Which MiniMax H3 workflow should I use for a dialogue shot?
Choose H3-Base FL2VA when the opening and closing compositions are fixed. Choose Ref2VA when an image, video or audio reference defines the character or performance. On the rental shelf, select MiniMax H3 (ComfyUI) and open the matching graph. Verify the checkpoint and inputs before timing a take. MiniMax H3 model card
First and last frames or reference material?
| Director’s instruction | Useful pilot | What to review |
|---|---|---|
| Start on the door and finish on the actor | FL2VA with your two approved frames | Composition at both ends, body motion through the cut |
| Keep the actor recognizable during dialogue | Ref2VA with the licensed performance references | Identity, mouth timing, voice consistency, background noise |
| Deliver vertical footage | Use references cropped to the final aspect ratio | Face and subtitle safe areas throughout the take |
How much time should a pilot allow?
CloudGPU previously measured 85–87 seconds of warm generation for a five-second 1344 × 768 clip on one RTX 4090 48 GB, using int8 files and the four-step turbo LoRA. On the documented 20-step configuration, the same duration took 352 seconds. Those timings exclude downloads and model loading. Record the first take and subsequent takes separately on your selected configuration. Test settings and GPU sizing
Export a clip with its audio track and play both before releasing the instance. Check the model licence for the intended distribution and keep the reference rights with the project. MiniMax H3 licence · Budget the episode