Research map / Image generation / Fast diffusion sampling
IQA-Adapter: Exploring Knowledge Transfer from Image Quality Assessment to Diffusion-based Generative Models
Khaled Abud, Sergey Lavrushkin, Alexey Kirillov, Dmitriy Vatolin
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 separate qualitative cross-attention instead of concatenating quality tokens with text tokens (ablation) | IQA-Adapter · Lexica prompts | SROCC with target: 0.26 -> 0.97 | Improved |
| add separate qualitative cross-attention (ablation) | IQA-Adapter · MS COCO 10k captions | FID: 29.04 -> 21.36 | Improved |
| add separate qualitative cross-attention (ablation) | IQA-Adapter · Lexica prompts | quality gain (%): 8.31 -> 8.95 | Improved |
| add qualitative negative guidance (ablation) | IQA-Adapter · Lexica prompts | quality gain (%): 8.95 -> 10.86 | Improved |
| add qualitative negative guidance (ablation) | IQA-Adapter · MS COCO 10k captions | FID: 21.36 -> 22.16 | Hurt |
| add sinusoidal positional encoding of input scores (ablation) | IQA-Adapter · Lexica prompts | quality metrics | No clear effect |
| add IQA-Adapter conditioning instead of gradient-based IQA guidance | SDXL · Lexica prompts | cross-metric quality gain: 7-9% | Improved |
| use gradient-based IQA guidance | SDXL · PartiPrompts | non-target metrics | Hurt |
| combine multiple IQA/IAA models (TOPIQ + LAION-AES) in the adapter (ablation) | IQA-Adapter · Lexica prompts | transferability | Improved |
| condition IQA-Adapter on high quality versus base SDXL | IQA-Adapter · Subjective study | win rate: 32% -> 60% | Improved |
| add IQA-Adapter (HYPER-IQA) to SDXL | SDXL · GenEval Position | accuracy: 10.50% -> 14.75% | Improved |
| add IQA-Adapter (TOPIQ FLIVE) to SDXL | SDXL · GenEval Overall | accuracy: 56.29% -> 56.76% | No clear effect |
Closest papers by meaning
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.91)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.89)
- Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.88)
- Causal-IQA: Towards the Generalization of Image Quality Assessment Based on Causal Inference (ICML 2024 · similarity 0.88)
- VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.88)
- LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations (ICML 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.
