Checklist AICC 2026 Round 5 · Task 1
Watermark Removal
Remove text and line watermarks from 64 × 64 bird images with a diffusion-based image-to-image model.
The task
Watermark removal is an image-restoration problem in which a model must separate essential image content from synthetic overlays. The task asks contestants to implement a diffusion-based approach; the statement points to DDPM, DDIM and Palette (image-to-image diffusion) as references.
The data are derived from CUB-200-2011 (Caltech-UCSD Birds-200-2011), resized to 64 × 64 pixels. The training set has 2,000 pairs of clean and watermarked images; the test set has 200 watermarked images. Two stochastic watermark styles are used: semantic text labels (for example 'CUB-2011', 'COPY') with luminance-based contrast colouring, and diagonal cross-lines.
Contestants reconstruct a clean version of each test image, preserving the bird, its colours and the background.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
dataset/train(2000 clean/watermarked image pairs),dataset/test(200 watermarked images) and a script that converts a folder of reconstructed images into the submission CSV.- You submit
- A single CSV with columns datapointID (e.g.
02001.png), subtaskID (1) and answer (the reconstructed PNG encoded as a Base85 string). - Scoring
- Mean squared error between reconstructed and ground-truth images, with pixel values normalised to [0, 1]; lower is better.
- Rules
- No pretrained models other than those provided in torchvision.
- No training on validation or test samples in any way (including unsupervised learning or computing statistics over them).
- No external datasets; no labelling of unlabelled or test data with a pretrained model (for example CLIP).
- Individual participation (maximum team size 1); at most 50 submissions per day.
- AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
- Format
- AICC Round 5, online on Kaggle, 27 Mar 2026 18:00 UTC – 29 Mar 2026 17:00 UTC.