Really impressive results for human-object interaction. They use a two-phase process where they optimize the diffusion noise, instead of the motion itself, to get to sub-centimeter precision while staying on manifold 🧠 HOIDiNi - hoidini.github.io

1/ Can we teach a motion model to "dance like a chicken" Or better: Can LoRA help motion diffusion models learn expressive, editable styles without forgetting how to move? Led by Haim Sawdayee, Chuan Guo, we explore this in our latest work. 🎥 haimsaw.github.io/LoRA-MDM/ 🧵👇

1/9 Excited to share EditP23! 🎨 Finally, a single tool for ALL your 3D editing needs: ✅ Pose & Geometry Changes ✅ Object Additions ✅ Global Style Transformations ✅ Local Modifications All driven by one simple 2D image edit. It's mask-free ✨ and works in seconds ⚡️. 🧵

Everyone uses CFG with w = 7.5… but why?? 🤔 For non-trivial prompts, choosing w is hard: too low → weak alignment, too high → artifacts. We show w shouldn’t be fixed — it should adapt! ✨ We present a tiny but smart MLP to adjust w over time → Better images/alignment 🚀 1/n

The Mahalanobis distance is the natural metric for Gaussian signals. But how can it be generalized to arbitrary probability densities? And how should a solution be tested? We address these questions in a new paper with Pierre-Étienne Fiquet Florentin Guth Jona Ballé, and Eero Simoncelli

“Make it red.” “No! More red!” “Ughh… slightly less red.” “Perfect!” ♥️ 🎚️Kontinuous Kontext adds slider-based control over edit strength to instruction-based image editing, enabling smooth, continuous transformations!

The initial noise in diffusion models is surprisingly correlated with the final image. Our NoisePrints paper exploits this to provide a lightweight, distortion-free, cryptographically secure watermark for proving authorship of generated images & videos, requiring no model access.

🚀 New preprint! We present NP-Edit, a framework for training an image editing diffusion model without paired supervision. We use differentiable feedback from Vision-Language Models (VLMs) combined with distribution-matching loss (DMD) to learn editing directly. webpage:

(1/4) [HOIDiNi] hoidini.github.io 🧵: Diffusion models are great at generating free-form human motion but tend to break down when objects enter the scene. Human–object interaction demands millimetric precision, and even tiny errors cause hands to float or penetrate surfaces

“What big teeth you have!” said Red.👩‍🦰 “All because my model suffers from semantic leakage,” said the Wolf.🐺 When Text-to-Image models blur boundaries, identities collapse. Meet 𝐃𝐞𝐋𝐞𝐚𝐤𝐞𝐫, a lightweight inference-time fix that mitigates semantic leakage! 👇

Visual Diffusion Models are Geometric Solvers We cast geometry as images: a plain diffusion model denoises into valid solutions. It is simple, general and effective. Shown on Inscribed Square, Steiner Tree, and Maximum Area Polygonization - all classic hard problems.