Research map / Image generation / Text-to-image diffusion
Cycle Consistency as Reward: Learning Image-Text Alignment without Human Preferences
Hyojin Bahng, Caroline Chan, Fredo Durand, Phillip Isola
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 DreamSim instead of LPIPS for image similarity (ablation) | CycleReward-I2T · DetailCaps-4870 | pairwise accuracy: 53.16 -> 58.02 | Improved |
| use DreamSim instead of CLIP for image similarity (ablation) | CycleReward-I2T · DetailCaps-4870 | pairwise accuracy: 57.90 -> 58.02 | Improved |
| use SBERT instead of BERT for text similarity (ablation) | CycleReward-T2I · DetailCaps-4870 | pairwise accuracy: 47.27 -> 51.74 | Improved |
| use SBERT instead of BERT for text similarity (ablation) | CycleReward-T2I · GenAI-Bench | pairwise accuracy: 55.20 -> 55.52 | Hurt |
| use Stable Diffusion 3 instead of SDXL-Turbo as text-to-image decoder (ablation) | CycleReward-I2T · DetailCaps-4870 | pairwise accuracy: 56.42 -> 58.02 | Improved |
| use InternVL2-26B instead of LLaVA-1.5-13B as image-to-text decoder (ablation) | CycleReward-T2I · DetailCaps-4870 | pairwise accuracy: 51.74 -> 57.21 | Improved |
| learn a reward model from cycle consistency instead of using raw scores | CycleReward-I2T · DetailCaps-4870 | pairwise accuracy: 56.46 -> 58.02 | Improved |
| jointly train on I2T and T2I preferences | CycleReward · DetailCaps-4870 | pairwise accuracy: 58.02 -> 60.50 | Improved |
| label ImageRewardDB with cycle consistency instead of human labels | ImageReward backbone · GenAI-Bench | pairwise accuracy: 56.70 -> 54.58 | No clear effect |
| DPO Qwen-VL-Chat with CyclePrefDB-I2T | Qwen-VL-Chat · DeCapBench | DCScore: 26.47 -> 30.63 | Improved |
| DPO Qwen-VL-Chat with CyclePrefDB-I2T | Qwen-VL-Chat · MME perception | score: 1460.2 -> 1485.7 | Improved |
| Diffusion DPO SD1.5 with CyclePrefDB-T2I | Stable Diffusion 1.5 · T2I-CompBench Spatial | score: 11.49 -> 16.55 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- ImageReward — ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation (NeurIPS 2023) · better on 1 of 2 shared results
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
- Leveraging Unpaired Data for Vision-Language Generative Models via Cycle Consistency (ICLR 2024 · similarity 0.90)
- Text as Partial Constraint: Core–Residual Alignment for Robust Vision–Language Learning (NeurIPS 2026 · similarity 0.89)
- Confidence-aware Reward Optimization for Fine-tuning Text-to-Image Models (ICLR 2024 · similarity 0.89)
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.89)
- Aligning Text to Image in Diffusion Models is Easier Than You Think (NeurIPS 2025 · similarity 0.88)
- Information Theoretic Text-to-Image Alignment (ICLR 2025 · similarity 0.88)
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.
