MiniMax H3 Prompt Rewriter turns a short idea into a structured video-and-audio prompt before the video model runs. The ComfyUI node pack offers several routes, with different language models, reference inputs and memory requirements. Choosing the right route matters more than downloading the largest file.
This guide reviews version 0.29.0, released October 13, 2026. The instructions follow the author's documentation and model cards; we have not installed the pack or measured its performance. H3Vid's online prompt generator is a separate tool and does not run these local adapters.
Choose the model for your actual inputs
| Route | Required base | Useful when |
|---|---|---|
| 27B Rewriter | Qwen3.6-27B | You want the trained text-to-audio-video rewrite |
| 8B Rewriter | Qwen3-VL-8B-Instruct | Your prompt depends on first or last pictures |
| Omni Rewriter | Qwen2.5-Omni-7B | Pictures, video and sound should inform the rewrite |
| Writer | A supported instruction-following GGUF | You prefer a smaller model using writing instructions |
The original 27B adapter is trained for T2VA. It expands action, shot timing, camera direction and sound into three named sections: integrated_multimodal_description, overall_soundscape and non_diegetic_music. It consumes text. Describing a reference photograph to it does not make it inspect that photograph.
The 8B adapter adds visual inputs for T2VA, I2VA, L2VA and FL2VA. Use first_frame for I2VA, last_frame for L2VA, and both for FL2VA. Its model card uses AV spellings for the same task families. Keep the adapter attached to the specified base and size; an arbitrary Qwen checkpoint is not interchangeable.
The Omni GGUF card explains its additional Ref2AV route. It can read pictures, clips and audio, and writes six fields for full-reference work: subject definitions, summary, retention analysis, detailed description, soundscape and music. The GGUF repository converts an existing adapter; it is not a separately trained H3 video model.
Plan memory before loading video weights
The node author's approximate smallest setups are about 13 GB VRAM for a 27B GGUF route, about 9 GB for the smaller multimodal routes, and about 5 GB for a small Writer. These are configuration estimates, not minimum guarantees for an entire H3 generation graph. Quantization, context length, offloaded layers and other resident models all affect the result.
The 27B safetensors route has a much larger download than its compact GGUF options. Do not substitute H3's existing text encoder to save that download: it is a different, truncated encoder whose job is to produce conditioning, not generate rewritten text. The adapter expects the language-model architecture it was trained against.
Treat prompt writing and video sampling as separate memory stages. Version 0.29.0's changelog adds an Unload model button and connects the pack's retained models to ComfyUI's unload-all behavior. This gives you an explicit way to release prompt-model memory before loading the video checkpoint.
Install and begin with a text-only template
Install through ComfyUI-Manager's registry entry, or clone the repository into ComfyUI/custom_nodes. For a reproducible documentation baseline, select tag v0.29.0. Use the same Python environment that launches ComfyUI when following its dependency instructions, then restart the application.
Open Workflow → Browse Templates and select this pack. Start with the simple writing template to confirm the prompt stage works before adding the much larger video-generation branch. The supplied prompt-to-video and references-to-video templates load H3 weights; the other text-oriented examples can stop at a completed prompt.
Choose a model entry appropriate to the route, enter a short scene idea, and set duration and aspect ratio to match the intended generation. For example, a ten-second shot of a cyclist stopping at a quiet lakeside calls for different pacing than a fifteen-second sequence with several cuts. These controls describe the requested composition; they do not extend the downstream generator's supported duration or resolutions.
Inspect the full rewritten_prompt before connecting it to the generator. Confirm that the requested action fits the available time, speech belongs to the correct subject, ambient sound belongs in the soundscape, and any musical score is deliberate. A well-formed response can still misinterpret your scene.
Keep references and labels in the same order
For a frame task, connect only the pictures that the selected task expects. For Omni Ref2VA, review the enabled reference strip and its ordering. A prompt naming Picture 1 must reach a generator whose first picture is that same asset.
The pack's Reference Slots node carries the enabled references downstream in the order used by the writing stage. Use that path rather than manually rebuilding a different image order after writing. Text-only routes can use captioning nodes to turn media into descriptions, but a caption remains an intermediate interpretation. It can omit a costume detail or confuse a sound source.
This is especially relevant to reference-to-video generation, where the same picture can serve as an identity reference rather than a required opening frame. Do not choose I2VA merely because the workflow includes an image; choose it when that image should define the first frame.
Reuse a prompt without silently freezing an old idea
repeat_last returns the last answer stored for that node. That is useful while adjusting the downstream video graph, but it can also return old text after you edit the idea. Check the node's status and disable the switch when you actually want a new rewrite.
The store lasts for the current ComfyUI session. It is not saved inside the workflow and a restart clears it. Save the final prompt through the library or as text if it must travel with the project. In 0.29.0, repeat_last with keep_model_loaded off also releases retained models, so reusing a prompt need not keep its language model occupying the GPU.
Your first useful milestone is a reviewed prompt with matching duration, task and reference order. Only then connect the generation branch and evaluate a short clip. Keep model revisions and the final text beside that result so later changes can be explained.
The cover is an AI-generated concept illustration, not a ComfyUI screenshot or MiniMax H3 output.