Roblox's PixVid lifts latent video diffusion into pixel space for 2x resolution
Roblox · hf · 2026-11-14
Roblox and collaborators present PixVid, a method that lifts a pretrained latent video diffusion model (finetuned from Wan2.1-14B) to generate every frame directly in pixel space, keeping the diffusion transformer and replacing the VAE decoder with a convolutional head.
Key findings:
- Predicting clean video instead of velocity, plus a larger noise shift, is key to quality
- Competitive with the latent baseline within 1.3M training clips
- Generates at 2x higher resolution (960×1664 vs 480×832) with the same per-frame token count
- Framewise pixel generation keeps depth-conditioned editing aligned under fast motion and responds to each action at the next frame, enabling interactive generation
More from Multimodal
- Repainting the Sistine Chapel with AI people, via Claude — mickeyxfriedman · 2026-11-14
- fofrAI adds realistic human motion to custom VFX tool using Carbon agents — fofrAI · 2026-11-14
- Full-duplex speech model Haru-4B trends on Hugging Face — higurashi-ai · 2026-11-14
- ByteDance releases Seedance 2.5 Flash video model at 35% off — videorouter · 2026-11-14
- OpenAI Images 2.6 tops DesignArena and Image Arena at 4s per image, same price — thursdai_pod · 2026-11-14
- MiniMax H3 Camera & Motion Control Tested on a RTX 4070 with Only 8GB VRAM — big-boss_97 · 2026-11-14