Focused Diffusion-GAN (FDGAN): trained checkpoints and generated sample sets behind the reported results

Final EMA weights (iteration 200,000) of the 23 diffusion models trained with the FDGAN code for 'Focused Diffusion-GAN: Object-Centric Image Generation Using Integrated GAN and Diffusion Frameworks' (Image and Vision Computing, under revision): ADM, FDGAN* (FDGAN with the adversarial stage disabled), FDGAN and the ablation rows, on Cityscapes-Pedestrian (11 runs), Traffic Signs (5) and COCO potted plant (7); and the 3,000-image generated sample sets behind the reported results of these models on Traffic Signs (11 sets) and COCO potted plant (16 sets). Checkpoints are packed per dataset (FDGAN_checkpoints_.tar, one folder per training run with ema_0.9999_200000.pt). Sample sets are packed per dataset (FDGAN_samples_.tar, one zip per set, each holding the 3,000 PNG images and a meta.json with the conditioning box of every image and the real training image it came from). The 45 Cityscapes-Pedestrian sample sets are in the companion record, DOI 10.5281/zenodo.23786063 (access on request, Cityscapes licence). Run and set names follow the job and sheet names of the code repository https://github.com/Latzi/focused-diffusion-gan, which holds the code, the job files, the evaluation notebooks and the metric sheets with every reported number. The published baselines of the paper are not part of this record. SHA256SUMS.txt lists the SHA-256 checksum of every tar of both records.