H3 Film Studio — Film Directions + 9 Retouch Engines

https://github.com/user-attachments/assets/386e3683-55e0-4e41-b825-72323731e8c0

Community release · Windows / NVIDIA / ComfyUI · Provided AS IS

This is my personal filmmaking and video-retouching Studio, shared so other creators can explore it, learn from it, and adapt it. It is not a one-click installer or a universal ComfyUI distribution. The simplest way to use it is to reproduce the setup below as closely as possible. Different hardware, checkpoints and custom-node versions may require manual changes.

What is it?

A local, browser-based Studio with two sections in one interface:

• Film Directions: AI-assisted film development (content, scenes, shots and clip prompts), references for characters/sets/audio, MiniMax H3 RAW video generation, motion context for successive clips, and a separate HD refine / continuity / assembly pipeline.
• Retouches: a separate non-destructive video editing workspace with nine visible editing modes: remove/replace people and objects, change the camera or background, edit a local region, make a precise retake, or insert a new object. Some features depend on additional models, LoRAs, masking or geometry nodes.
• Reference generation: Krea2-based image/reference workflows, optionally with SeedVR2 and UltraSharp upscaling.
• Local LLMs: Ollama for film directors and some retouch prompts; optional llama.cpp variants for supported directing configurations.

The project's original configuration uses a Singularity MiniMax H3 Ref2VA INT8 checkpoint and an H3 workflow with separate RAW and HD phases. A modified model is not guaranteed to work with exactly the same encoder, VAE, LoRA, input schema, sampler or latent format.

Nothing here requires a cloud subscription to the Studio itself; however, downloading large models and running video generation requires considerable disk space, memory and a compatible GPU.

---

Studio tour — what you can actually do

H3 Film Studio combines film planning, reference management, video generation and editing inside a locally hosted browser interface. The two main workspaces serve different purposes: Film Directions creates and organizes a film, while Retouches modifies existing video clips. Both rely on the ComfyUI and local AI models installed on your computer.

Film Directions — from an idea to an assembled film

1. Start or resume a film project. Describe the premise, style and creative intention. The project keeps track of its scenes, shots, clips and production decisions so you can pick up where you left off.
2. Direct the AI Directors. The local LLM-assisted pipeline proposes content, scenes, shots and generation prompts. Review the proposals, request changes and validate decisions instead of treating the first generated script as final. The Studio separates the wider story, individual scenes, the shot plan and the exact H3 clip prompt.
3. Build a reference bank (RefBank). Keep visual references for characters, places, objects and appearance, together with audio-related references where supported. Import your own material or use the included Krea2 reference-generation and image-editing workflows. Available reference-upscale paths include SeedVR2 and UltraSharp where installed.
4. Generate a video clip as H3 RAW. Film Directions sends the selected prompt and references to ComfyUI and records the generated RAW MP4; its native RAW latent can also be saved for the later HD pipeline. The original setup is based on MiniMax H3 / Singularity and short clip-by-clip production, rather than generating an entire movie in one pass.
5. Maintain continuity between clips. The workflow can use preceding motion context and the project's continuity data so a following clip can connect to what came before. This is a continuity mechanism, not a guarantee of seamless identity or motion in every output.
6. Refine the selected clip in HD. The supplied workflow has a separate RAW → latent upscaling → HardLock → HD refine → visible output path, using its project-specific nodes. The original setup targets 1920-wide video. Keep the exact model, VAE and workflow configuration if you want to reproduce the original behavior.
7. Build the cumulative FINAL. Once a clip is processed, the assembly path joins it with previously approved output to extend the film. The interface provides a timeline, clip-specific prompts/references, RAW and HD viewing, rendering status, production journal and the latest cumulative FINAL preview.
8. Send a RAW clip to Retouches when needed. Select a clip and open the Retouches workspace for a targeted alternative take. Review the resulting MP4 before approving it; the original film source is kept separately.

What the interface contains: a Film Director decision desk, project management, screenplay/shot views, RefBank, active prompt and reference display, a clip viewer, RAW/HD controls, a timeline, job progress, and a cumulative FINAL preview.

