MIND model slashes image generation FID by half vs. DiT baseline

A new paper from Duoduo Xue, Zhiyu Zhu, and Junhui Hou introduces MIND (Data Manifold‑aware Image diffusioN moDel), a diffusion framework that explicitly models the geometry of the data manifold. Unlike standard diffusion models that rely solely on continuous score estimation, MIND integrates discrete patch tokenization directly into the score function. This hybrid approach leverages the structural quantification strengths of discrete tokens while retaining the parallel generation flexibility of continuous diffusion. The authors enable end‑to‑end differentiable training through a novel soft top‑k aggregation mechanism, and they introduce dual‑branch high‑frequency feature embedding layers to counteract the spectral bias of transformer backbones on low‑dimensional inputs. For inference, a multi‑stage transition sampling scheme dynamically adjusts the sampling strategy based on timestep.

On ImageNet 256×256, MIND demonstrates dramatic improvements. After only 80 epochs of training, the base model achieves an FID of 22.73 without guidance—nearly halving the 43.47 FID of the vanilla DiT‑B/2 baseline. On average, MIND reduces FID by 15.95 over DiT and 9.06 over SiT. With guidance, the MIND‑B variant (just 130M parameters) achieves an FID of 2.06, surpassing the 3.1B‑parameter LlamaGen‑3B model. Scaling up, MIND‑XL (715M parameters) pushes FID down to 1.95. These results indicate that explicit manifold modeling can dramatically improve sample quality and parameter efficiency. The authors plan to release the code publicly, opening the door for further research and practical applications in high‑fidelity image generation.