Diffusion models make images by repeatedly refining noise. An earlier method, REPA, trains a small add-on network to make an image generator’s internal features resemble those from a pretrained image-recognition model, then removes that network after training. RECON keeps it so later image-generation layers can reuse its visual cues. In an ImageNet test generating 256-by-256 images for specified classes, RECON scored 3.25 on gFID after 400,000 training steps, compared with 7.90 for REPA. gFID compares feature distributions that Inception-v3, an image-recognition network, extracts from generated and real images; lower means the distributions are closer, not that people preferred the samples.
During generation, the retained network uses the image generator’s current features to estimate the visual cues a pretrained image-recognition model would extract. It condenses those estimates into one visual summary and gives it to later layers alongside their information about small regions of the image. Those layers can then use the summary while refining the image.
During training, the researchers use the vision encoder’s actual features from clean images as a teacher signal. The retained add-on network, called a projection head, estimates those features from the image generator’s current features, but its estimates match the vision encoder less well at high noise. Training later layers on the clean-image signal at every noise level could therefore teach them to rely on a cue the projection head cannot reproduce during generation. RECON supplies that teacher signal only below a normalized noise level of 0.5, where 0 means a clean image and 1 means pure noise.
RECON lowered gFID, but its samples covered less variety
For the reported class-conditional score, the authors evaluated 50,000 generated samples with a 250-step sampler and no classifier-free guidance. RECON’s gFID fell further, to 2.21, after 2.4 million training steps.
A separate 80-epoch SiT-XL/2 ImageNet comparison with REPA found that RECON raised precision from 0.70 to 0.81 while recall fell from 0.65 to 0.56. The paper defines precision as the fraction of generated images that fall within the real-data feature range, and recall as how much of that range the generated set covers.
The authors submitted the study to arXiv on Oct. 14, 2026, as a preprint.
Sources
Illustration: MrAlanKoh, “Example of Denoising Diffusion models,” CC BY-SA 4.0, via Wikimedia Commons; unmodified. It illustrates diffusion denoising generally, not RECON’s method or results. Image and rights record; license terms.