Technical workflow, simplified:

Film idea -> AI Directors -> Scenes -> Shot plan -> H3 clip prompt
                      |                  |
                      +-> RefBank (Krea / imported references)
                                         |
                          H3 RAW MP4 + saved RAW latent
                                         |
                         Motion Context for the next clip
                                         |
                         Latent upscale -> HardLock -> HD refine
                                         |
                              Visible clip -> cumulative FINAL

The files already supplied for this route include workflows/H3Studio_1920_API_CLEAN.json and the three workflows/Krea_reference*.json variants. They are workflows, not the model weights or custom-node implementations. See the installation sections below for those dependencies.

Retouches — edit an existing clip without replacing the original

Retouches is a separate video-editing workspace. Choose a completed project clip or import an MP4, inspect it in the player, select a mode, describe your change in natural language and optionally point to a subject or a time range. Where relevant, use an image from RefBank or import a new appearance reference. The Studio prepares a ComfyUI job, lets you review the new MP4 and keeps the original video as a separate file.

The nine modes visible in the V3.1 Retouches interface are:

Additional RAW-engine code: the source also contains Prolongation (extend) and Performance processing paths for action, dialogue and emotional changes. Their availability depends on the version and entry points being used; they are not counted as extra buttons in the nine-mode V3.1 grid. The distinction matters when comparing screenshots, older versions and source code.

Typical Retouches workflow:

Project RAW clip or imported MP4
                |
     Select one of 9 editing modes
                |
     Describe the change in plain language
     (optional: image / region / time range)
                |
         ComfyUI retouch render
                |
      New MP4 -> review -> approve / retry / continue editing
                |
          Original clip remains intact

Conveniences: optional mask previews before an expensive generation, quick short tests on supported modes, independent render history, and a review flow for chaining a further edit from an approved version. Specific retouches need additional third-party nodes, LoRAs or models; check the tables below before launching a mode.

Important V3.1 handoff limitation: approving a retouched MP4 and returning to Film Directions does not necessarily replace the original RAW latent used by the HD refine workflow. The interface itself warns that this conversion is a separate unfinished step. Do not assume a retouched MP4 will automatically become the exact latent source for HardLock/HD. Some retouch → HD/FINAL operations also have known audio or performance issues; inspect the actual outputs before relying on them.

Screenshots and demo video

GitHub can display screenshots and an uploaded MP4 directly in this README. This release has no demo media embedded yet: upload your own screenshot or MP4 as a GitHub attachment, then paste its generated https://github.com/user-attachments/assets/... URL into the following template. No local Windows path will work as a public URL.

---

Quick start — follow this order

The following is the easiest route for the original Windows configuration. It does not promise identical results on other installations.

1. Install NVIDIA drivers and ComfyUI Windows Portable for NVIDIA. H3 needs a recent ComfyUI with native MiniMax H3 support; the Studio workflow was authored around ComfyUI 0.30.0+, but its third-party nodes may require a newer build.
2. Extract ComfyUI so the following paths exist (these are the paths expected by the unmodified Studio):
   • C:\AI\ComfyUI_windows_portable\ComfyUI\main.py
   • C:\AI\ComfyUI_windows_portable\python_embeded\python.exe
3. Download/extract the Studio release into a separate directory (for example C:\H3studioFILM_V3_1). Do not put the Studio at the root of ComfyUI.
4. Install the custom nodes: copy the two included custom_nodes folders into ComfyUI and install the external extensions listed below.
5. Download the models and LoRAs for the features you want to use. Place each file in the exact ComfyUI model folder shown. Do not rename the original H3 checkpoints without also updating references in workflows/code.
6. Install Ollama and pull the text model(s) used by your chosen Director/Retouch mode. Some existing retouch flows specifically expect gemma4:26b-a4b-it-q4_K_M or another installed gemma4:26b* Ollama tag. Check your locally available model tags with ollama list.
7. Make FFmpeg available in PATH (or ensure the embedded ComfyUI Python has imageio-ffmpeg). Install Git for Windows if you will clone nodes.
8. Start ComfyUI yourself first, for example by running the portable package's run_nvidia_gpu.bat. Wait until http://127.0.0.1:8188/object_info responds. The Studio launcher does not start, stop or restart ComfyUI.
9. From the Studio release directory, double-click LANCER-STUDIO-UNIFIE-7861.cmd. Open http://127.0.0.1:7861/ for Film Directions and Retouches.

