Research map / Image generation / Text-to-image diffusion
Histogram-constrained Image Generation
Haoming Liu, Yuanhe Guo, Yijia Cao, Shenji Wan, Hongyi Wen
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 OT guidance during sampling instead of direct OT on the finished image (ablation) | HIG color-constrained generation · Color-constrained generation | CLIP-Score / Aesthetics: 26.94 -> 27.19 CLIP, 6.52 -> 6.78 Aesthetics | Improved |
| add guidance during sampling versus direct OT, histogram alignment (ablation) | HIG · Color-constrained generation | HistKL: 0.00 -> 1.09 | Mixed |
| add post-hoc OT after guided sampling (ablation) | HIG (guidance) · Color-constrained generation | HistKL: 1.09 -> 0.00 | Improved |
| add post-hoc OT after guided sampling (ablation) | HIG (guidance) · Color-constrained generation | CLIP-Score / Aesthetics: 27.19 -> 26.91 CLIP, 6.78 -> 6.66 Aesthetics | Hurt |
| apply HIG color guidance versus unconstrained SDXL generation | SDXL · Color-constrained generation | HistKL: 10.87 -> 1.09 | Improved |
| apply HIG guidance versus unconstrained SDXL on prompt compliance | SDXL · Color-constrained generation | CLIP-Score: 29.47 -> 27.19 | Hurt |
| use multi-option binning instead of single-option binning for information embedding (ablation) | HIG information embedding · Information embedding, 256 tokens | exact match: 97.3% -> 97.7% | Improved |
| use multi-option binning instead of single-option binning for information embedding (ablation) | HIG information embedding · Information embedding, 512 tokens | exact match: 91.0% -> 90.3% | Hurt |
| increase embedded text from 32 to 512 tokens (ablation) | HIG information embedding · Information embedding | exact match (single-option): 100.0% -> 91.0% | Hurt |
| apply JPEG compression to embedded images (ablation) | HIG information embedding · Robustness | exact match: below 90% once text exceeds 32 tokens | Hurt |
| apply random scaling [0.5, 2.0] to embedded images (ablation) | HIG information embedding · Robustness | exact reconstruction rate: around 90% for text lengths up to 128 tokens | No clear effect |
Closest papers by meaning
- FlexControl: Computation-Aware Conditional Control with Differentiable Router for Text-to-Image Generation (ICML 2025 · similarity 0.86)
- CtrLoRA: An Extensible and Efficient Framework for Controllable Image Generation (ICLR 2025 · similarity 0.86)
- ControlAR: Controllable Image Generation with Autoregressive Models (ICLR 2025 · similarity 0.85)
- A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.85)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.85)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · 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.
