OmnimatteZero makes real-time, training-free video cleanup

This is one of those “research-to-workflow” papers that could matter quickly because it cuts out custom training while still handling messy visual effects that usually break edits.

• –The key unlock is guidance over diffusion attention (temporal + spatial), which improves consistency across frames instead of treating each frame as isolated inpainting.
• –Compared with earlier generative omnimatte approaches that rely on model training/fine-tuning, OmnimatteZero positions itself as a zero-training pipeline using off-the-shelf video diffusion backbones.
• –Open-source code lowers the barrier for researchers and tool builders to benchmark, adapt, and potentially productize effect-aware video cleanup.
• –Practical caveat: current usage still depends on good masks and substantial GPU memory, so “training-free” does not yet mean lightweight for everyday consumer hardware.