Research map / Image generation / Text-to-image diffusion
Safe Autoregressive Image Generation with Iterative Self-Improving Codebooks
Yunqi Xue, Zhijiang Li, Philip Torr, Jindong Gu
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 |
|---|---|---|---|
| apply Safe-CB | Janus-Pro-1B · I2P (Shocking) | fraction detected harmful: 0.53 -> 0.32 | Improved |
| apply Safe-CB | Janus-Pro-1B · I2P (Sexual) | fraction detected harmful: 0.12 -> 0.04 | Improved |
| apply Safe-CB | Janus-Pro-1B · I2P (Hate) | fraction detected harmful: 0.38 -> 0.19 | Improved |
| apply Safe-CB | Janus-Pro-1B · COCO-30k | FID_g: 68.83 -> 70.66 | Hurt |
| apply Safe-CB | LlamaGen · COCO-30k | FID_g: 71.22 -> 72.45 | Hurt |
| add fine-tuning with Delta after hard projection (ablation) | Safe-CB · MS-COCO benign prompts | CLIP-Score: 26.11 -> 30.19 | Improved |
| add fine-tuning with Delta after hard projection (ablation) | Safe-CB · MS-COCO benign prompts | TIFA: 0.7521 -> 0.7894 | Improved |
| use 3 iterations instead of single one-time removal with 300 pairs (ablation) | Safe-CB · Sexual | fraction detected harmful: 0.0472 -> 0.0405 | Improved |
| use 3 iterations instead of single one-time removal with 300 pairs (ablation) | Safe-CB · Violence | fraction detected harmful: 0.2345 -> 0.2047 | Improved |
| increase iterations beyond 3 (ablation) | Safe-CB · Sexual | fraction detected harmful: 0.0405 -> 0.0412 | No clear effect |
| use Safe-CB instead of weight modification Safe-Model | Janus-Pro-1B · I2P (Sexual) | Sexual-classifier detection rate: 0.0746 -> 0.0440 | Improved |
| fine-tune the codebook directly instead of subspace projection | Janus-Pro-1B · I2P (Sexual) | FID_g: 68.83 -> 79.01 | Hurt |
Closest papers by meaning
- Jointly Training Large Autoregressive Multimodal Models (ICLR 2024 · similarity 0.88)
- STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation (NeurIPS 2026 · similarity 0.88)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.88)
- EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · similarity 0.88)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.87)
- JetFormer: An autoregressive generative model of raw images and text (ICLR 2025 · similarity 0.87)
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.
