⚠️ This is a beta version, and more features will be added as the community comes up with useful ideas and improvements.
⚠️ Warning: Although this is not an especially advanced workflow, It uses custom nodes and, depending on the setup, may require basic knowledge of Python dependency management, virtual environments, the ComfyUI file system, and troubleshooting. If you are unfamiliar with these topics, please make a backup of your ComfyUI setup before installing new nodes or updating ComfyUI.
For beginners, I generally recommend using the portable version, as it is easier to maintain and back up. It is also important to have a solid understanding of how to properly use and prompt the MiniMax H3 model.
📝 Sage Attention nodes are not included. You can enable Sage Attention globally by adding the --use-sage-attention flag to your startup script.
💡You can find the system prompt I use to generate my MiniMax prompts here.
The Prompt Enhancer is not included because I haven't been able to find a ComfyUI-compatible model powerful enough to consistently handle the system prompt. I personally use Qwen3.6 running on Ollama.
⚠️ This workflow is built with the standard ComfUI layout system in mind, Nodes 2.0 will mess up the size of the nodes and front end functions like the bypasser switches, we are all grown ups here, let the kids play with toys 🤭
Main Features
Anything-to-video: Text-to-video, image-to-video, audio + image-to-video, video-to-video, and more.
Turbo mode switch
Upscaling and frame interpolation
Character identity lock
Lip-sync
Motion transfer:V2V / body performance
Camera movement / cinematography transfer
Voice cloning
Video editing
Multi-reference fusion: audio, video, and images
Native stereo audio generation: dialogue, SFX, and music generated jointly
First / last frame conditioning
Text and brand rendering
Video extension / continuation
Style transfer
Relighting / object swap / background replacement: as an editing subset
Multi-shot / multi-scene storytelling
Multilingual dialogue
📝 Personal Notes
The recommended video length is 20 seconds maximum. Beyond that, results can become inconsistent, although I've managed to successfully push the length up to 30 seconds on an RTX 4090.
For voice cloning, use reference audio around 5 seconds long. Longer references may cause the reference audio itself to leak into the final video.
The sample videos were generated using the LTX upscaler included in the workflow, along with a high-quality first-frame image, to achieve the best possible output quality.
Description
First version release
FAQ
Comments (15)
https://github.com/Tulut114/tulut_comfyui_Gadget
I have a simple node that can work with an LLMAcpp backend, but I don't recommend integrating it unless you have enough RAM or VRAM, otherwise it will definitely cause an OutOfMemoryError (OOM). You can use a PreviewAny node to receive the output and then turn off LLM.
I also got a node, but it offloads the LLM model from memory before continuing the execution, by sending an empty message and a "Keep alive" value of 0 to the ollama server.
@LatentHeart Because I originally wrote this node just for my own use. When you said that the plug-in could not be found, you can see that it had been uploaded to the warehouse when I sent you a message. My own main computer has enough capacity, and so do other devices, so I haven't considered the issue of oom. However, I have upgraded. I integrated llamacpp's path directly into the node, added low VRAM mode and triple insurance. Now it will pull up the backend on its own and kill the backend after optimization is complete. I'll update when I feel motivated to get out of bed😂😴
It is a very nice workflow. Thank you!
One question: when I try i2v, the video first frame does not completely match with the image I uploaded. Why is that?
Are you using the official prompting format. H3 requires a detailed and very specific prompting format to get the most out of it. That being said I’ve had this happen even using the proper Prompt format
@kannime I start the prompt with this:
"Reference: Use the uploaded image as the identity reference. The shot begins from <Picture 1>. Preserve her exact facial features, expression, hairstyle, skin texture, body proportions, clothing, camera angle, room layout, lighting, and background. Do not change her face or identity at any point."
Cool, I like it. The only thing I missed was a node for adding loras, so I just added one after the turbo-lora-node, it didn't break but you can probably add one more "proper" than I did.
Yeah I will add a Power lora loder when more LoRAs come out ;)
Loved your LTX 2.3 workflow and this one looks promising as well! Thanks
Thanks :)
ComfyUI refuses to install ComfyUI-LTXVideo and I can't make this work...
I had the same problem, probably an issue with newer comfyui version and outdated(?) pack... I replaced the vaedecode with another tiled decode and it works
i adopted your idea in my workflow of using LTX upscale model to do the upscale while using the Minimax H3 for initial generation. Thanks for your workflow. it really well organized . Nicely done. Once i need to use the ref2VA, i will definitely try this... meanwhile i usually use FL2VA.. my display card can't handle ref2VA too much.