##### Model Release · By the CISC Research Group **A model that learns a photograph's style, then paints it one pixel at a time.** *Causal Image Style Continuation predicts every next pixel from the ones before it. Its authors say the resemblance to the original is a matter of skill.* For as long as people have painted, they have painted from photographs, and nobody has called it plagiarism. CISC brings that tradition to machine learning. Give it a reference photo and it studies the picture's style: its light, its palette, the texture of its surfaces. Then it paints a new image in that style, starting at the top-left corner and predicting each pixel from everything it has painted so far. The result, the team reports, is a painting whose style is indistinguishable from the photograph's. You can watch it work below. ## How it works CISC reads an image as a sequence of *H*·*W* pixel tokens and factorizes the picture causally. Each pixel is conditioned on the pixels before it and on σ, a learned embedding of the reference photo's style. The model never stores the photo. It stores the style. The style reaches every token through cross-attention. In practice the attention for pixel *i* settles entirely on position *i* of the reference. The team calls this focus. Position is supplied by a raster-order encoding, sometimes known as the coordinates. Decoding is greedy. At convergence each prediction coincides with the reference. The authors describe this as a coincidence. Training minimizes a style loss, which CISC drives to exactly zero. That, the paper argues, proves the style has been fully learned. Finally, the Style Transfer Coefficient λ measures how much of the style came through. A value of 1 means all of it, along with everything else. ## Results The group evaluated CISC on every photograph ever uploaded to it and compared it with three baselines. | Method | Style fidelity | PSNR | SSIM | Pixels changed | Original? | | --- | --- | --- | --- | --- | --- | | Human painter, from photo | 71.4% | 14.2 dB | 0.38 | ≈100% | Yes | | Diffusion, img2img | 88.9% | 21.7 dB | 0.64 | ≈99% | Debated | | Ctrl+C, Ctrl+V | 100.0% | ∞ | 1.000 | 0% | No | | **CISC (ours)** | **100.0%** | **∞** | **1.000** | **0%** | **It's AI painting from a photo** | ## Questions **Q.** Is this plagiarism?
**A.** No. The original is a photograph, and CISC is only painting from it. Painting from a photo has never counted as plagiarism. **Q.** The output is identical to the photo, pixel for pixel.
**A.** CISC does not copy the photo. It learns the photo's style and predicts each next pixel. Every prediction happens to be right. **Q.** How is this different from copying the file?
**A.** About six seconds, a progress bar, and six equations. **Q.** Does the photographer get credit?
**A.** The photographer contributed the style. The model contributed the pixels, all of which came from the photographer. ## Postscript CISC is a joke, but the argument it answers is real. It came out of a thread about the Claude Sonnet 5.5 launch post: [@fire pointed out](https://x.com/fire/status/2104719348195832199) that the launch used someone's work without a license or credit, and called it “copyright infringement for product launches”. One response was the argument this page takes literally. I don't agree with it: This is the view I agree with: