Skip to content

MiniMax-H3 Ref2VA: expose reference video resolution (+ long-term: latent tail carry-over for chunk chaining) #1951

Description

@laurentvv

Hi, I generate long videos with the MiniMax-H3 Ref2VA model entirely through sd-cli -M vid_gen --ref-video <frames_dir> --ref-video-audio <wav> on Vulkan (AMD RX 6950 XT): the tail of chunk N (last frames + audio, extracted with ffmpeg) becomes the reference of chunk N+1 — the CLI equivalent of hradec's ComfyUI "HR Endless Sampler" node. It works well, thanks for the great H3 support!

Two things would make chaining cheaper/seamless:

1. Reference resolution control

As far as I can read in the source, ref-video frames are resized to a nominal size based on 768 px (aspect kept, capped 768x1344, aligned 32; the source size is kept only if smaller). There is no way to deliberately encode the reference below that nominal size.

ComfyUI's HR Endless Sampler has video_continuation_res (decode -> resize -> re-encode the continuation block) exactly to shrink the reference latent: less ref-attention VRAM, bigger chunks, faster VAE encode (on CPU backends this is significant — on my 16 GB VRAM / 32 GB RAM machine the CPU VAE encode of a 12-frame reference takes ~14.5 min).

Would you consider a --ref-video-res WxH (or a scale factor) that clamps the reference to a user-chosen size? (Going up from a small source is already possible implicitly; the ask is going down on purpose.)

2. Latent tail carry-over (harder, long-term)

Chaining currently re-encodes the reference each chunk (decode -> resize -> encode), so a tiny generation loss sits at every seam and can accumulate over very long chains. ComfyUI-H3-Motion-Context slices the previous chunk's tail in latent space (bit-exact, zero drift).

I understand this may not fit the CLI pipeline (the ref would have to skip the VAE encode stage entirely), but I'd like to put it on the radar — even a "reuse previous output latents as reference" hook would remove seams completely.

Happy to A/B test patches on this machine (Vulkan, 16 GB VRAM) — the single-chunk recipe is validated and documented on our side.

Environment: stable-diffusion.cpp master-841 win vulkan-x64, Windows 11, AMD RX 6950 XT 16 GB (RDNA2), model minimax_h3_ref2va_pruned-Q4_K_M.gguf.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions