SprintML/ml-in-pl-bonus

Model Tracer (bonus task): which generator made this image?

Task. Identify which specific model version generated a given image.

Goal

Decide whether an image was generated by one of eight autoregressive image generators, RAR-B, RAR-L, RAR-XL, RAR-XXL or VAR-d16, VAR-d20, VAR-d24, VAR-d30, or whether it comes from an unknown model (outlier). Attribute every image to its source model, or mark it as an outlier.

Challenge

You get a limited number of labeled examples from the eight known models. The challenge is to build a method that generalizes to unseen images. A critical part is out-of-distribution (OOD) detection: images that do not belong to any known model must be recognized and labeled outlier. Models of the same family look alike (for example rarl and rarxl, or var20 and var24), so telling sizes apart takes more than a glance.

Evaluation

Submissions are scored by accuracy, the share of correctly predicted labels over all test images:

$$\text{accuracy}=\frac{\text{number of correct predictions}}{\text{total number of predictions}}\backslash text\{accuracy\}\; =\; \backslash frac\{\backslash text\{number\; of\; correct\; predictions\}\}\{\backslash text\{total\; number\; of\; predictions\}\}$$

Every test image counts equally.

The data (Dataset.zip)

One archive with three splits (11,050 images in total, all 256 x 256 PNG):

split	labels	images
train/	one folder per model: var16, var20, var24, var30, rarb, rarl, rarxl, rarxxl	800 (100 per model)
val/	the same eight folders plus outlier/	450 (50 per folder: 400 images of known models and 50 outliers)
test/	none (hidden), files 1.png, 2.png, ...	9,800

The test split is used only for scoring. On Kaggle, 30% of it determines the public leaderboard and the other 70% the final private ranking.

Your task

Assign exactly one label to every test image. Allowed labels:

var16, var20, var24, var30, rarb, rarl, rarxl, rarxxl, outlier

White-box access to the generators

You are not limited to pixels. The image generators are provided, so you can generate your own images and inspect model-level signals such as token probabilities, likelihoods and losses. Participants are encouraged to use this model-level information, not only the pixels.

The provided generator code (the two scripts that expose these signals explicitly):

file	what it does
VAR/generate_var_images.py	generates class-conditional images with a pretrained VAR model and its VQ-VAE, and computes token-level cross-entropy and token probabilities. Set MODEL_DEPTH in the script (16, 20, 24 or 30; it must match the checkpoint). Run from VAR/: python generate_var_images.py. Outputs go to VAR/outputs/: sample_*.png (images), summary.csv (per-image loss and probability statistics), details.npz (per-token losses, probabilities and tokens) and metadata.json (the run configuration).
RAR/1d-tokenizer/generate_rar_images.py	generates class-conditional images with a pretrained RAR generator and its MaskGIT-VQ tokenizer, re-tokenizes the generated images, runs the generator on those tokens and computes the per-token negative log-likelihood and mean token probability. Run from RAR/1d-tokenizer/: python generate_rar_images.py. Outputs go to RAR/1d-tokenizer/outputs_rar/ (images, summary.csv, details.npz, metadata.json).

Supporting files:

file or folder	content
VAR/download_var_checkpoints.py, RAR/1d-tokenizer/download_rar_checkpoints.py	download the large model weights (VAR d16, d20, d24, d30 and its VAE; RAR-B, RAR-L, RAR-XL, RAR-XXL and the MaskGIT-VQ tokenizer) from their original Hugging Face repositories. A GPU is needed to run the generators.
VAR/checkpoints/var_ae_ft.pth, RAR/1d-tokenizer/checkpoints/rar_ae_ft_delta.pth	the tokenizer (autoencoder) checkpoints that ship with this repository
VAR/models/, RAR/1d-tokenizer/modeling/	the model definitions (architectures, tokenizers, quantizers)
VAR/outputs/, RAR/1d-tokenizer/outputs_rar/	example outputs of the two scripts
the remaining files (training scripts, demos, configs, requirements)	unchanged from the upstream VAR and RAR projects

The code keeps the licenses of the original projects (see VAR/LICENSE and RAR/1d-tokenizer/LICENSE).

Submission format

A CSV named submission.csv with the header image_name,label:

image_name,label
1.png,rarxxl
2.png,outlier
3.png,rarl
4.png,var16

Requirements: every test image exactly once; every image_name ends in .png; labels exactly as listed above (spelling matters); file size under 10 MB. sample_submission.csv shows the format (its labels are random).

Files in this repository

Dataset.zip (the three splits), sample_submission.csv, VAR/ and RAR/ (the generator code described above) and this README.

Competition

The task runs as a Kaggle competition: https://www.kaggle.com/competitions/model-tracer-bonus-task

References

1. Bihe Zhao, Louis Kerner, Michel Meintz, Tameem Bakr, Franziska Boenisch, Adam Dziedzic. Data Provenance for Image Auto-Regressive Generation. https://adam-dziedzic.com/static/assets/papers/DataProvenance.pdf
2. Chao Wang, Kejiang Chen, Zijin Yang, Yaofei Wang, Weiming Zhang. AEDR: Training-free AI-generated image attribution via autoencoder double-reconstruction. https://arxiv.org/abs/2507.18988
3. Zhenting Wang, Vikash Sehwag, Chen Chen, Lingjuan Lyu, Dimitris N. Metaxas, Shiqing Ma. How to trace latent generative model generated images without artificial watermark? ICML 2024. https://openreview.net/forum?id=TwZ2sY6eJj
4. Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, Liang-Chieh Chen. Randomized Autoregressive Visual Generation. 2024. https://yucornetto.github.io/projects/rar.html