Research map / Image generation / Text-to-image diffusion
Reflect-DiT: Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection
Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Arsh Koneru, Yusuke Kato, Kazuki Kozuka, Aditya Grover
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 |
|---|---|---|---|
| replace best-of-20 sampling with in-context reflection (Reflect-DiT) | SANA-1.0-1.6B · GenEval overall | accuracy: 0.62 -> 0.81 | Improved |
| replace best-of-20 with Reflect-DiT at same sample budget | SANA-1.0-1.6B · GenEval overall | accuracy: 0.75 -> 0.81 | Improved |
| replace best-of-20 with Reflect-DiT | SANA-1.0-1.6B · GenEval (Position) | accuracy: 0.54 -> 0.66 | Improved |
| use Reflect-DiT with 20 samples instead of SANA-1.5-4.8B best-of-2048 | SANA-1.0-1.6B · GenEval overall | accuracy: 0.80 -> 0.81 | Improved |
| replace SFT or DPO fine-tuning plus best-of-N with Reflect-DiT | SANA-1.0-1.6B · GenEval overall, N=4 | accuracy: 0.736 -> 0.776 | Improved |
| replace Diffusion-DPO with Reflect-DiT | SANA-1.0-1.6B · GenEval overall, N=20 | accuracy: 0.765 -> 0.807 | Improved |
| increase in-context feedback from K=1 to K=3 (ablation) | Reflect-DiT · GenEval overall, N=20 | accuracy: 0.766 -> 0.807 | Improved |
| increase in-context feedback from K=3 to K=5 (ablation) | Reflect-DiT · GenEval overall, N=20 | accuracy: 0.807 -> 0.816 | Improved |
| increase image embedding size from 4x4 to 8x8 (ablation) | Reflect-DiT · GenEval overall, N=20 | accuracy: 0.786 -> 0.807 | Improved |
| increase image embedding size from 6x6 to 8x8 (ablation) | Reflect-DiT · GenEval overall, N=20 | accuracy: 0.801 -> 0.807 | Improved |
| replace best-of-20 with Reflect-DiT | SANA-1.0 · DPG-Bench Hard-56 | score: 41.0 -> 51.8 | Improved |
| replace best-of-20 with Reflect-DiT | SANA-1.0 · DPG-Bench All | score: 85.6 -> 86.1 | Improved |
Later papers that compared with this method
1 later accepted paper reports Reflect-DiT as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- RAISE: Requirement-Adaptive Evolutionary Refinement for Training-Free Text-to-Image Alignment (CVPR 2026 · better on 1 of 1 shared result)
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
- SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer (ICML 2025 · similarity 0.92)
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.91)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.91)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.91)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.91)
- EC-DIT: Scaling Diffusion Transformers with Adaptive Expert-Choice Routing (ICLR 2025 · similarity 0.90)
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.
