DLMMDD Synthetic Image Attribution Challenge: 1st Place Solution

This repository contains the notebooks used for the winning entry in the DLMMDD Workshop Synthetic Source Attribution Challenge. The best private leaderboard accuracy was 0.997333. A separate submission reached the best public score of 0.999333.

Official Results

The scores below were recorded after the competition closed. When a notebook produced several files, the table lists each file that received an official score.

Notebook	Method	CSV submitted	Public LB	Private LB
DLMMDD_EVA02_Large_+_ConvNeXtV2_Balanced.ipynb	Three-fold EVA02-Large and ConvNeXt V2-Large ensemble with exact balanced assignment	submission_diverse_balanced_exact.csv	0.999333	0.996666
DLMMDD_prior_logit_ensemble.ipynb	Five-fold calibrated EVA02-Large ensemble with Sinkhorn balancing	submission_balanced_sinkhorn.csv	0.998000	0.997333
DLMMDD_prior_logit_ensemble.ipynb	Five-fold calibrated EVA02-Large ensemble with exact balanced assignment	submission_balanced_exact.csv	0.998000	0.997333
[5folds] DLMMDD_ensemble_with_tta_and_calibration.ipynb	Five-fold calibrated EVA02-Large ensemble before balancing	submission_calibrated_ensemble.csv	0.996000	0.997333
DLMMDD_EVA02_Large.ipynb	Three-fold EVA02-Large processed-validation model	submission_pad_processedval_eva02_large448_96gb_orig.csv	0.996666	0.994666
DLMMDD_EVA02_Large448_96GB_LBPush_5Fold.ipynb	Five-fold EVA02-Large, original test view	submission_pad_processedval_eva02_large448_96gb_5fold_orig.csv	0.995333	0.995333
DLMMDD_ConvNext_v2_large.ipynb	Three-fold ConvNeXt V2-Large ensemble	submission_convnextv2_large.csv	0.989333	0.993333

Dataset Access

Join the Kaggle competition before running the notebooks. The dataset is private to accepted competition participants, so each user must authenticate with a personal Kaggle credential.

1. Join the competition on Kaggle.
2. Create or download kaggle.json from the Kaggle account settings.
3. Upload the credential in Colab, or place it in the standard Kaggle configuration directory for a local run.
4. Run the authentication and competition-download cells.

The Colab notebooks expect data under:

/root/.cache/kagglehub/competitions/dlmmdd-workshop-synthetic-source-attribution-challenge

Edit the configuration paths when using another location. Do not commit a Kaggle credential to this repository.

Checkpoints

Training can be skipped when compatible checkpoints are available:

• Three-fold EVA02-Large checkpoint archive
• Three-fold ConvNeXt V2-Large checkpoint archive

The five-fold private-best path also needs EVA02-Large folds 3 and 4. Generate them with DLMMDD_EVA02_Large448_96GB_LBPush_5Fold.ipynb, or place a complete five-fold archive in the location configured by the calibrated-inference notebook.

For Colab, upload each archive to /content/ or /content/drive/MyDrive/ and edit the path variables. For a local run, extract the archives into the project directory and point the notebook configuration to those directories.

Private-Best Path

The private-best submission uses EVA02-Large only. ConvNeXt and pseudo-label fine-tuning are not part of this path.

1. Run DLMMDD_EVA02_Large.ipynb to train folds 0, 1, and 2. It writes checkpoints to dlmmdd_pad_processedval_eva02_large448_96gb_outputs.
2. Put those checkpoints in the output directory used by DLMMDD_EVA02_Large448_96GB_LBPush_5Fold.ipynb. With force_retrain=False, the notebook reuses folds 0-2 and trains folds 3-4.
3. Archive the resulting five checkpoints as dlmmdd_pad_processedval_eva02_large448_96gb_outputs5folds.zip, or edit the archive path in [5folds] DLMMDD_ensemble_with_tta_and_calibration.ipynb.
4. Run the calibrated-inference notebook. It fits one temperature per fold on the matching processed-validation split, predicts four test views, and averages the five calibrated log-probability matrices with equal weight.
5. Keep the generated ensemble_logprobs.npy and ensemble_ids.csv in the same row order. Run DLMMDD_prior_logit_ensemble.ipynb to create the raw, Sinkhorn-balanced, and exact-balanced submissions.

The four inference views combine scales 1.0 and 0.92, each with and without a horizontal flip. The fitted fold temperatures were 0.558371, 0.598287, 0.585643, 0.594193, and 0.586100.

The test set contains 3,000 images from ten sources. The competition composition and the observed training counts establish a target of 300 test images per class. Sinkhorn scaling adjusts the class marginals softly and changed three raw predictions. Exact assignment changed seven predictions while enforcing exactly 300 predictions per class. Both submissions retained the private score of 0.997333.

Best-Public Path

The best-public submission follows a separate branch:

1. Run DLMMDD_EVA02_Large.ipynb and DLMMDD_ConvNext_v2_large.ipynb, or load their three-fold checkpoint archives.
2. Run DLMMDD_EVA02_Large_+_ConvNeXtV2_Balanced.ipynb.
3. Submit submission_diverse_balanced_exact.csv.

This notebook calibrates each fold, evaluates four test views, and performs an out-of-fold weight sweep. The recorded sweep selected a ConvNeXt V2-Large weight of 0.40 and an EVA02-Large weight of 0.60. Exact balanced assignment then changed five predictions. This branch achieved the highest public score but not the highest private score.

Artifact Integrity

The two generated test-prediction arrays are not committed. Recreate them with the calibrated-inference notebook and compare their SHA-256 hashes before applying marginal balancing:

Artifact	SHA-256
[5folds] DLMMDD_ensemble_with_tta_and_calibration.ipynb	e4eb095a27e7104b487cc11f99816f59833967a48fceeb532187350fac86003f
DLMMDD_EVA02_Large448_96GB_LBPush_5Fold.ipynb	441ca3c3cad22d1f19f754021711e28c7370f7c62458fba6a3138cda84e8d199
ensemble_logprobs.npy	48defd0eb038ca8424f1460b5f63a3374e0c85dafdd6bcf1836a3e52d9af386c
ensemble_ids.csv	af9d7adce4699b68f7541040e1c3f2a5b1b6c5590d260afd7c598de6f5dd95ed

Hardware and Runtime

All recorded runs used Google Colab. Runtime comparisons are approximate because the notebooks ran on different GPU models and included different amounts of inference and file handling.

Stage	Hardware	Recorded or estimated time
EVA02-Large folds 0-2	NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB	169.5 min from epoch logs
EVA02-Large folds 3-4	NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB	about 113.7 min from epoch timings
Five-fold EVA02-Large training total	NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB	about 283.2 min
Five-fold calibration and four-view inference	NVIDIA L4	59.6 min for the inference cell
ConvNeXt V2-Large folds 0-2	NVIDIA A100-SXM4-80GB	93.9 min from epoch logs
Diverse ensemble and marginal balancing	NVIDIA L4 Colab session	consolidated runtime not captured

The 59.6-minute calibrated-inference figure excludes the separate dataset download and checkpoint extraction cells. Marginal balancing runs on the saved 3000 x 10 score matrix and takes little time compared with model inference.