Upload README.md with huggingface_hub

README.md

@@ -22,9 +22,12 @@ no prompt, no noise — so it is fast and small enough to run fully on-device, i
 
 ## What it is
 
-- **fp16, ~870 MB.** On the GPU the Core AI output matches the fp32 reference (cosine 1.000008).
+- **fp32, ~1.7 GB.** Output matches the torch reference (cosine 1.000012). fp32 because the
+  pruned SD-2.1 UNet's attention/group-norm overflow in fp16 (NaN on smooth tiles).
 - **Image → image, one step.** Input a low-resolution tile, get a 4× tile back. No text, no noise.
 - **456 M parameters** (pruned SD-2.1 UNet + half VAE decoder).
+- The graph outputs the **raw** SR; AdcSR's per-image color-match is applied host-side by
+  `SuperResolver` after tiling (baking it per-tile blows up uniform tiles).
 
 ## I/O contract (per tile)
 
@@ -40,9 +43,8 @@ let sr = try await SuperResolver(model: .adcsrX4)   // downloads this repo on fi
 let big = try await sr.upscale(cgImage)             // ×4; tiles any-size input + feather-blends
 ```
 
-`SuperResolver` splits any-size input into overlapping 128-px LR windows, runs each, and blends.
-Load uses explicit GPU specialization options (GraphModel does this) — the runtime's *default*-
-options load path has a GPU-delegate JIT bug, but an explicit `.gpu` load is clean.
+`SuperResolver` splits any-size input into overlapping 128-px LR windows, runs each, and blends
+(and caps very large inputs so the result stays a reasonable size).
 
 ## License & attribution