YeTI: You Only Need Two Noisy Images for Real-World sRGB Noise Generation

ECCV 2026 · Paper (arXiv:2607.09193) · Code (GitHub) · Dataset

Jaekyun Ko^1,2*, Byung Wan Lim^1*, Soomin Lee¹, Dongjin Kim¹, Tae Hyun Kim^1†
1Department of Computer Science, Hanyang University  ²Mobile eXperience (MX) Division, Samsung Electronics
^*equal contribution  ^†corresponding author

Model Description

YeTI is a real-world sRGB noise generation framework that learns to synthesize realistic, signal-dependent sensor noise from only two noisy observations of the same scene — no clean ground truth or camera metadata required. It uses a Reconstruction Autoencoder (RAE) to disentangle scene structure from noise characteristics, and a one-step Conditional Diffusion Transformer (C-DiT) trained with consistency objectives to model the latent noise distribution. At inference time, YeTI takes a single noisy image and generates additional realistic noisy samples of the same scene, which can be used to train downstream (self-supervised) denoisers such as AP-BSN and MM-BSN.

Full details, training recipe, and evaluation protocol are in the paper and the official code repository.

Intended Uses & Limitations

Intended use: academic research on real-world sRGB noise modeling / synthesis, and as a data-augmentation source for training self-supervised image denoisers. Requires the YeTI code repository to load and run — these checkpoints are not standalone transformers/diffusers models.

Limitations: trained primarily on smartphone sRGB noise (SIDD) and MAI2021 data; noise realism may degrade on sensors, ISO ranges, or capture pipelines far outside the training distribution. Released for academic / non-commercial research use only — see License below.

Files in this repository

File	Description	Size
rae.ckpt	Reconstruction AutoEncoder — disentangles structure / noise latents	~148 MB
c_dit.ckpt	Conditional Diffusion Transformer — main noise generation model	~856 MB
apbsn.ckpt	AP-BSN self-supervised denoiser, trained on YeTI-generated noisy data only	~46 MB
apbsn_mix.ckpt	AP-BSN denoiser trained on a 50:50 mix of YeTI-generated and real noisy data	~46 MB
mmbsn.ckpt	MM-BSN self-supervised denoiser, trained on YeTI-generated noisy data only	~68 MB
mmbsn_mix.ckpt	MM-BSN denoiser trained on a 50:50 mix of YeTI-generated and real noisy data	~68 MB
config.json	Machine-readable checkpoint index	—

Training Data

Trained on the public SIDD (Smartphone Image Denoising Dataset) and MAI2021 benchmarks. Preprocessed archives matching this codebase's expected format are hosted at BWLim/YeTI (dataset).

How to use

These checkpoints are PyTorch Lightning .ckpt files meant to be used with the official YeTI code repository, which defines the model architectures (yeti/archs) and LightningModules (yeti/models) needed to load them.

# 1) Clone the code
git clone https://github.com/ByungWanLim/YeTI.git
cd YeTI

# 2) Download these weights into ckpt/
huggingface-cli download BWLim/YeTI --local-dir ckpt

# 3) Run validation / generation, e.g.
python main.py --config configs/val/C-DiT/val_lit_c-dit.yaml --ckpt ckpt/c_dit.ckpt

See the repository README for the full training / evaluation / noise-generation usage (environment setup, dataset preparation, and all run_*.sh scripts).

Citation

@article{ko2026yeti,
  title   = {YeTI: You Only Need Two Noisy Images for Real-World sRGB Noise Generation},
  author  = {Ko, Jaekyun and Lim, Byung Wan and Lee, Soomin and Kim, Dongjin and Kim, Tae Hyun},
  journal = {arXiv preprint arXiv:2607.09193},
  year    = {2026}
}

License

These model weights are released for academic / non-commercial research use only. See LICENSE for full terms. For commercial licensing inquiries, please contact the corresponding author (Tae Hyun Kim, Hanyang University).