Self-trained optical flow diffusion model warps LDM frames for coherent video
pixlpa · x · 2026-09-23
Developer pixlpa shares a self-built video pipeline: a diffusion model trained on optical flow maps plus a small latent diffusion model trained on frames. The LDM generates an image, the flow model warps it for temporal continuity, and LDM inpainting fills disoccluded areas—conditioned on a 64px current-frame image and previous 4 flow frames.
Related event: Dev Trains Motion-First Optical Flow Diffusion Model for Video Generation(2 posts)→
More from Multimodal
- Ant's inclusionAI Releases Ming Image-0.1 Design: 6B MIT-Licensed Model for Text-Rich Design — AdinaYakup · 2026-09-23
- Ant Group open-sources 6B image model Ming Image-0.1 under MIT, tops UI/UX leaderboard — AdinaYakup · 2026-09-23
- An AI-Generated Song Is Sweeping Spanish Instagram—and It's Actually Good — vboykis · 2026-09-23
- Reddit User Seeks Tool to Label Speaker Identity for Every Subtitle Line in a 15-Character Show — dtdisapointingresult · 2026-09-23
- Blender scenes baked to Gaussian splats: 14GB project becomes 98MB in-browser — willeastcott · 2026-09-23
- Founder predicts a breakout fully AI-generated hit within 6-12 months — idanbeck · 2026-09-23