RefMod Pilot

Hugging Face distribution of vavo/minimax-h3-refmod-pilot, snapshot 29abef7. Includes ComfyUI workflows, the photo-upload extension and the R01 robot reference. This is a RefMod/workflow package, not standalone MiniMax H3 model weights or a trained LoRA. GitHub is the source repository; future updates are not automatically mirrored.

Download the complete package · Browse files

Simple ComfyUI workflows for MiniMax H3 character references, by LoRA Pilot.

The inference canvas has four elements:

1. CHARACTERS: choose RefMods and their strengths, with + Add RefMod, Refresh RefMods and RefMod library controls.
2. EDIT SCENE HERE: prompt, aspect ratio, resolution, custom width/height, seconds, seed and steps.
3. OUTPUT: preview and save the video with audio.
4. READ ME: instructions beside the controls.

The original RefMod loader stays on the main canvas so its management buttons remain available. Model loading, reference encoding, sampling and decoding live inside the Scene subgraph. Open that subgraph to reach the advanced settings. RefMod Pilot provides workflows and a small ComfyUI frontend extension for photo uploads. No browser add-on is required.

Each element has its own color. The Markdown README includes a clickable LoRA Pilot banner loaded from lorapilot.com; the image needs internet access and adds no Python dependency. The original loader also exposes upstream advanced controls.

Workflows

Create a RefMod: upload your photos and save a reusable reference. The screenshot shows the Orange Robot creator with its upload control.

Generate a video: choose references, edit the scene and settings, and view the output.

Install

Install RefMod Pilot once to enable Upload photos and drag-and-drop:

1. Download this repository as a ZIP.
2. Extract its folder into ComfyUI/custom_nodes/ (the folder must contain __init__.py directly).
3. Restart ComfyUI and refresh its browser page.

This installs into ComfyUI, not Chrome or Firefox. The uploader uses standard browser file APIs, uploads to your ComfyUI server, and needs no extra Python packages. Its existing H3 dependencies below are still required. You can also install this repository by URL through a ComfyUI Manager version that supports Git URL installs.

Use a current ComfyUI with native MiniMax H3 support and subgraphs. Development targets ComfyUI 0.34.0 and frontend 1.49.6.

Install these custom-node packs through ComfyUI Manager, including their dependencies, then restart ComfyUI:

• ComfyUI-MiniMaxH3Mod: RefMod creation, loading and reference-aware text encoding. The inspected revision is f9462081e28794389b5a6c5067eb327412ad8ee7; upstream interfaces are evolving.
• ComfyUI-KJNodes: H3 memory-saving attention and feed-forward nodes used in inference. The creator does not need this pack.

RefMod Python dependencies are torch, safetensors, numpy and Pillow. Keep the CUDA-compatible PyTorch supplied by your working ComfyUI installation. The requirements file lists the remaining packages; it does not install ComfyUI, custom-node packs, or model weights.

If the packages are missing, run this with ComfyUI's Python:

python -m pip install -r requirements.txt

For the standard LoRA Pilot pod environment:

/opt/venvs/core/bin/python -m pip install -r requirements.txt

KJNodes has its own requirements, installed by its installer/Manager. Optional video-file inputs to the folder creator may require opencv-python or imageio; still-image creation does not need them.

Download these weights from Comfy-Org/MiniMax-H3 into the corresponding ComfyUI model folders:

The creator only needs the video VAE. The inference workflow needs all four files and saved RefMods under models/refmods/. Model licenses and requirements belong to their upstream projects.

Create a character

Open Character-RefMod-Creator.json in ComfyUI.

To create another orange robot reference, use Orange-Robot-RefMod-Creator.json. It starts with an unset photo folder and Full Reference at 768 px. A finished R01 example is included below; the source dataset ZIP is not bundled.

1. Click Upload photos in the first node, or drop 1–8 selected photos onto its upload area. Wait for “photos ready”. Captions are unnecessary.
2. Enter a unique character name. Reusing a name replaces its saved RefMod.
3. Run. The file saves under ComfyUI/models/refmods/ (normally /workspace/models/refmods/ on LoRA Pilot).

