tvtsr-fused

TVTSR (fused) — Transfer VAE Training ×4 super-resolution

Pre-merged, ready-to-load weights for TVTSR (Fine-structure Preserved Real-world Image Super-resolution via Transfer VAE Training, arXiv 2507.20291):

• fused_unet.safetensors — the SD-2.1-base UNet with the TVT LoRA merged in (plain diffusers layout, so no runtime LoRA merge is needed).
• vae.safetensors — the TVT VAE-D4 (4× spatial compression instead of SD's 8×), exported from the original Lightning checkpoint to a plain AutoencoderKL safetensors.

These are convenience artifacts: content-identical to merging the official TVTSR release yourself, just pre-fused and converted so they load directly in any framework (diffusers, candle, …) with no LoRA-merge or Lightning-checkpoint step.

Provenance & credits

• TVTSR — the LoRA and the VAE-D4 are by the TVT authors: paper · code · original weights (Apache-2.0).
• Base UNet — a modified derivative of Stability AI's stable-diffusion-2-1-base.

License

Because fused_unet.safetensors is a modified derivative of SD-2.1-base, it is governed by the CreativeML Open RAIL++-M license, including its use-based restrictions (Attachment A) — by using it you agree to those terms. The TVT LoRA and VAE-D4 are Apache-2.0. Redistribute only with this license and attribution intact.

SD-2.1-base license / use restrictions: https://huggingface.co/stabilityai/stable-diffusion-2-1-base/blob/main/LICENSE.md