Diffusion sampling gets a dedicated corrector: FID drops from 9.58 to 2.07 on ImageNet, and unguided generation approaches CFG

During diffusion sampling, the denoiser's input at each step comes from its own imperfect earlier predictions rather than the noisy real training images it saw during training, and this training-inference mismatch lets errors accumulate along the denoising trajectory, producing artifacts and incoherent structures. Previously, sample quality was improved mainly by classifier-free guidance (CFG), which raises quality but reduces diversity and causes oversaturation; another route was increasing sampling steps or using more accurate numerical integration, but gains plateau quickly once integration error is small. On ImageNet class-conditional generation at matched numbers of function evaluations, corrective sampling lowers FID from 9.58 to 2.07 for DiT-XL/2, from 8.55 to 1.92 for SiT-XL/2, and from 5.53 to 1.31 for SiTREPA-XL/2, against CFG references of 2.51, 2.11, and 1.35.

The corrector takes the denoiser's clean-sample prediction as input and predicts the same clean target, which makes correction an explicit learning task whose training inputs are supplied online by the base model. Standard denoising training only learns a mapping from noisy observations to clean images, whereas the corrector builds on that initial prediction and focuses on the remaining inaccuracies. Across DiT, SiT, and JiT, the denoiser and a separate corrector are initialized from publicly released checkpoints and jointly fine-tuned; for SiTREPA this takes 50K updates, about 1.25% of its 4M pretraining updates.

Training supervises only a single correction, but at inference the corrector can be applied repeatedly to its own outputs, iteratively refining the prediction at a fixed noise level so early artifacts are suppressed before the sampling update. Previously, extra inference compute was typically spent on more denoising steps, whereas correction spends it on improving each step's prediction. At matched total inference cost, iterative correction is more effective than simply increasing denoising steps; in text-to-image generation without CFG, GenEval rises from 28.30 to 50.09 and DPG-Bench from 70.71 to 76.82 at roughly 245 function evaluations.

The extra inference cost of correction can be removed by distillation: a student is trained to match the teacher's predictions after several correction iterations and needs only one forward pass per sampling step. Weight sharing lets the denoiser and corrector share a backbone with separate lightweight input embedders, cutting parameters from 1.35B to 0.68B, but each correction iteration still needs an additional forward pass. On SiT-XL/2 class-conditional ImageNet-256 generation, the student reaches FID 2.04, close to the teacher's corrected 1.92, while reducing function evaluations by about 67%.

The report covers ImageNet class-conditional and unconditional generation, text-to-image generation, and experiments on DiT, SiT, SiTREPA, and JiT; the corrector is initialized from publicly released checkpoints and jointly fine-tuned with a small number of updates relative to pretraining. A reader interested in unguided generation or the diversity cost of CFG can start with the class-conditional and unconditional ImageNet FID, precision, and recall comparisons, then the text-to-image GenEval and DPG-Bench results.

The corrector is trained only on the base denoiser's predictions yet can be applied repeatedly to its own outputs, and this extrapolation currently has only an empirical explanation, with theoretical understanding left open. Correction increases denoising error while reducing the gap between training and inference distributions, and how the two jointly affect sample quality is not fully resolved. Correction sometimes produces overly regular structures, such as disk-like pizza toppings or strawberry seeds arranged in a regular grid. When the denoiser is frozen and only the corrector is trained, FID rises again with more correction iterations, whereas joint training continues to improve.