Research map / Image generation / Autoregressive and latent generation
Colorization Transformer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 conditional transformer layers instead of input-addition conditioning | Axial Transformer colorization core · ImageNet validation | FID: 19.98 $\pm$ 0.20 -> 19.37 $\pm$ 0.09 | Helped |
| Add conditional attention, conditional MLP, and conditional layer norm (ablation) | ColTran core · ImageNet, 64 × 64 coarse-color validation | per-pixel log-likelihood | Helped |
| Add conditional attention (ablation) | ColTran core · ImageNet, 64 × 64 coarse-color validation | per-pixel log-likelihood | Helped |
| Use conditional scales and shifts together rather than scales only or shifts only (ablation) | Conditional transformer layers · ImageNet, 64 × 64 coarse-color validation | per-pixel log-likelihood | Helped |
| Condition attention on queries and keys as well as values (ablation) | Conditional self-attention · ImageNet, 64 × 64 coarse-color validation | per-pixel log-likelihood | Helped |
| Use learnable spatial pooling rather than fixed mean pooling in conditional layer norm (ablation) | Conditional layer norm · ImageNet, 64 × 64 coarse-color validation | per-pixel log-likelihood | Helped |
| Add the auxiliary parallel prediction model with λ=0.01 (ablation) | ColTran core · ImageNet holdout, 64 × 64 coarse samples | FID | Helped |
| Increase the auxiliary parallel-model loss weight above 0.01 (ablation) | ColTran core · ImageNet holdout, 64 × 64 coarse samples | FID | Hurt |
| Provide the grayscale image as input to each upsampler (ablation) | ColTran upsamplers · ImageNet holdout, 256 × 256 | FID: 27.0 -> 16.4 | Helped |
| Use separate color and spatial upsamplers instead of one shared upsampler (ablation) | ColTran upsamplers · ImageNet holdout, 256 × 256 | FID: 16.6 -> 16.4 | Helped |
| Reduce the autoregressive colorization resolution to 28 × 28 (ablation) | ColTran core · ImageNet validation | FID: 19.37 -> 22.06 | Hurt |
| Use top-K sampling with K=4 or K=8 instead of K=512 (ablation) | ColTran · ImageNet colorization | FID | Helped |
Closest papers by meaning
- Learning Diverse Image Colorization (CVPR 2017 · similarity 0.89)
- Instance-Aware Image Colorization (CVPR 2020 · similarity 0.89)
- Automatic Controllable Colorization via Imagination (CVPR 2024 · similarity 0.88)
- Enabling Region-Specific Control via Lassos in Point-Based Colorization (AAAI 2025 · similarity 0.85)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · similarity 0.84)
- Generative Quanta Color Imaging (CVPR 2024 · similarity 0.84)
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.