First launch on a fresh ComfyUI setup: The launcher may synchronize its existing H3StudioRetouchExtras files and stop intentionally, asking you to restart ComfyUI. This is expected: close and restart only ComfyUI after the GPU queue is empty, then launch the Studio again. It must not hot-reload custom Python nodes while rendering.

PowerShell launch example

# Example: adapt the ZIP filename to the downloaded release.
$zip  = "$env:USERPROFILE\Downloads\H3studioFILM_V3_1_CONTINUATION_SPLIT_FIX.zip"
$dest = "$env:USERPROFILE\Downloads\H3studioFILM_V3_1_public"
Expand-Archive -LiteralPath $zip -DestinationPath $dest -Force
Set-Location $dest

# ComfyUI must already be running on port 8188.
.\LANCER-STUDIO-UNIFIE-7861.cmd

Do not extract over an older Studio installation if you want to keep its original files, projects and configurations. Choose a distinct directory.

Required software

The code imports packages such as torch, numpy, Pillow, opencv-python, av and imageio-ffmpeg depending on the selected feature. Install missing packages in ComfyUI's embedded Python, not an unrelated system Python, after following the individual extension requirements. For example:

& "C:\AI\ComfyUI_windows_portable\python_embeded\python.exe" -m pip show torch numpy pillow opencv-python av imageio-ffmpeg

This checks availability; it does not install or upgrade anything.

---

Custom nodes & extensions

Custom-node folders belong under:

C:\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes\

Already inside the current Studio archive

Copy the directories, not just their individual Python files:

In File Explorer, copy each folder into ComfyUI custom_nodes. The unified launcher only auto-synchronizes the second group; copying both groups before starting ComfyUI is the clearest first-install method.

Install external nodes

You can install through ComfyUI Manager where available, or clone the upstream projects. Prefer the versions known to work in your own ComfyUI; upstream updates can change input schemas.

Manual examples from the ComfyUI custom_nodes directory:

Set-Location "C:\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes"
git clone https://github.com/shootthesound/ComfyUI-H3Studio.git
git clone https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler.git
git clone https://github.com/drozbay/MaskVidExperiments.git
git clone https://github.com/scraed/LanPaint.git
git clone https://github.com/cseti007/ComfyUI-CrossViewWarp.git
git clone https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler.git

If an extension already exists, do not clone it again. Verify its version instead. Read its README and install any Python requirements with the ComfyUI embedded Python, then restart ComfyUI when idle.

Check whether ComfyUI actually loaded the nodes

With ComfyUI running, open:

http://127.0.0.1:8188/object_info

In the Studio directory, you can also run:

.\VERIFIER-NODES-COMFYUI-V10_97.ps1

A Python file existing on disk does not prove the corresponding class loaded successfully. Read ComfyUI's terminal import errors and restart after node installations.

---

Models to download — exact filenames

All AI weights are excluded from the original archive. Download the files you need separately. The tables show original workflow filenames; a different checkpoint often requires changes to the Studio's hardcoded references and may require a different sampler, conditioning or VAE.

Paths below are relative to C:\AI\ComfyUI_windows_portable\ComfyUI\models\. If ComfyUI uses extra_model_paths.yaml, place them in the corresponding configured folders instead.

A. Main Film Directions / H3 RAW → HD pipeline

Important: The H3 workflow contains an older metadata reference to minimax_h3_ref2va_pruned_nvfp4.safetensors, but its actual UNETLoader widget selects Singularity v1.3 INT8. Follow the active selection, not that stale metadata. This is an 8-step RAW / HD-refine project-specific workflow; the Turbo LoRA name does not mean every operation uses four steps.

B. Other H3 model variants used by some retouch routes

These variants are not drop-in replacements for every feature. In particular, a Ref2VA and an FL2VA graph are not necessarily identical.

C. Krea2 + image references

Krea2 reference upscale (only if using those nodes):

D. SAM3.1 + retouch LoRAs

Version note: The Character Swap author currently also offers a newer LoRA. The Studio's existing code explicitly references the older h3_character_swap_pro4500_1000.safetensors file; using the new one without editing its references is not equivalent.

E. Geometry / MoGe2 for camera-change workflows

Camera-change graphs also need MoGe-compatible inference (LoadMoGeModel / MoGeInference or equivalent compatible nodes) and a MoGe2 ViT-L model. Start with the CrossViewWarp documentation and the MoGe project. The original ZIP does not contain the exact MoGe2 weight file or a verified version/URL, so this is an integration-specific download, not a guaranteed one-file instruction. Check the node's model picker rather than inventing a filename.