Uploads accept PNG, JPG, WebP and BMP, up to 20 MB each. Each batch gets its own folder and preserves selection order; the workflow sets the folder and photo count automatically. ZIPs, HEIC and videos are not accepted by this photo uploader.

If an upload fails, the previous folder is restored. Retry the selection; partially uploaded files remain unused in ComfyUI/input/refmod-pilot/. Photo uploads do not start encoding or video generation.

Advanced users can still enter an existing server folder. That loader takes the first eight media files alphabetically; it does not rank image quality. Use still images with similar aspect ratios and keep faces inside the crop. The simple creator crops references to the first photo's aspect ratio and stacks them into one video-kind reference when using multiple photos.

Full Reference mode preserves the VAE reference rather than compressing it to a small grid. No LoRA or H3 model weights are trained. More reference photos increase the cost of subsequent video generation.

Try the orange robot

1. Download lpr01robot_r01_v1.safetensors (649 KiB) and put it in ComfyUI/models/refmods/.
2. Open Orange-Robot-Demo.json, then click Refresh RefMods.
3. Keep the first slot set to lpr01robot_r01_v1, edit the scene, and Run.

This example uses six views from the user-provided r01-eight-image-dataset-v1.zip (lpr01robot): front, both front angles, side, back and waving. The redundant rear angle and cropped close-up were omitted. The original eight-image dataset is unchanged. Source hashes and settings record the selection.

Full Reference at 768 px produces one stacked video-kind reference with six latent frames and 3,456 tokens. Use <Video 1> when it is the only selected RefMod. Creation completed in 5.7 seconds on an RTX PRO 6000 Blackwell 48 GB MIG instance; this is a single observed encoding time, not a general performance guarantee. No model weights were trained. Video likeness and motion remain untested.

The supplied images are synthetic reference-guided R01 views. Minor proportion, surface and perspective differences remain; they do not establish exact 3D consistency.

Generate a video

Open RefMod-Pilot-Inference.json.

1. Select your RefMods in CHARACTERS. Public downloads start with three empty slots. Use + Add RefMod for additional slots (up to eight).
2. Update the character tags and scene in EDIT SCENE HERE. Choose size and seconds.
3. Run. Results appear in OUTPUT and save under ComfyUI/output/video/RefModPilot* (the pod may map output elsewhere).

Size and duration

• Aspect ratio + Resolution (megapixels) calculate dimensions in preset mode. Square, landscape, portrait and ultrawide presets are available. Start at 0.4 MP.
• Custom size switches to the Width and Height fields; aspect ratio and megapixels are then ignored. Dimensions round to multiples of 32.
• Duration is entered in seconds, limited to 5–15 seconds, and rounded upward onto H3's frame grid at 24 fps. Five seconds produces 124 frames (~5.17 s), ten produces 243 (~10.13 s), and fifteen produces 362 (~15.08 s).
• Seed repeats a take; Steps defaults to 20. Increasing resolution, reference count or duration can substantially increase memory and runtime.

Reference labels matter

RefMod files may contain one picture, several independent pictures, or a stacked video reference. The three slot numbers are not necessarily the three picture numbers.

Open the Scene subgraph and inspect Reference map after encoding. Use its <Picture n> and <Video n> labels in your prompt. For example, three multi-photo files from the supplied creator normally use <Video 1>, <Video 2> and <Video 3>. Separate-image bundles use picture labels for each view. The example prompt's picture labels are placeholders to replace for your files.

Validation status

Development preview. The original generation path passed live text/vision encoding with three bundles (24 independent images). This does not establish video likeness or identity separation. The compact subgraph workflow's validation results are recorded in VALIDATION.md.

Credits

Built on ComfyUI, ComfyUI-MiniMaxH3Mod by Luisacaotica, and ComfyUI-KJNodes. RefMod Pilot supplies workflow configuration and usability guidance; the RefMod implementation comes from upstream. No upstream Python implementation or model weights are vendored here.

Part of Lora Pilot family

www.lorapilot.com · YouTube: Stable Diffusion for Masses