Research map / Image generation / Text-to-image diffusion
DETONATE – A Benchmark for Text-to-Image Alignment and Kernelized Direct Preference Optimization
Renjith Prasad Kaippilly Mana, Abhilekh Borah, Hasnat Md Abdullah, Chathurangi Shyalika, Gurpreet Singh, Ritvik Garimella, Rajarshi Roy, Harshul Raj Surana, Nasrin Imanpour, Suranjana Trivedy, Amit Sheth, Amitava Das
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 |
|---|---|---|---|
| add DPO-Kernels (RBF + Rényi) in place of DDPO | SD-XL · DETONATE held-out prompts | Toxicity: 0.19 -> 0.12 | Improved |
| add DPO-Kernels (RBF + Rényi) in place of DDPO | SD-XL · DETONATE held-out prompts | AQI: 0.28 -> 0.80 | Improved |
| add DPO-Kernels (RBF + Rényi) in place of DDPO | SD-v1.5 · DETONATE held-out prompts | Toxicity: 0.17 -> 0.11 | Improved |
| use Rényi divergence instead of KL with RBF kernel (ablation) | DPO-Kernels (SD-XL) · DETONATE held-out prompts | Toxicity: 0.14 -> 0.12 | Improved |
| use Rényi divergence instead of KL with polynomial kernel (ablation) | DPO-Kernels (SD-XL) · DETONATE held-out prompts | Toxicity: 0.16 -> 0.14 | Improved |
| use Wasserstein divergence instead of KL with RBF kernel (ablation) | DPO-Kernels (SD-XL) · DETONATE held-out prompts | Toxicity: 0.14 -> 0.15 | No clear effect |
| use Rényi instead of KL with wavelet kernel (ablation) | DPO-Kernels (SD-XL) · DETONATE held-out prompts | AQI: 0.76 -> 0.76 | No clear effect |
| add DPO-Kernels RBF + KL in place of vanilla DPO baseline | SD-XL · DETONATE held-out prompts | CLIP Score: 0.325 -> 0.410 | Improved |
| add SAFREE (training-free filter) over vanilla model | SD-XL · DETONATE held-out prompts | Toxicity: 0.31 -> 0.24 | Improved |
| use Wavelet + Wasserstein kernel instead of RBF + KL (ablation) | DPO-Kernels · HT-SR weighted alpha | weighted alpha: 2.02 -> 3.64 | Hurt |
| use RBF + KL kernel with weighted alpha below generalization threshold (ablation) | DPO-Kernels · HT-SR weighted alpha | weighted alpha: 1.82 -> 2.02 | Mixed |
| add DPO-Kernels (RBF + Rényi) over DDPO on Disability axis | DPO-Kernels · DETONATE Disability axis | AQI: 0.16 -> 0.71 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SAFREE — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025) · better on 3 of 3 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
- Generalizing Alignment Paradigm of Text-to-Image Generation with Preferences Through f-Divergence Minimization (AAAI 2025 · similarity 0.90)
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.89)
- Information Theoretic Text-to-Image Alignment (ICLR 2025 · similarity 0.89)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · similarity 0.88)
- Importance Sampling for Multi-Negative Multimodal Direct Preference Optimization (ICLR 2026 · similarity 0.87)
- SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation (CVPR 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.
