ComfyUI-RH-MiniMax-H3 Introduction
ComfyUI-RH-MiniMax-H3 adds video-and-audio generation nodes to ComfyUI. You can use a text description, keyframe images, or reference media to guide a generated clip. You can also provide a video and generate audio for it. The extension is based on the official MiniMax-H3 project.
The author has brought the model components into the ComfyUI process, so these workflows do not require a separate SGLang server or Diffusers pipeline. The interface centers on two generation nodes and three model loaders. That gives you a smaller set of nodes to work with while leaving options for references, faster generation, and memory management.
Start with an example workflow if you want to see how the nodes connect before building your own.
How ComfyUI-RH-MiniMax-H3 Works
Think of a workflow as a set of instructions and materials for a short production. Your prompt describes the scene; images, video, or audio can provide additional direction. The model loaders prepare the components that interpret those inputs and create the result.
- Choose what will guide the clip. Use text alone, add keyframe images, supply ordered references, or provide a source video, depending on the task.
- Load the model components. One loader selects the video-and-audio generation model, another loads the Qwen3-VL text encoder, and a third loads the video and audio VAEs. The text encoder helps interpret your instructions; the VAEs handle the model’s video and audio representations.
- Run a generation node.
RHMiniMaxH3VideoGenhandles text, keyframe, and source-video workflows.RHMiniMaxH3RefGenhandles workflows guided by references. - Use the results. Both generation nodes output
frames,audio, andav_latent. The frames and audio are the outputs you will usually use in an artwork workflow;av_latentis an intermediate output for workflows that need it.
For example, you might describe a rainy street in a prompt and generate video with audio. For a more directed result, you could add keyframe images. If you already have a video and need sound, use the video-to-audio workflow instead.
ComfyUI-RH-MiniMax-H3 Features
Generation Modes
- Text-to-video-and-audio (T2VA): Describe the clip in words and generate visual frames and audio.
- Keyframe-guided video-and-audio (FL2VA): Use images as visual guides alongside your instructions.
- Reference-guided video-and-audio (Ref2VA): Provide ordered image, video, or audio references. Audio-only references are also allowed, so an image is not required for every reference-based task.
- Video-to-audio (V2A): Supply an existing video when your main goal is to generate audio for it.
The type of material you provide changes what the model has to work from: a prompt leaves more of the visual direction open, while keyframes or references give it specific material to follow.
Focused Nodes
The two generation nodes cover the main creative tasks, while the three loaders select their supporting model components. The nodes appear under RunningHub/MiniMax H3 in ComfyUI.
Older, more granular nodes remain registered so existing workflows can still load and run. They are deprecated and hidden from normal node search, the node tree, and slot-drag suggestions. For new work, use the current generation and loader nodes.
Optional Turbo LoRA
A turbo LoRA is an optional model add-on intended for generation with fewer sampling steps. Select one in the model loader when its listed use matches your workflow; leave the LoRA selection empty to use the base converted weights. The available options include 4-step and 8-step variants, as well as a 4-step variant identified for 768p work.
Attention Backend
The loaders and generation nodes offer an attention_backend setting. This controls which supported attention implementation the extension uses; it is primarily a compatibility and performance choice, not a creative prompt setting.
| Setting | What it does |
|---|---|
auto | Follows ComfyUI’s current optimized attention setting. A sensible starting point when you do not need a specific backend. |
sdpa | Uses PyTorch SDPA. |
sage | Uses SageAttention when it is available. |
ck | Uses Comfy Kitchen INT8 attention, the same kernel selected by --use-ck-attention. This setting produces an error if the backend is unavailable. |
ComfyUI-RH-MiniMax-H3 Models
The extension uses a complete converted INT8 ConvRot model bundle. “INT8 ConvRot” identifies the converted weights used by this extension; it does not mean every file in the bundle is interchangeable. The extension automatically detects the complete converted bundle, and it also supports the legacy ComfyUI/models/MiniMax-H3 directory.