F. Local language models

• Install Ollama, then check the Ollama model library for models available for your installation.
• Film Directions lists names such as gemma4:12b, gemma4:26b-a4b-it-q4_K_M, qwen3.5:9b and others. Pull the exact available tag your chosen backend expects, e.g. ollama pull <valid-model-tag>.
• Some Retouches flows require Gemma 4 26B (and do not silently fall back to a Qwen model). A smaller model selected in Film Directions does not automatically replace that retouch requirement.
• Two optional Qwen3.8 27B GGUF configurations use external llama-server executables and specific .gguf files; these are not bundled. Only configure them if you already understand those backends. See environment variables H3FILM_KIWI_ROOT, H3FILM_LLAMA_SETTINGS and H3FILM_LLAMA_PORT in the existing code.

Model licenses: Model files are hosted by their respective publishers, with their own licenses, access conditions and regional restrictions. Follow the license on every linked model page. This Studio does not redistribute H3, Krea, SAM or third-party LoRA weights.

---

Extra workflows

The archive already includes:

• workflows/H3Studio_1920_API_CLEAN.json
• workflows/Krea_reference.json
• workflows/Krea_reference_img2img.json
• workflows/Krea_reference_sref.json
• retouches_new_workflows/RAW_ENGINE_SOURCE_COPY.json

The following external retouch workflows are downloaded by the supplied installer rather than embedded in the original ZIP:

From the Studio release directory, the existing download script is:

.\INSTALLER-WORKFLOWS-RETOUCHES.ps1

You can instead download the JSON files manually and save them under retouches_workflows/ with the local filenames shown. A further LanPaint MiniMax H3 AV inpainting example is referenced by the code for some masked-video paths. This example is not a substitute for the main Film Directions RAW/HD workflow.

---

File structure

H3studioFILM_V3_1_public/
├── LANCER-STUDIO-UNIFIE-7861.cmd       # Recommended Studio launcher
├── Start-H3Studio-UNIFIE-7861.ps1      # PowerShell launcher; original ComfyUI path
├── launch_h3_unified.py               # Preflight + single 7861 web server
├── h3film_studio_v3.py                # Film Directions
├── studio_web.py                       # Studio web API
├── studio_ui/                         # Browser interface
├── prompts/ + director_teams/         # Directors / prompts
├── retouches_*.py + studio_retouches.py
├── workflows/                         # Included H3 + Krea JSON workflows
├── retouches_new_workflows/           # Included RAW source / manifest
├── retouches_workflows/               # Downloaded retouch JSONs (not in initial ZIP)
├── custom_nodes/
│   ├── H3RawLatentPersist/
│   └── H3StudioRetouchExtras/
└── tests/                             # Structural and regression tests

The ZIP also includes historical launch scripts and release notes. For normal use, prefer only LANCER-STUDIO-UNIFIE-7861.cmd; do not start separate old Retouches and Film servers in parallel.

Original network addresses and fixed configuration

Using a different folder or port? This release intentionally does not include an interactive setup wizard. Adapt the paths/URLs in the launch scripts and the corresponding Python modules; changing only the COMFYUI_DIR environment variable is not sufficient while the primary PowerShell launcher still uses fixed paths. The code is shared so you can make those changes yourself.

Changing models or workflows

You can substitute your own compatible models, LoRAs and workflows, but the Studio may reference exact filenames and node inputs in both Python and JSON. Inspect at least:

• workflows/H3Studio_1920_API_CLEAN.json and the three Krea workflows;
• h3_runtime_simple.py and h3film_studio_v3.py;
• retouches_engine.py, retouches_clean_plate.py, retouches_masked_core.py, retouches_crossview_core.py and retouches_roycho_complete_v63.py.

A new model may require a different text encoder, VAE, latent format, attention backend, sampler, LoRA or workflow graph. Don't assume that swapping a .safetensors filename is enough. If you want to change the Studio, try using your favorite AI coding assistant: share your errors, installed ComfyUI /object_info, workflows and the relevant source files, and ask it to help you adapt the project to your setup. Keep backups of the original working files.

