Project-page gallery assets

Native-resolution images and thumbnails for the project-page galleries: 512 paired samples in each of ImageNet conditional, ImageNet unconditional, text-to-image conditional, and empty-prompt text-to-image generation.

ImageNet correction checkpoints

ImageNet correction checkpoints

Ready-to-load denoiser–corrector bundles for class-conditional and unconditional ImageNet-256.

Checkpoint	Updates	Evaluation / sampling recipe
JiT-L/16	100K	45 denoising steps, K=9
SiT-XL/2	100K	83 denoising steps, K=2
SiT-REPA-XL/2	50K	123 denoising steps, K=5, noise interval [0.5, 0.7]
SiT-REPA-XL/2	100K	42 denoising steps, K=5, full trajectory
SiT-REPA-XL/2 (unconditional), sit-repa-uncond-100k	100K	42 denoising steps, K=5, full trajectory

All recipes above use CFG=1. Sampling and training code is coming soon.

Download without a GitHub login:

python src/utils/download.py sit
# Or: jit / sit-repa-50k / sit-repa-100k / sit-repa-uncond-100k

Run from a clone of https://github.com/sunovivid/correction-training.
The downloader restores one model.safetensors and config.json per model in
checkpoints/NAME/, verifies their integrity, and resumes interrupted transfers.

These bundles contain inference and initialization weights, not optimizer
restart states. JiT operates on pixels; SiT and SiT-REPA use a frozen SD VAE,
which the sampling code downloads separately. Checkpoint metadata supplies the
corrector parameterization, trained gate, and sampling configuration.