Research map / Image generation / Text-to-image diffusion
Test-Time Alignment of Text-to-Image Diffusion Models via Null-Text Embedding Optimisation
Taehoon Kim, Henry Gouk, Timothy Hospedales
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 |
|---|---|---|---|
| optimise null-text embedding (Null-TTA) instead of noise/latents, vs DNO | SD-v1.5 · PickScore target | PickScore: 0.289 -> 0.315 | Improved |
| Null-TTA vs DAS on held-out ImageReward | SD-v1.5 · PickScore target | ImageReward: 0.871 -> 0.946 | Improved |
| increase nmax from 25 to 115 (ablation) | Null-TTA · HPSv2 target | HPSv2: 0.347 -> 0.428 | Improved |
| increase nmax from 25 to 115 (ablation) | Null-TTA · HPSv2 target | time per image: 4m33s -> 17m07s | Hurt |
| optimise only null-text embedding rather than full network | Null-TTA · HPSv2 target | GPU memory (MB): 20449 -> 17585 | Improved |
| use Null-TTA instead of DNO at similar reward | SD-v1.5 · HPSv2 target | wall-clock time: 19m38s -> 8m40s | Improved |
| apply Null-TTA to SDXL | SDXL · PickScore target | PickScore: 0.253 -> 0.282 | Improved |
| increase nmax from 25 to 45 (ablation) | Null-TTA · SDXL PickScore target | PickScore: 0.266 -> 0.282 | Improved |
| Null-TTA vs DNO on non-differentiable JPEG target (zeroth-order gradient) | SD-v1.5 · JPEG compressibility | JPEG reward: -30.132 -> -33.741 | Mixed |
| Null-TTA vs DNO on held-out metrics under JPEG target | SD-v1.5 · JPEG compressibility | Aesthetic: 3.988 -> 4.649 | Improved |
| Null-TTA vs DNO/DAS in user study | SD-v1.5 · User study HPDv2 | mean rank: 2.04 -> 1.81 | Improved |
| Null-TTA vs DAS in multi-objective PickScore+HPSv2 | SD-v1.5 · Multi-objective | Pareto frontier | Improved |
Closest papers by meaning
- Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.90)
- DyMO: Training-Free Diffusion Model Alignment with Dynamic Multi-Objective Scheduling (CVPR 2025 · similarity 0.89)
- Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment (CVPR 2025 · similarity 0.89)
- Aligning Text to Image in Diffusion Models is Easier Than You Think (NeurIPS 2025 · similarity 0.89)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.88)
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 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.
