Minimax H3 Fast-Omni Studio - Reference - Image to video - Extend clip
🆕 Version 1.4.0 — Experimental LoRA Support
Don't forget to update the argon_omni_studio folder located within the custom_nodes directory of your ComfyUI folder. The easiest way is to delete the existing folder and replace it with this one.
You can now use up to 3 LoRAs directly in Omni Studio, each with:
A file selector.
An independent on/off switch.
Adjustable strength.
A field for trigger words, added to the generation prompt when the slot is active.
The new controls are available in French and English. All three slots are disabled by default, so you can keep using the workflow as usual.
⚠️ Please read before testing
This feature is experimental. Adding a LoRA loader does not make every LoRA compatible with FastH3 Omni. Style, motion and Turbo LoRAs may fail to load, have no visible effect, or produce unexpected results. A successful load does not guarantee correct results.
Turbo LoRAs do not automatically change the sampling settings or speed up generation. This update does not automatically adjust the step count or sampling schedule.
Compatibility with individual LoRAs has not been verified through GPU generation tests. You’re welcome to experiment, but you do so at your own risk. I can’t guarantee compatibility or take responsibility for errors, failed generations or unwanted results caused by third-party LoRAs.
Start with one LoRA at a time, and feel free to share which ones work for you, along with your settings! 🙂
Main Features
Dedicated Argon Omni Studio interface.
Full French / English interface switching, including workflow labels and studio controls.
Image-to-video generation with first-frame guidance.
Optional last-frame guidance.
Up to 9 image references: 3 characters, 3 locations and 3 props.
Up to 3 audio references.
French dialogue and speaker assignment.
Video continuation with previous-frame and optional motion/audio references.
Automatic saving of individual segments and assembled videos.
24 FPS output, with configurable duration and resolution.
INT8 model configuration with 8-step sampling.
Download (External site)
Location : models/diffusion_models https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/diffusion_models/fasth3_omniref_8step_pruned_int8_convrot.safetensors?utm_source=chatgpt.com
Location : models/text_encoders https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors?utm_source=chatgpt.com
Location : models/vae https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors?utm_source=chatgpt.com
Location : models/vae https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors?utm_source=chatgpt.com
Installation
Download and extract the ZIP.
Close ComfyUI.
Copy
argon_omni_studiointoComfyUI/custom_nodes/.Restart ComfyUI and refresh the browser with Ctrl + F5.
Import
Argon_FastH3_Omni_Studio.json.Install the required FastH3 Omni model files from the official FastVideo Hugging Face repository.
Open Omni Studio from the Argon Director node.
Important: Importing the JSON alone does not install the custom interface.
Recommended First Test
Start with one scene image, one character reference, 5 seconds and 1024 × 576 resolution.
Once generation works, test scene continuation and assembled video export.
Compatibility
Requires an up-to-date ComfyUI installation with native MiniMax H3 reference-to-video support.
The included configuration uses INT8 models. Memory requirements depend on the number of references, resolution, duration and GPU.
Included
Workflow JSON, custom Python nodes, JavaScript/CSS interface, French and English documentation, installation guides and validation information.
Model weights are not included.
Notes
Character references and scene guides improve continuity but do not guarantee pixel-perfect identity consistency.
The interface has undergone frontend and Python contract testing. A complete GPU generation test for this specific release has not been independently documented.
Credits
Independent ComfyUI workflow and studio interface built around the FastH3 Omni / MiniMax H3 ecosystem.
Model weights are provided separately by their respective creators and remain subject to their original licenses.
If you have any suggestions or ideas to improve this workflow, feel free to let me know! Your feedback is always welcome. 😊
If you enjoy my work and would like to support future updates, a small donation would be greatly appreciated! ❤️
☕ Support me: https://paypal.me/ein95charly
Thank you for your support!