sr-diffusion

Research checkpoint storage for the sr-diffusion project.

GitHub: https://github.com/BitIntx/sr-diffusion

This is a public source-available, non-commercial research project. It trains a vision-only x4 latent diffusion super-resolution pipeline from scratch and does not use a pretrained text-to-image diffusion model.

Current artifacts are study/research checkpoints. They are not a production SR model.

License

• Checkpoints, generated samples, metrics, and other non-code artifacts: CC BY-NC 4.0.
• Source code: PolyForm Noncommercial License 1.0.0.

Commercial use is not permitted without separate written permission. This includes paid hosted inference, resale, and integration into commercial products.

Training data is not redistributed in this repository. Dataset license constraints should be reviewed before training or redistributing derived weights.

Artifacts

Path	Source	SHA256
LICENSE	LICENSE	c635a1fa2c80
CHECKPOINT_LICENSE.md	CHECKPOINT_LICENSE.md	da3b25759abb
configs/diffusion_photo10k_b32.yaml	diffusion_photo10k_b32.yaml	0fa6c0eeec85
checkpoints/stage3_diffusion_b32_best_eval_noise.pt	best_eval_noise.pt	ea4b458d668c
checkpoints/stage3_diffusion_b32_step_0025000.pt	step_0025000.pt	38d2f44adf65
metrics/stage3_b32_step_0024000_metrics.json	step_0024000_metrics.json	7df33601c289
metrics/stage3_b32_step_0025000_metrics.json	step_0025000_metrics.json	bebb82e594da

Stages

• Stage 1: factor-4 VAE / Autoencoder over 512px HR crops.
• Stage 2: deterministic LR-to-HR-latent pretraining with the Stage 1 VAE frozen.
• Stage 3: conditional latent diffusion U-Net, planned.