Sources, method & scope

How the comparison images were made

1. High-resolution reference: A native 2048 × 2048 px crop was taken from each officially supplied JPEG, without enlargement. It is the basis for comparing original contours, brushwork, and texture; it is not a lossless record of the physical painting.
2. Input image (512 px): The crop was downsampled to 512 × 512 px using Lanczos resampling. No additional noise, missing areas, or artificial damage was added.
3. AI upscaling (4×): RealESRGAN_x4plus enlarged the input to 2048 × 2048 px.
4. AI retouch: Qwen-Image-2.1 received the 512 px input and its AI-upscaled image, with instructions to preserve the shapes, lines, and textures appropriate to each medium. The complete generated image was saved unchanged as the candidate.

   BF16 / 40 steps / CFG 1 / fixed seed 20261022 / 2048 × 2048 px output / VAE tiling disabled

5. Blend: The saved AI retouch candidate was mixed with the AI-upscaled image at 0, 10, 25, 50, and 100% using 8-bit RGB values. At 0%, the AI-upscaled image is retained; full-image 100% is the AI retouch candidate.

The published processing results have no added alignment or geometric warping. No pixels from the high-resolution reference were pasted back into them.

Reading the masks and display

The blend has three scopes: full image, a manually selected interior region with a hard boundary, and the same interior region with a soft boundary that reduces the contribution over 24 px inward. The selected areas are inside the flower head in the photograph, the white background in the oil painting, and white color fields in the abstract painting, avoiding major contours and the signature. These masks were defined manually while viewing the AI-upscaled image; they are not automatic defect detection.

Black retains the AI-upscaled image, white applies the selected strength, and gray applies a lower strength. A soft mask smooths the boundary; it does not correct misalignment in the candidate. Even when the outside is retained, fine lines and brushwork within the selected area can change.

All stages share the same 2048 × 2048 px coordinates. For display, the 512 px input is enlarged 4× using nearest-neighbor resampling. This makes its pixels comparable without adding detail. The full-image view shows composition; 100% and 200% views show detail at the same location.

The five strengths were fixed in advance. There was no automatic search for a strength or mask that minimized error against the high-resolution reference.

Findings and evaluation scope

• Photograph: Low strengths add a slight impression of detail in places, but do not recover the reference image’s web-like fine hairs or the arrangement of individual fine lines.
• Oil painting: A fabric-like texture is added, but the original individual brush marks and grooves are not recovered. Low strengths can still mix different positions and tones of brushwork.
• Abstract painting: Line positions and signature strokes change. Full-image blends retain double edges, while interior masks still do not show a texture improvement that matches the original paint surface and cracks.

This comparison covers one image per category and one candidate per image at a fixed seed. It is not a model-wide success rate. Invented detail and shape changes can remain even at low strengths. A small improvement in pixel error alone was not treated as evidence of recovered detail or acceptable image quality.

Artist acceptance, blinded human evaluation, print quality, generalization to other images or seeds, and reductions in working time or total cost were not measured.

Image sources and rights statements

These are the sources and provider rights statements for the supplied images used in this comparison. The viewer shows crops and processed derivatives of those images.

These credits do not imply endorsement of this method or its results by the artists, museums, or USFWS. The rights statements above refer to the specified supplied images.