Research map / Image generation / Text-to-image diffusion
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis
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.
Scales rectified flow transformers (MM-DiT-style) with a noise-biased sampling scheme for high-resolution text-to-image synthesis, reporting scaling trends and comparisons against other diffusion models.
Problem
Rectified flow has better theoretical properties than diffusion but has not been decisively established as a state-of-the-art approach for high-resolution text-to-image synthesis at scale.
Approach
Combines rectified flow training (straight-path noise-data interpolation) with a multimodal diffusion transformer (separate text and image weights with joint attention) and a re-weighted noise sampling emphasizing intermediate timesteps, then scales model size.
Key results
- Rectified flow with reweighted noise sampling outperforms prior diffusion formulations in the paper's ablations on high-resolution text-to-image synthesis.
- Scaled MM-DiT models show smooth validation loss improvement with model size and training steps.
- The largest model is reported to be competitive with closed-source text-to-image systems in human preference evaluation.
Limitations
- Only the architecture figure and appendix fragment are available, so full experimental details are not verifiable here.
- Comparisons rely on human preference evaluation with closed-source competitors whose details are not public.
Experimental setup
Benchmarks: GenEval T2I-CompBench human preference evaluation
Metrics: human preference win rate GenEval score validation loss
Baselines: SDXL Pixart-alpha DALL-E 3 Midjourney v6 Ideogram v1
Builds on: rectified flow diffusion transformer latent diffusion
Closest papers by meaning
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.90)
- Text-to-Image Rectified Flow as Plug-and-Play Priors (ICLR 2025 · similarity 0.90)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · similarity 0.89)
- OmniFlow: Any-to-Any Generation with Multi-Modal Rectified Flows (CVPR 2025 · similarity 0.88)
- Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow (ICLR 2025 · similarity 0.88)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 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.