Research map / Image generation / GANs
COCO-GAN: Generation by Parts via Conditional Coordinating
Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, Hwann-Tzong Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use discrete rather than continuous coordinate sampling (ablation) | COCO-GAN (N2,M2,S16) · CelebA 64x64 | FID: 6.13 -> 4.87 | Improved |
| let the discriminator see the full image (optimal D) (ablation) | COCO-GAN · CelebA 64x64 | FID: 4.87 -> 4.05 | Improved |
| let the generator produce the full image (optimal G) (ablation) | COCO-GAN · CelebA 64x64 | FID: 4.87 -> 6.12 | No clear effect |
| use one generator per spatial coordinate instead of a shared conditional generator (ablation) | COCO-GAN · CelebA 64x64 | FID: 4.87 -> 7.26 | Hurt |
| generate and discriminate by parts rather than full images (COCO-GAN vs projection D backbone) | Projection D · CelebA 128x128 | FID: 19.55 -> 5.74 | Improved |
| use COCO-GAN generation by parts reducing macro patch to part of image | Projection D backbone · CelebA 128x128 (batch size 128) | training memory (MB): 17,184 MB -> 8,992 MB (47.7% reduction) | Improved |
| shrink micro patch size down to 4x4 pixels (ablation) | COCO-GAN · CelebA 128x128 | FID: overall FID slightly degrades | Hurt |
| use a macro patch smaller than 1/16 of the full image (32x32 macro, 16x16 micro) (ablation) | COCO-GAN · CelebA 128x128 | image quality | Hurt |
| add CoordConv to COCO-GAN (ablation) | COCO-GAN · CelebA | visual quality | No clear effect |
| freeze later generator layers and train only the first two layers during post-training (ablation) | COCO-GAN post-training · LSUN 256x256 beyond-boundary | artifact stability | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- IntroVAE — IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis (NeurIPS 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Generating Multiple Objects at Spatially Distinct Locations (ICLR 2019 · similarity 0.88)
- Rethinking conditional GAN training: An approach using geometrically structured latent manifolds (NeurIPS 2021 · similarity 0.88)
- Instance-Conditioned GAN (NeurIPS 2021 · similarity 0.87)
- RoCGAN: Robust Conditional GAN (IJCV 2020 · similarity 0.87)
- Spatial Steerability of GANs via Self-Supervision from Discriminator (TPAMI 2024 · similarity 0.87)
- Semi-supervised FusedGAN for Conditional Image Generation (ECCV 2018 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
