We’re hiring a summer research intern at the MIT–IBM Computing Research Lab in Boston. We are looking for a PhD student near graduation. We’re interested in non-autoregressive architectures for language generation and reasoning, including diffusion models, flows, and EBMs.

First time reviewing at NeurIPS, I usually submit to CV or Graphics. 1. Author responses were very long (LLMs), it was a nightmare to respond to. It needs a lower word limit. 2. Papers were not good (but this is true for CV as well). 3. Not everything can be solved in revision.

Excited to share that ProxyPose has been accepted to NeurIPS 2026! 🎉 We track the 6-DoF motion of any query pixel using video-to-video translation. Grateful to @taubnerfelix @pooja_ravioli @kyroskutulakos @DaveLindell for making this happen! Details: ruihangzhang97.github.io/pro…

Still very excited to receive the #ECCV2026 Best Paper Award for Heat Kernel Textures: the Geodesic Gaussians That Do Not Splat [circle-group.github.io/resea…], along with @simo_foti, @CanerKorkmaz0 and Stefanos Zafeiriou. See Simone's post below for a detailed break-down.

Pixel diffusion models spend as much compute on blank sky as on faces. Every patch is a token in every layer. Our fine-tuned MiniT2I runs on region tokens instead. Half as many tokens, same quality, twice as fast, trained at different budgets for elastic inference.

Introducing Phota Identity V2, the fastest, most accurate model for persistent identity. Train once with 5 photos in 1 minute, then use that identity across supported models, including NB and GPT Image. The hardest part of generative visual AI isn't creating one great image. It's creating the same person or character again and again. Built for premium content, digital twins, and brand storytelling at scale. Create with more expressive faces, natural skin, and the freedom to put anyone in any scene.