Structural (per-frame) conditioning for H3 — supported, planned, or out of scope?
I'd like to know whether H3 has, or is planned to have, a conditioning path for a per-frame
structural signal — a depth / clay / pose sequence aligned to the target frames — as opposed to
the conditioning inputs I can find documented.
From the model card, the five modes (T2VA, I2VA, FL2VA, L2VA, Ref2VA) all appear to condition on
text plus first frame / last frame / reference images. I don't see a per-frame control path
there. If I have missed one, pointing me at it is a complete answer.
The reason I'm asking: the workflow I want is a 3D animatic driving the video — render a
clay/white-model sequence, feed it frame-aligned, and let the model handle texture and lighting
only. That works well precisely because the geometry and the timing are already given: there is
nothing for the model to invent, so faces and motion stop drifting. It is the one approach I have
seen produce consistently clean results.
So the practical question is which way to go:
- If a per-frame control path exists or is planned for H3, I will build on H3.
- If not, then this is a model-level capability rather than a node-level one, and I should use a
model that has it.
Either answer is useful to me, and I would rather ask than guess, because it changes the whole
toolchain downstream.