Had an incredible opportunity to give two lectures on diffusion models at MLSS-Sénégal 🇸🇳 in early July!. Slides are available here:. Big thanks to @eugene_ndiaye for the invitation!.

I recently presented our work, “Inference-Time Guided Generation with Diffusion and Flow Models,” at HKUST (CVM 2025 keynote) and NTU (MMLab), covering three classes of guidance methods for diffusion models and their extensions to flow models. Slides:

#ICLR2025 Come join our StochSync poster (#103) this morning! We introduce a method that combines the best parts of Score Distillation Sampling and Diffusion Synchronization to generate high-quality and consistent panoramas and mesh textures.

🚀 We’re hiring!.The KAIST Visual AI Group is looking for Summer 2025 undergraduate interns. Interested in:.🌀 Diffusion / Flow / AR models (images, videos, text, more).🧠 VLMs / LLMs / Foundation models.🧊 3D generation & neural rendering. Apply now 👉

RT @_akhaliq: ORIGEN. Zero-Shot 3D Orientation Grounding in Text-to-Image Generation

Introducing ORIGEN: the first orientation-grounding method for image generation with multiple open-vocabulary objects. It’s a novel zero-shot, reward-guided approach using Langevin dynamics, built on a one-step generative model like Flux-schnell. Project:

GPT-4o vs. Our test-time scaling with FLUX (2/2). GPT-4o cannot precisely understand the text (e.g., misinterpreting “occupying chairs” on the left), while our test-time technique generates an image perfectly aligned with the prompt. Check out more 👇.🌐

GPT-4o vs. Our test-time scaling with FLUX (1/2). GPT-4o still cannot count objects (see the ten, not seven, tomatoes on the left), but our test-time technique makes it work with FLUX. What you need is not a new model, but a test-time technique!. 🌐

Unconditional Priors Matter!. The key to improving CFG-based "conditional" generation in diffusion models actually lies in the quality of their "unconditional" prior. Replace it with a better one to improve conditional generation!. 🌐

Thanks @_akhaliq! Our test-time technique makes image flow models way more controllable—better at matching text prompts, object counts, and object relationships; adding or removing concepts; and improving image aesthetics—all without finetuning!. Project:

🚀 Inference-time scaling for FLUX! Significant improvements in reward-guided generation with flow models, including text alignment, object counts, etc.—all at a compute cost under just $1! . 📄 Paper: 🔗 Project:

#CVPR2025.🚀Check out **VideoHandles** by Juil (@juilkoo), the first method for test-time 3D object composition editing in videos. 🔗 Project: 📄 arXiv:

#NeurIPS2024 .DiT generates not only higher-quality images but also opens up new possibilities for improving training-free spatial grounding. Come visit @yuseungleee 's GrounDiT poster to see how it works. Fri 4:30 p.m. - 7:30 p.m. East #2510. 🌐

#NeurIPS2024.Thursday afternoon, don't miss @USeungwoo0115's poster on Neural Pose Representation, a framework for pose generation and transfer based on neural keypoint representation and Jacobian field decoding. Thu 4:30 p.m. - 7:30 p.m. East #2202. 🌐

#NeurIPS2024.We'll be presenting SyncTweedies on Wednesday morning, a training-free diffusion synchronization technique that enables generation of various types of visual content using an image diffusion model. Wed, 11 a.m. - 2 p.m. PST.East #2605. 🌐

#SIGGRAPHAsia2024.🔥 Replace your Marching Cubes code with our Occupancy-Based Dual Contouring to reveal the "real" shape from either a signed distance function or an occupancy function. No neural networks involved. Web:

A huge thank you to Jiaming Song (@baaadas) for delivering a wonderful guest lecture in our "Diffusion Models and Applications" course! He shared valuable insights on video generative models and the future of generative AI. 🎥 📚

#DiffusionModels.🎓 Our "Diffusion Models and Their Applications" course is now fully available! It includes all the lecture slides, recordings, and hands-on programming assignments. Hope it helps anyone studying diffusion models. 🌐

#DiffusionModels.Big thanks to @OPatashnik for the insightful guest lecture in our diffusion models course on using attention layers in diffusion networks for image manipulation. Don’t miss the recording!.🎥:

#NeurIPS2024.We introduce GrounDiT, a SotA training-free spatial grounding technique leveraging a unique property of DiT: sharing semantics in the joint denoising process. Discover how we **transplant** object patches to designated boxes. Great work by @yuseungleee and Taehoon.