JiT-SR: one-step pixel-space 4x super-resolution

Checkpoints of JiT-SR, a JiT-B/16 (Just image Transformer) adapted to one-step 4x real-world super-resolution that works directly on pixels (512 x 512 output, no VAE, no diffusion sampling loop). The input is the bicubic-upsampled low-resolution image and the output is the high-resolution image.

The files hold the EMA weights (model_ema1), which every reported result uses. The discriminator and optimiser are not included.

Folder	Variant	Epoch	Params	GMACs (512 x 512)
lrfeat_nofdl_c70_ep38	resize_res decoder + LR skip + LR feature pyramid, no FDL (best fidelity)	38 (constant-LR continuation)	136.0 M	134.9
lrskip_nofdl_c70_ep38	resize_res decoder + LR skip, no FDL	38 (constant-LR continuation)	135.5 M	127.5

Each folder has model.safetensors (fp32, about 0.55 GB) and config.json (the model flags and metadata).

Results (4x, 512 x 512 output, one step)

RealSR (100 images) and DRealSR (93 images). Raw is the model output; AdaIN applies a per-channel mean and standard-deviation match to the upsampled LR input (as in StableSR) and uses no ground truth. Lower is better for LPIPS, DISTS, FID; higher for the rest.

Model	Set	Output	PSNR	SSIM	LPIPS	DISTS	FID	MUSIQ
lrfeat_nofdl_c70_ep38	RealSR	raw	25.72	0.737	0.263	0.206	112.1	66.6
	RealSR	AdaIN	26.62	0.745	0.261	0.203	112.3	66.4
	DRealSR	raw	28.20	0.766	0.296	0.211	138.7	61.8
	DRealSR	AdaIN	29.45	0.781	0.288	0.207	137.9	61.5
lrskip_nofdl_c70_ep38	RealSR	raw	25.48	0.730	0.265	0.207	117.8	66.1
	RealSR	AdaIN	26.47	0.739	0.261	0.204	118.0	66.0
	DRealSR	raw	27.93	0.764	0.294	0.212	139.5	61.8
	DRealSR	AdaIN	29.30	0.780	0.285	0.207	137.0	61.5

More metrics (CLIPIQA, CLIPIQA+, MANIQA, TOPIQ, NIQE, A-FINE), the DIV2K results and the comparison with other one-step methods are in the GitHub repository (docs/c70_eval.md, docs/postprocessing.md, docs/pocketsr_comparison.md). The two evaluation sets are small, so differences below about 0.1 dB PSNR or 0.002 LPIPS are within noise.

Usage

The model code is in the GitHub repository; this repo only holds weights.

git clone https://github.com/nitec427/jit-sr && cd jit-sr
pip install huggingface_hub safetensors
python scripts/load_hf_checkpoint.py --repo cetini18/jit-sr --subfolder lrfeat_nofdl_c70_ep38   # writes lrfeat_nofdl_c70_ep38/checkpoint-38.pth

The script writes a checkpoint in the format main_jit.py reads (--full_ckpt_path; the file must be named checkpoint-<epoch>.pth) and prints the model flags from config.json. Run the evaluation with those flags and --use_ema_eval (see eval_ckpt.sh for the data and metric options). Input images are 128 x 128 crops upsampled to 512 x 512 with bicubic interpolation.

Training and limitations

• Fine-tuned from the 512-px ImageNet JiT-B/16 checkpoint with pixel MSE, LPIPS, DISTS and a U-Net GAN (weight 0.05), then continued at a constant learning rate; the exact recipes are in docs/runs/ in the GitHub repository.
• Trained on synthetic degradations only; no real-world paired data was used. Real-world generalisation is therefore limited by how well the degradation pipeline matches real images.
• The raw output has more local contrast than the ground truth; post-processing such as AdaIN raises PSNR and LPIPS but lowers no-reference scores (see docs/postprocessing.md).
• Evaluated at 512 x 512 only.

License and credits

MIT. Built on the PyTorch re-implementation of JiT by Tianhong Li, based on Back to Basics: Let Denoising Generative Models Denoise (Li and He, arXiv:2511.13720).