OmniRef for H3: multi-reference conditioning, and why 8 steps was not the whole story
H3 V2 already compressed audio-video generation into 8 steps. That was an acceleration story - great for text-to-audio-video, less useful when your problem was "the character in this reference must appear in that shot". This episode looks at OmniRef, which attacks the other axis: multi-reference conditioning.
Watch the episode: https://youtu.be/RWMIgOFwAqY
What OmniRef adds
The 8-step V2 path made generation fast. OmniRef makes it reference-aware: instead of one conditioning image, you give it several, and the model reconciles them - identity from one, styling from another, scene or wardrobe from a third.
Why that is a different problem from speed
Speed is a single number. Reference consistency is a judgement: does the output still look like reference A while matching the lighting of reference B? Those are separate failures, and they show up at different moments in a render.
What to check when you test multi-reference
- Which reference wins when two of them disagree - identity or styling?
- Does an extra reference improve the result, or just make it muddier?
- At what count does adding references stop helping?
- Do you still get the 8-step speed with multiple references loaded?
Known limits
- More references means more ways to be wrong: conflicts between references surface as mush, not as a clear error.
- Reference quality dominates. A soft reference will drag the result soft regardless of how many others you stack.
- There is no downloadable workflow JSON for this episode. OmniRef is an online experience in this context; the article is about what to expect when you run it, not a package to install.
- Not a substitute for the 8-step acceleration path - they solve different problems and can be combined.
Where this leaves the pipeline
The useful sequence is: get speed with the 8-step path, then add references only where consistency actually breaks. Loading five references "just in case" makes the prompt harder to debug, not the output better. Start with two, find the failure mode, then add the one reference that fixes it.
Top comments (1)
Multi-reference conditioning is where serialized episode production really unlocks. Preserving wardrobe, prop, and lighting identity across scene angles matters even more than raw step count!