---

Troubleshooting

Studio says “ComfyUI not responding on :8188” : Start ComfyUI first, confirm http://127.0.0.1:8188/object_info, then start the Studio. Do not start a second ComfyUI instance on the same port.

ComfyUI introuvable / missing python_embeded\python.exe : You did not extract the portable build to the original C:\AI\ComfyUI_windows_portable\ location. Either use that location or edit Start-H3Studio-UNIFIE-7861.ps1 to match your machine.

Nodes absents de ComfyUI EN MEMOIRE or unknown node class : Compare the exact missing class to the custom-node list, install it, examine ComfyUI import errors, and restart ComfyUI when there is no render in progress. The Studio may intentionally stop once after synchronizing node files. Remember that the V3.1 ZIP is missing a separate group of project-specific nodes until they are added to the release.

UNETLoader / CLIPLoader / VAELoader says model not found : Check the exact filename and ComfyUI model folder. Confirm the active workflow selects Singularity (not the old NVFP4 metadata). Restart or refresh ComfyUI if the model list was cached. Not all quantizations are interchangeable.

Ollama runs, but Director/Retouches fails : Run ollama list. An installed Ollama server is not the same as having the right model tag. Some retouch paths expect a Gemma 4 26B tag; they will not use Film Directions' smaller model automatically.

Port 7861 already in use : Close the old Studio server terminal, not the ComfyUI render process. The unified launcher intentionally refuses to create a duplicate Studio server.

Slow generation / out of VRAM : Video models and HD refinement use substantial VRAM and RAM. First check that your ComfyUI attention backend, extension versions, cache state and model/LoRA selection match the intended graph. The number of sampler steps may differ between RAW, Roycho/retake and HD refinement. Avoid blind CUDA/PyTorch upgrades.

RAW audio is correct but HD/FINAL audio differs or breaks : The V3.1 package is shared AS IS. The author has observed audio and slowdown issues in certain retouch → HD / assembly scenarios; this README does not claim these are fixed. Keep the original RAW MP4 and audio as backups, check each intermediate file, and only replace the output after verifying audio. Avoid destroying existing renders or cache files to “fix” a workflow.

Downloads from Hugging Face are missing, gated or no longer available : Open the linked model repository page, read its instructions and license, and pick the file matching the exact name used by the workflow. Some publishers require account acceptance or region-specific authorization. The project maintainer does not host the third-party weights.

---

AS IS, community use and credits

This is a personal project shared “AS IS”, without warranty of any kind, without guaranteed compatibility or technical support. The author is sharing the Studio, prompts, workflows, local scripts and redistributable code as generously as possible; installing third-party software and weights is the user's responsibility. This is not a turnkey commercial product, and no particular GPU, model, operating system or output quality is guaranteed.

Want to modify it? Please do! Use a different LLM, model, LoRA or ComfyUI workflow, and ask your favorite AI assistant to help adapt the code and diagnose errors. Share fixes or improvements back with the community if you like. The repository is provided as a starting point, not a promise to maintain everyone's installation.

Third-party licenses and model access

The Studio's source code and the external components are separate things. Each upstream project and model has its own licensing terms. Free download does not automatically grant permission to redistribute a file. Keep the original licenses/attributions for any upstream code bundled in a release; otherwise link to its source instead. Do not bundle third-party weights by default.

MiniMax H3 has a territory-restricted community license. At the time of writing, the standard community license excludes use in the EU, UK, US and South Korea without separate authorization. Check the official MiniMax H3 license and Q&A, and use the official license request process if applicable. Third-party H3 finetunes/LoRAs may inherit the same restrictions. Krea, SAM, third-party datasets and other weights can have different terms. Linking to a model is not a claim that every visitor may legally use it.

Credits / upstream projects used or referenced include:

• ComfyUI, ComfyUI-H3Studio, H3 Latent Upscaler
• MaskVidExperiments, LanPaint, ComfyUI-CrossViewWarp, SeedVR2
• MiniMax H3, Singularity, Krea-2, SAM3.1
• Akatz Person Remover, Akatz Character Swap, Cseti CrossView-Warp LoRA

Source license: The Studio repository must include an appropriate LICENSE file chosen by the author before it is advertised as open source under a specific license. This README does not select a license on the author's behalf, and that license cannot override upstream restrictions.

---

Made for creators who like to experiment. Take it, explore it, adapt it — and have fun making films.