Dev combines custom motion model with image diffusion to make datamoshed animation
pixlpa · x · 2026-09-15
Developer pixlpa shared an experiment combining his own generative motion model (now with image conditioning) with a small image diffusion model that inpaints disoccluded areas and periodically regenerates the distorted image, producing glitchy, abstract "datamoshed" animation. He notes improvements in motion conditioning are bringing form and movement closer together in generative video.
Related event: Self-Trained Motion Model Plus Image Diffusion Creates Datamosh Animations(4 posts)→
More from Multimodal
- MiniMax H3 hits 2x real-time video denoising with SGLang + VDN-H3 on 8x B200 — MiniMax_AI · 2026-09-15
- Ant Group's Realtime-Venus: full-duplex interaction with asynchronous model delegation — antgroup · 2026-09-15
- Open-source LongCat-Avatar turns one photo plus audio into minutes of lip-synced talking video — Roger_M_Taylor · 2026-09-15
- Odyssey Teases World Models Part 3 Release, Following Interactive Video Push — soleio · 2026-09-15
- BVB benchmark tests agent video understanding by rebuilding 288 real videos in Blender — SonglinYang4 · 2026-09-15
- fal unveils H3 Max Camera Controls: single view becomes navigable 3D scene in under 3 seconds — isidentical · 2026-09-15