The decision table
| Your input | Checkpoint | Task |
|---|---|---|
| Prompt only | FL2VA | T2VA |
| Opening frame | FL2VA | FL2VA |
| Ending frame | FL2VA | FL2VA |
| Opening and ending frames | FL2VA | FL2VA |
| Character or product reference images | Ref2VA | Ref2VA |
| Motion or camera reference video | Ref2VA | Ref2VA / V2V |
| Voice or sound reference plus visual input | Ref2VA | Ref2VA |
A keyframe is not the same as a reference
A first or last frame owns a temporal endpoint. H3 must begin at that composition, end at it, or build the motion between two supplied states. A reference image instead contributes identity, object, setting, palette, or another declared property without necessarily appearing as frame zero.
Ref2VA input limits
More references are not automatically better. One identity owner, one environment owner, and one motion or voice owner are easier to diagnose than several files competing for the same property.
- Up to 9 images.
- Up to 3 video clips, each 2 to 15 seconds, with no more than 15 seconds total video.
- Up to 3 audio clips, each 2 to 15 seconds, with no more than 15 seconds total audio.
- Audio cannot be the only reference type.
- No more than 12 files across all reference types.
Deployment consequences
The checkpoint families have separate transformer weights. In SGLang, serve FL2VA and Ref2VA with different model-variant values and usually different ports. In ComfyUI, the T2V and I2V templates load the FL2VA diffusion model while the R2V template loads the Ref2VA model.
- Download only the family needed by a dedicated service.
- Keep route names aligned with the loaded variant.
- Reject reference requests at an FL2VA endpoint instead of silently ignoring materials.
- Test every conditioning path with an input whose effect is easy to verify.