arraypress/upscale-adcsr

AdcSR ×4 for Core AI — the "detail" upscaler (it INVENTS detail)

Read this first: AdcSR is a one-step diffusion super-resolver. It does not only sharpen what is in the picture — it adds plausible detail that was never there (texture in hair, foliage, fabric). That looks best on generated images and casual pictures, and is wrong for real photos you need to trust (evidence, documents, faces you must recognise, measurements). upscale offers it only as an opt-in --detail mode that says so; its default is SwinIR-L (arraypress/upscale-swinir-l-real).

What it is

• adcsr_x4_float32.aimodel: input lr [1,3,128,128] in [-1,1], output sr [1,3,512,512] in [-1,1] (×4), per-tile colour match baked in. float32 (~1.8 GB): the pruned SD 2.1 UNet overflows in float16 (NaN on smooth tiles).
• Model: AdcSR (CVPR 2025, Apache-2.0 code) — Adversarial Diffusion Compression of a Stable Diffusion 2.1 based super-resolver.

Provenance and what was verified

• Republished unchanged from the community Core AI conversion mlboydaisuke/AdcSR-CoreAI at commit 39ae70aeebbe09b7cc2740c81bfc5fd71c084f41. Its card reports parity with the PyTorch reference (cosine 1.000012).
• We could not re-check that parity ourselves: the reference needs stabilityai/stable-diffusion-2-1-base, which is no longer downloadable from Hugging Face (Stability withdrew it), so neither a fresh export nor a side-by-side PyTorch run was possible.
• What we did measure (M3 Max, upscale bake-off, ×4 shrink-and-back on 5 generated + 5 stock photos, LPIPS lower is better): generated 0.181 (best of seven upscalers), real 0.241; with blur, noise and JPEG damage, real 0.262 (best). PSNR is 3–4 dB below the faithful upscalers — the invented detail. 1024 → 4096 in 26.6 s on the GPU.

Licences

• AdcSR code and weights: Apache-2.0 — LICENSE-AdcSR-Apache-2.0.
• Built on Stable Diffusion 2.1: CreativeML Open RAIL++-M License dated November 24, 2022 (© Stability AI) — LICENSE-OpenRAIL++-M, including Attachment A, the use restrictions, which pass on to anyone using or redistributing this model. Commercial use is allowed within those restrictions.
• Text copied from Stability AI's original LICENSE-MODEL (their repository is no longer online; the file was taken from two independent vendored copies that are byte-identical).