| Component | Role | When you use it |
|---|---|---|
MiniMax-H3-FL2VA-int8_convrot.safetensors | Main generation weights for the text/keyframe video-and-audio path. | T2VA, FL2VA, and the supported VideoGen workflows. |
MiniMax-H3-Ref2VA-int8_convrot.safetensors | Main generation weights for the reference-guided path. | Ref2VA workflows using image, video, or audio references. |
qwen3-vl-32b-int8_convrot.safetensors | Text-encoder weights. | Interpreting instructions used with the generation workflows. |
MiniMax-H3-video_vae.safetensors | Video VAE component. | Handling the model’s video representation. |
MiniMax-H3-audio_vae.safetensors | Audio VAE component. | Handling the model’s audio representation. |
The bundle also contains configuration files for the FL2VA and Ref2VA paths. Use the complete bundle rather than assuming a single weight file is sufficient. It is available from Hugging Face and its ModelScope mirror. The download is approximately 95 GiB, and the original documentation recommends at least 110 GiB of free disk space.
The optional turbo LoRAs serve different workflows:
| Turbo LoRA | Intended use |
|---|---|
minimax_h3_fl2v_turbo_4step_v0.1.safetensors | FL2VA or T2VA with 4-step turbo generation. |
minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors | FL2VA or T2VA with 4-step turbo generation for 768p. |
minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors | FL2VA or T2VA with 8-step turbo generation. |
minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors | Ref2VA with 4-step turbo generation. |
Choose a LoRA that matches the generation path shown in its intended use. No LoRA is required to run the base converted weights.
Troubleshooting ComfyUI-RH-MiniMax-H3
A model does not appear or cannot be loaded
- Check that you have the complete INT8 ConvRot bundle, including its configuration folders, rather than only one
.safetensorsfile. - Check that the bundle is under
ComfyUI/models/MiniMax-H3-INT8-CONVROT/, or under the supported legacyComfyUI/models/MiniMax-H3directory. - Confirm the download completed. The bundle is large, so an incomplete transfer may leave files missing.
- Restart ComfyUI and check the loaders again.
A reference video or its audio cannot be used
- Confirm that your workflow uses
RHMiniMaxH3RefGenfor reference-guided generation. - Check that
ffmpegandffprobeare available to the environment running ComfyUI; they are required for Ref2VA video and audio references. - Try the image/audio reference example to compare its node connections with your workflow.
The ck attention setting produces an error
- Change
attention_backendtoautoand try the workflow again. - If you specifically need
ck, check that the required Comfy Kitchen INT8 attention backend is available in your ComfyUI environment. Unlikeauto,ckerrors when its backend is unavailable.
An older workflow is difficult to edit
Existing workflows using legacy granular nodes can still load and run, but those nodes are hidden from normal discovery because they are deprecated. For a new workflow, start with one of the current examples. The author also provides a tools/migrate_workflow.py script for migrating older workflow JSON files.
Frequently Asked Questions
Do I need a separate SGLang server?
No. The extension runs its model components inside the ComfyUI process.
Can I use an audio reference without an image reference?
Yes. Ref2VA allows audio-only references.
Do I have to select a turbo LoRA?
No. Leave the LoRA selection empty to use the base converted weights.
What do the generation nodes return?
Both return frames, audio, and av_latent.
Learn More about ComfyUI-RH-MiniMax-H3
- Text-to-video-and-audio example workflow: A starting point for prompt-led generation.
- Video-to-audio example workflow: Shows the workflow for generating audio from a source video.
- Image/audio reference example workflow: Shows how to use references with Ref2VA.
- Official MiniMax-H3 project: Background on the project on which the extension is based.
- Hugging Face INT8 ConvRot bundle and ModelScope mirror: Sources for the converted model package.
The extension’s code is licensed under Apache-2.0. The model weights have separate upstream terms; review those terms before using them in your work. NOTICE.md contains attribution and third-party notices.
