Continuous diffusion had a good run—now it’s time for Discrete diffusion! Introducing Anchored Posterior Sampling (APS) APS outperforms discrete and continuous baselines in terms of performance & scaling on inverse problems, stylization, and text-guided editing.

Differentiability propels continuous diffusion. But discrete diffusion is inherently not differentiable. APS enables differentiability 1️⃣ Quantized Expectation → gradient-like guidance for tilting base distribution 2️⃣ Anchored Remasking → adaptive decoding for unmasking order

🎥 Text-guided inpainting and garment styling Our approach (APS) generates prompt-aligned garment completions compared to prior methods. Red boxes highlight prompt-misalignment in competing approaches

🎨 APS goes beyond inverse problems. APS for training-free stylization which is an area largely dominated by continuous diffusion. Left = input (reference style image), Right = stylized outputs (with text describing content).

APS highlights the following achievements: ✔️ Sharper textures ✔️ Better measurement consistency ✔️ State-of-the-art PSNR/LPIPS compared to prior discrete methods

Why APS? ✅ Works with pretrained discrete diffusion models ✅ Handles linear + nonlinear degradations ✅ Efficient test-time scaling ✅ Unlocks tasks previous samplers struggle with

🔗 Project page: 🤗 HF Link: 💻 Code (coming soon): 📄 Paper:

Work done in collaboration with team members at @Google: Andreas Lugmayr (@AndreasLugmayr), Yasamin Jafarian, Srivatsan Varadharajan, Prof. Ira Kemelmacher-Shlizerman (@kemelmi) and my advisors from UT Austin: Prof. Constantine Caramanis (@cmcaram), Prof. Sanjay Shakkottai

I have done work in similar space. Here's my work. Very interesting read !! @litu_rout_