Rochester and NVIDIA Build Pixel-Space Diffusion Model for Precise Local Image Edits
URochester_Hajim · reddit · 2026-10-30
Image generators like ChatGPT and Gemini often wreck unrelated details when asked to fix one small thing. Researchers at the University of Rochester and NVIDIA attribute this to the latent-space diffusion architecture behind most state-of-the-art models: concepts are compressed into compact representations for efficiency, sacrificing pixel-level control.
Their PixelDiffusion (PixeLDIT) approach runs diffusion directly in pixel space, letting users make precise per-pixel edits without unintentionally altering the rest of the image.
- News: rochester.edu/newscenter/how-to-fix-ai-generated-images-pixel-by-pixel-728132/
- Paper: arXiv 2609.24919
- Project page: pixeldit.github.io/pixeldit2
Related event: Rochester and NVIDIA Unveil Pixel-Space Diffusion for Precise Image Editing(2 posts)→
More from Multimodal
- Pika H3 Flex ported to Mac via MLX: 5s video in ~5 minutes on M4 Max — Opening-Ad5541 · 2026-10-30
- Style Atlas adds a character node for Krea 2 style mixing, now open source — Dear-Spend-2865 · 2026-10-30
- Last Week in Multimodal AI: PixText, MOSS Edit in ComfyUI, Anima refs and Hy-Nomas port — Vast_Yak_4147 · 2026-10-30
- Krea 2 + H3 morphing in ComfyUI with Suno soundtrack guided by video analysis — Accurate_Public4674 · 2026-10-30
- One-shot building a Gaussian Splatting viewer in Runway App Studio for 1k credits — jongranskog · 2026-10-30
- Abstract monochrome silhouette prompt template with stark outlines and shading — LudovicCreator · 2026-10-30