Feature Request: Expose last_image parameter and support multi-frame reference conditioning
Nobody has claimed this yet.
Assessment
- Difficulty
- 5/5
- Estimated time
- Over a week
- Newbie friendliness
- 35/100
Research direction
Start in pipeline_mova.py at prepare_latents (line 261) and the pipeline call entry point. Trace how call currently supplies the last_image argument, then verify the minimal FLF2V request is complete when that parameter reaches prepare_latents; the broader multi-frame conditioning work requires resolving the audio-video bridge and segment-conditioning questions described here.
Written by the indexing model from the issue text.
Description
Feature Request
- Expose last_image in pipeline call
prepare_latents already accepts a last_image parameter (pipeline_mova.py:261), but call hardcodes it to None. This
is a minimal change — just wire the parameter through to enable first+last frame conditioning (FLF2V mode).
- Multi-frame reference conditioning (longer-term)
WAN 2.2 supports FLF2V (first-last-frame to video) and per-frame reference conditioning (Time to Move), which enables
much longer coherent video generation by providing visual anchors throughout the sequence.
Adding similar support to MOVA would be especially valuable for audio-conditioned generation, where longer clips are
needed to match dialogue or music. Currently the only way to extend duration is to increase num_frames, which quickly
hits VRAM limits at higher resolutions.
Key considerations for MOVA specifically:
- How per-frame visual references would interact with the audio-video bridge cross-attention
- Whether reference frames could help maintain lip sync coherence over longer durations
- Potential for segment-based generation with overlapping context (difficult today because audio conditioning can't be
cleanly split)
Context
The base WAN architecture already supports this — MOVA's prepare_latents even has the last_image code path. The main
gap is in the pipeline interface and extending beyond two reference frames.
- Dominant language
- Python
- Stars
- 1.1k
- Forks
- 97
- PR merge metrics
- No merged PRs in 30d
Getting set up
This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from OpenMOSS/MOVA
-
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
-
Difficulty 3/5 1-2 days Newbie friendliness 48/100
-
Difficulty 4/5 3-5 days Newbie friendliness 38/100
-
Difficulty 3/5 1-2 days Newbie friendliness 48/100
-
Difficulty 5/5 Over a week Newbie friendliness 25/100
Similar issues
-
bug
Difficulty 1/5 Under an hour Newbie friendliness 88/100
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
Maintainers usually reply within 1 day
-
documentation
Difficulty 2/5 1-3 hours Newbie friendliness 84/100
sergey-miryanov/gcmon#230 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
Azure-Samples/azure-search-openai-demo#3249 ·
Maintainers usually reply within 1 day
-
[BUG] Failed archive extraction deletes the dataset cachePossibly taken @dajiaohuang claimed this today. Open
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
alibaba/GraphScope#4674 ·