all the folders, avoiding the tedious operation of going in and out of folders.

🌳 Preview based on File Tree Structure & File operations

• Supports file tree-based preview.
• Supports automatic refreshing.
• Supports basic file operations, such as multiple selection for deleting/moving/copying, and creating new folders.
• Hold down the Ctrl, Shift, or Cmd key to select multiple items.
  • Supported multi-select operations include: delete, move, copy, pack download, add tags, remove tags, move to another folder, copy to another folder, drag and drop.
  • You can keep the multi-select state by clicking the "Keep Multi-Select" button in the lower right corner, allowing you to perform multiple operations on the selected file collection conveniently.
• Drag-and-drop into folders and safer move/copy (continue on error).

🆚 image comparison (similar to Imgsli)

• Provides a side-by-side comparison of two images.
• Provides a comparison of image generation information at the same time.

🧠 Topic/Tag Analysis

• Tag relationship graph visualization for topic clusters.

🗂️ Smart Organize

AI-powered automatic file organization

• Semantic Clustering: Automatically groups similar images based on prompt semantics using AI embeddings
• Auto-Generated Folder Names: AI generates meaningful folder names in your preferred language
• Preview Before Action: Review the proposed organization before confirming - skip or adjust as needed
• Background Processing: Large folders are processed in the background, you can continue working
• Flexible Options: Choose between move or copy, set minimum cluster size, include subfolders recursively

Requirements: Same as Topic Search - requires OPENAI_BASE_URL, OPENAI_API_KEY, and Python dependencies numpy, hnswlib

📸 See Smart Organize Preview below for screenshots and video demo.

🌐 Multilingual Support

• Currently supports Simplified Chinese/Traditional Chinese/English/German.
• If you would like to add a new language, please refer to i18n.ts and submit the relevant code.

🔐 Privacy and Security

• Supports custom secret key for authentication.
• Supports configuring access control for the file system, which will be enabled by default when the service allows public access (Only when used as an extension of sd-webui).
• Supports customizing the allowed paths for access control.
• Supports controlling access permissions. You can run IIB in read-only mode.
• Click here to see details

📦 Packaging/Batch Download

• Allows you to download multiple images at once.
• The data source can be search results, a regular image grid view page, walk mode, etc. Images can be added to the processing list through drag-and-drop or "Send To".

⌨️ Keyboard Shortcuts

• Allows for deleting and adding/removing tags, with customizable trigger buttons in the global settings page.

If you like this project and find it helpful, please consider giving it a ⭐️. This would be very important for me to continue developing and maintaining this project. If you have any suggestions or ideas, please feel free to raise them in the issue section, and I will respond as soon as possible. Thank you again for your support!

Sponsor me on WeChat

Installation / Running

As an extension for SD-webui:

1. Open the Extensions tab in SD-webui.
2. Select the Install from URL option.
3. Enter https://github.com/zanllp/sd-webui-infinite-image-browsing.
4. Click on the Install button.
5. Wait for the installation to complete and click on Apply and restart UI.

As a standalone program that runs using Python. (without SD-webui):

Refer to Can the extension function without the web UI?

If you need to view images generated by ComfyUI/Fooocus/NovelAI, please refer to https://github.com/zanllp/sd-webui-infinite-image-browsing/issues/202.

If you need a Dockerfile, you can refer to this link. https://github.com/zanllp/sd-webui-infinite-image-browsing/discussions/366

As a desktop application (without SD-webui and Python):

The executable version also supports ComfyUI/Fooocus/NovelAI.

Download and install the program from the releases section on the right-hand side of the repository page. If the antivirus detects a virus, it can be ignored as a false positive. There are two versions of the compiled version for Windows, with the pyinstaller version having a lower false positive rate.

If you need to compile it yourself, please refer to https://github.com/zanllp/sd-webui-infinite-image-browsing/blob/main/.github/workflows/tauri_app_build.yml.

As a Library Usage:

Use iframe to access IIB and use it as a file browser for your application. Refer to https://github.com/zanllp/sd-webui-infinite-image-browsing/blob/main/vue/usage.md

As an AI Agent Skill

IIB can be used with AI agents like Claude Code, Cursor, and OpenClaw. See AI Agents Documentation for details.

Preview

Image Search

During the first use, you need to click and wait for the index generation. For my case with 20,000 images, it took about 45 seconds (with an AMD 5600X CPU and PCIe SSD). For subsequent uses, it will check whether there are changes in the folder, and if so, it needs to regenerate the index. Usually, this process is very fast.

Image search supports translation, see https://github.com/zanllp/sd-webui-infinite-image-browsing/issues/39 for more detail. Feel free to share files for other languages to facilitate everyone's use.

Full Screen Preview (Side-by-Side Layout)

Full Screen Preview

In full-screen preview mode, you can also view image information and perform operations on the context menu. It supports dragging, resizing and expanding/collapsing .

https://user-images.githubusercontent.com/25872019/235327735-bfb50ea7-7682-4e50-b303-38159456e527.mp4

If you, like me, don't need to view the generation information, you can choose to simply minimize this panel, and all contextual operations will still be available.

Image comparison

Transfer files between different tab panes.

https://github.com/zanllp/sd-webui-infinite-image-browsing/assets/25872019/e631e3c3-1cbf-49bc-8577-f2963a6c9e4d

Right-click menu

You can also trigger it by hovering your mouse over the icon in the top right corner.

Walk mode

https://user-images.githubusercontent.com/25872019/230768207-daab786b-d4ab-489f-ba6a-e9656bd530b8.mp4

Dark mode

Smart Organize

AI-powered automatic file organization - groups similar images and creates meaningful folders.

https://github.com/user-attachments/assets/c1279556-d255-4e71-b230-48523a4859bf

Natural Language Categorization & Search (Experimental)

This feature groups images by semantic similarity of prompts and supports natural-language retrieval (similar to the retrieval stage in RAG). It’s experimental: results depend on the embedding/chat models and the quality of prompt metadata.

How to Use (for end users)

1. Open “Natural Language Categorization & Search (Experimental)” from the startup page
2. Click Scope and select one or more folders (from QuickMovePaths)
3. Categorize: click Refresh to generate topic cards for the selected scope
4. Search: type a natural-language query and click Search (auto-opens the result grid)

The selected scope is persisted in backend KV: app_fe_setting["topic_search_scope"]. Next time it will auto-restore and auto-refresh once.

API Endpoints

• Build/refresh embeddings: POST /infinite_image_browsing/db/build_iib_output_embeddings
  • Request: folder, model, force, batch_size, max_chars
• Cluster (categorize): POST /infinite_image_browsing/db/cluster_iib_output_job_start then poll GET /infinite_image_browsing/db/cluster_iib_output_job_status?job_id=...
  • Request: folder_paths (required, array), threshold, min_cluster_size, force_embed, title_model, force_title, use_title_cache, assign_noise_threshold, lang
• Prompt retrieval (RAG-like): POST /infinite_image_browsing/db/search_iib_output_by_prompt
  • Request: query, folder_paths (required, array), top_k, min_score, ensure_embed, model, max_chars

How it Works (simple explanation)

• 1) Prompt extraction & normalization
  • Reads image.exif and keeps content before Negative prompt:
  • Optionally removes “boilerplate” terms (quality/photography parameters, etc.) to focus on topic semantics (IIB_PROMPT_NORMALIZE*)
• 2) Embeddings
  • Calls OpenAI-compatible /embeddings
  • Stores vectors in SQLite table image_embedding (incremental, to avoid repeated costs)
• 3) Clustering
  • Online centroid-sum clustering, plus a post-merge step for highly similar clusters
  • Optionally reassigns members of small clusters into the closest large cluster to reduce noise
• 4) Title generation (LLM)
  • Calls /chat/completions with tool/function calling to force structured JSON output
  • Stores titles/keywords in SQLite table topic_title_cache
• 5) Retrieval
  • Embeds the query and ranks images in the selected scope by cosine similarity, returning TopK

Caching & Incremental Updates

1) Embedding cache (image_embedding)

• Where: table image_embedding (keyed by image_id)
• Skip rule (incremental update): an image is skipped if:
  • same model
  • same text_hash
  • existing vec is present
• Re-vectorization cache key: text_hash = sha256(f"{normalize_version}:{prompt_text}")
  • prompt_text is the extracted + (optionally) normalized text used for embeddings
  • normalize_version is a code-derived fingerprint of normalization rules/mode (not user-configurable)
• Force rebuild: pass force=true to build_iib_output_embeddings or force_embed=true to cluster_iib_output_job_start

2) Title cache (topic_title_cache)

• Where: table topic_title_cache keyed by cluster_hash
• Hit rule: when use_title_cache=true and force_title=false, titles/keywords are reused
• Cache key (cluster_hash) includes:
  • member image IDs (sorted)
  • embedding model, threshold, min_cluster_size
  • title_model, output lang
  • normalization fingerprint (normalize_version) and mode
• Force title regeneration: force_title=true

Configuration (Environment Variables)

All calls use an OpenAI-compatible provider:

• OPENAI_BASE_URL: e.g. https://your-host/v1
• OPENAI_API_KEY: your API key
• EMBEDDING_MODEL: embeddings model used for clustering and semantic search
• AI_MODEL: default chat model (fallback)
• TOPIC_TITLE_MODEL: chat model used for cluster titles (falls back to AI_MODEL)
• IIB_PROMPT_NORMALIZE: 1/0 enable prompt normalization
• IIB_PROMPT_NORMALIZE_MODE: balanced (recommended) / theme_only

Optional: TwelveLabs Marengo embedding backend

Embeddings normally go to an OpenAI-compatible /embeddings endpoint. As an opt-in alternative you can embed prompt text with TwelveLabs Marengo, which embeds text/image/audio/video into one shared latent space:

• Set EMBEDDING_MODEL=marengo3.0.
• Set TWELVELABS_API_KEY to your TwelveLabs key (falls back to OPENAI_API_KEY if unset, for backward compatibility). OPENAI_BASE_URL is ignored for embeddings, and the OpenAI-compatible chat path used for cluster titles is unaffected.
• Install the optional dependency: pip install 'twelvelabs>=1.2.8'.

This is fully opt-in and non-breaking — the default OpenAI-compatible path is unchanged unless you select a Marengo model. You can grab a free API key (generous free tier) at https://twelvelabs.io.

Note: There is no mock fallback for AI calls. If the provider/model fails or returns invalid output, the API will return an error directly.