ComfyUI Node
H3 Rewrite Prompt
Builds the chat prompt the MiniMax-H3 prompt-rewriter LoRA was trained on — system prompt, task header, and reference-frame placeholders — plus the reference images batched in placeholder order. Feed both into Generate Text on a Qwen3-VL-8B encoder carrying the rewriter LoRA. ref2va is not supported by the adapter.
H3 Rewrite Prompt
- first_frame
- last_frame
- prompt
- reference_images
◄prompt►
◄taskt2va►
◄resolution16:9►
◄duration10►
◄system_prompt►
CategoryNynxz/H3/Prompt
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| prompt | STRING | The short request to rewrite. | |
| task | COMBO | t2va | t2va: text only. i2va: first frame. l2va: last frame. fl2va: both. |
| resolution | COMBO | 16:9 | Target aspect. Upstream defaults to 16:9 for t2va and adaptive for the image-conditioned tasks. |
| duration | INT | 104–15 | Target duration in seconds. The rewriter paces the shot timeline to fit it. |
| first_frameopt | IMAGE | Required by i2va and fl2va. | |
| last_frameopt | IMAGE | Required by l2va and fl2va. | |
| system_promptopt | STRING | Overrides the trained system prompt. Leave empty — the adapter was trained on the built-in one, and replacing it moves the rewriter off the format it learned. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| prompt | STRING | Already chat-templated, so Generate Text's use_default_template has no effect on it. |
| reference_images | IMAGE | In placeholder order. None for t2va — leave it unwired. |