Research map / Image generation / Text-to-image diffusion
CHATS: Combining Human-Aligned Optimization and Test-Time Sampling for Text-to-Image Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CHATS, which trains separate preferred and dispreferred diffusion/flow models with a joint DPO-style objective and samples with a proxy-prompt CFG-like combination of the two models.
Problem
Text-to-image diffusion models use preference alignment (DPO) and classifier-free guidance separately, leaving their combination unexplored; this limits text-image alignment, quality, and human aesthetic consistency.
Approach
Model preferred and dispreferred distributions with two separate networks initialized from the same reference, trained with a joint DPO-derived loss (sum of reward terms), and sample by combining their noise predictions in a CFG-like formula using a proxy prompt that linearly merges the conditional and null embeddings.
Key results
- On SD1.5, CHATS reaches HPS v2 average of 26.17 on HPS prompts versus 22.04 for Diffusion-DPO-style baseline numbers in the GenEval table; GenEval overall 44.11 vs 41.99 for Diffusion-DPO.
- On GenEval with SDXL, CHATS achieves overall 60.12 vs 59.27 for Diffusion-DPO; on In-house T2I, 66.42 vs 62.22.
- Ablation on In-house T2I (Photo HPS v2): single model with full data 28.64 vs full CHATS 29.30; CHATS trained on 7,459 OIP pairs beats Diffusion-DPO on OIP across models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Standard with CHATS | SD1.5 · GenEval Overall | score (%): 38.52 -> 44.11 | Helped |
| replace Diffusion-DPO with CHATS | SDXL · GenEval Overall | score (%): 59.27 -> 60.12 | Helped |
| replace Diffusion-DPO with CHATS | In-house T2I · GenEval Overall | score (%): 62.22 -> 66.42 | Helped |
| replace Standard with CHATS | SD1.5 · DPG-Bench Overall | score (%): 57.92 -> 64.44 | Helped |
| replace Diffusion-DPO with CHATS | SDXL · DPG-Bench Overall | score (%): 74.66 -> 76.08 | Helped |
| replace Diffusion-DPO with CHATS | In-house T2I · DPG-Bench Overall | score (%): 82.18 -> 83.72 | Helped |
| train one model on preferred samples only rather than all preference samples (ablation) | single-model flow-matching finetuning · Photo | HPS v2: 28.64 -> 28.70 | Helped |
| use two separate models for preferred and dispreferred distributions (ablation) | single-model flow-matching finetuning · Photo | HPS v2: 28.70 -> 28.99 | Helped |
| add reference policy to two-model finetuning (ablation) | two-model finetuning · Photo | HPS v2: 28.99 -> 29.15 | Helped |
| add cooperative CHATS sampling with α=0.5 (ablation) | two-model finetuning · Photo | HPS v2: 29.15 -> 29.30 | Helped |
| finetune on OIP instead of PaP v2 (ablation) | CHATS with SDXL · Photo | HPS v2: 28.67 -> 29.62 | Helped |
| finetune on OIP instead of PaP v2 (ablation) | CHATS with In-house T2I · Paintings | HPS v2: 28.70 -> 29.95 | Helped |
Surprising findings
- Diffusion-DPO performs better when finetuned on the large PaP v2 dataset, while CHATS performs better on the small high-quality OIP dataset.
- Naive DPO loss causes collapse of the dispreferred model, motivating a reformulated objective.
Limitations
- Two forward passes at batch size 1 instead of one batched CFG pass reduce inference throughput.
- Requires training and storing two models, with the dispreferred model handling the unconditional branch.
- Theoretical derivation relies on approximations (q approximating p_theta, Jensen bound, p_theta(z_t) approx equal assumption).
Experimental setup
Datasets: Pick-a-Pic v2 OpenImagePreferences
Benchmarks: HPS v2 GenEval DPG-Bench
Metrics: HPS v2 score GenEval overall score DPG-Bench overall score ImageReward PickScore
Baselines: Standard SD1.5 Standard SDXL Standard In-house T2I Diffusion-DPO NPO Diffusion-NPO
Builds on: Diffusion-DPO classifier-free guidance DPO rectified flow
Closest papers by meaning
- Rich Human Feedback for Text-to-Image Generation (CVPR 2024 · similarity 0.91)
- ChatGen: Automatic Text-to-Image Generation From FreeStyle Chatting (CVPR 2025 · similarity 0.90)
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.90)
- Generalizing Alignment Paradigm of Text-to-Image Generation with Preferences Through f-Divergence Minimization (AAAI 2025 · similarity 0.90)
- Automated Filtering of Human Feedback Data for Aligning Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.89)
- Towards Effective Usage of Human-Centric Priors in Diffusion Models for Text-based Human Image Generation (CVPR 2024 · similarity 0.89)
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.