File size: 2,656 Bytes
3a65b94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
---
license: openrail++
library_name: coreai
pipeline_tag: image-to-image
tags:
- super-resolution
- adcsr
- coreai
- apple-silicon
---

# 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](https://github.com/Guaishou74851/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`](https://huggingface.co/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).