Rochester & NVIDIA unveil pixel-space diffusion model for precise, localized image edits
URochester_Hajim · reddit · 2026-10-29
Fixing one small detail in an AI-generated image often breaks something unrelated — a flaw rooted in latent-space diffusion, which compresses concepts and regenerates globally. Researchers at the University of Rochester and NVIDIA built a pixel-space diffusion model that lets users make precise pixel-level edits without unintentionally altering other details.
- Paper: arXiv 2609.24919
- Project page and code are open source (pixeldit.github.io/pixeldit2)
More from Multimodal
- Claude Skill churns out impressive educational explainer videos — AIandDesign · 2026-10-30
- Dev builds interactive 3D paleontology book with Claude Opus: fossils animate to life — techartist_ · 2026-10-30
- Openjutsu: open-source ComfyUI one-click video character swap using MiniMax H3, fully local — Devajyoti1231 · 2026-10-30
- "Darkness": a stunning dusk city flyover AI video demo — sdrowegnarts · 2026-10-30
- Falcon: Streaming Audio-Visual Speaker Extraction Cuts Params 91.9% and Latency 11× — _akhaliq · 2026-10-30
- Open-source StableGen 0.4 turns prompts into rigged, textured characters inside Blender — sakalond · 2026-10-30