New diffusable representation learning objective converges to pointwise mutual information, with joint tokenizer+diffusion scaling laws
kastnerkyle · x · 2026-10-20
Alan Baade, with Fei-Fei Li, Gordon Wetzstein, Justin Johnson and Vince Sitzmann, introduces a principled, scalable diffusable representation learning objective that works where DINO/LeJEPA collapse. Theory shows image representations converge so that their dot product with another image's representation equals pointwise mutual information. The team derives joint tokenizer+diffusion scaling laws, compute-optimal versus raw diffusion — positioning the tokenizer as a new scaling axis.
More from Multimodal
- Turkish AI film festival screens 65 AI-assisted films, eyeing week-long edition by 2027 — lmoroney · 2026-10-20
- Sunra launches MCP to bring leading image, video and audio models into Claude and ChatGPT at maker's API price — EXM7777 · 2026-10-20
- DLSSNR fine-tuning pipeline released: 16GB VRAM for full training, 8GB for LoRA — bdsqlsz · 2026-10-20
- Entire K-Pop Music Video One-Shotted with Opus 5.5 and Hedra MCP — AIandDesign · 2026-10-20
- One-shot K-Pop music video made with Opus 5.5 and Hedra MCP for virtual idol V3X — RebeccaBellan · 2026-10-20
- Stanford's LILIE: language-informed lossless image encoder cuts T2I gFD by 14 points — jcniebles · 2026-10-20