Research map / Image generation / Text-to-image diffusion
AMO Sampler: Enhancing Text Rendering with Overshooting
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.
Proposes AMO, a training-free overshooting sampler for pretrained rectified flow text-to-image models that alternates ODE overshooting with noise re-injection, modulated per patch by text-to-image cross-attention, to improve text rendering.
Problem
Text-to-image rectified flow models such as SD3, Flux, and AuraFlow frequently misspell, omit, or fail to render text in generated images, and fine-tuning fixes for this are costly and can degrade general image quality.
Approach
An overshooting sampler for rectified flow that over-simulates the learned ODE then re-injects noise (equivalent to Euler plus a Langevin correction term that fixes compounding discretization error), with the overshooting strength adaptively set per image patch by its attention to the text tokens to avoid over-smoothing.
Key results
- AMO raises human-judged correct rate from 32.5% to 43.0% on SD3 and from 74.0% to 82.5% on Flux versus the Euler sampler.
- AMO improves FID on all three models (e.g., AuraFlow 128.4 to 117.5) while keeping inference cost comparable to Euler.
- Plain overshooting without attention modulation yields 81.5% correct rate on Flux but over-smooths images and degrades FID to parity-lost levels; pure overshooting alone without noise compensation gives 0% correct rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Euler sampling with AMO | SD3 · DrawTextCreative, ChineseDrawText, TMDBEval500 | CR: 32.5 % -> 43.0 % | Helped |
| replace Euler sampling with AMO | Flux · DrawTextCreative, ChineseDrawText, TMDBEval500 | CR: 74.0 % -> 82.5 % | Helped |
| replace Euler sampling with AMO | AuraFlow · DrawTextCreative, ChineseDrawText, TMDBEval500 | CR: 1.0 % -> 3.0 % | Helped |
| replace Euler sampling with AMO | SD3 · DrawTextCreative, ChineseDrawText, TMDBEval500 | FID: 81.1 -> 80.9 | Helped |
| replace Euler sampling with AMO | Flux · DrawTextCreative, ChineseDrawText, TMDBEval500 | OCR-A: 0.313 -> 0.381 | Helped |
| replace Euler sampling with AMO | AuraFlow · DrawTextCreative, ChineseDrawText, TMDBEval500 | OCR-F: 0.238 -> 0.258 | Helped |
| use overshooting instead of discretizing the SDE | RF model · Text rendering accuracy, 20 steps | accuracy: 51.5 % -> 68.5 % | Helped |
| use overshooting instead of discretizing the SDE | RF model · Text rendering accuracy, 50 steps | accuracy: 71.0 % -> 81.5 % | Helped |
| use overshooting instead of discretizing the SDE | RF model · Text rendering accuracy, 100 steps | accuracy: 76.0 % -> 81.5 % | Helped |
| add overshooting without noise compensation (ablation) | AMO on Flux · Flux | CR: 74.0 % -> 0.0 % | Hurt |
| add noise compensation to overshooting (ablation) | AMO on Flux · Flux | CR: 0.0 % -> 81.5 % | Helped |
| remove attention modulation (ablation) | AMO on Flux · Flux | CR: 82.5 % -> 81.5 % | Hurt |
Surprising findings
- Overshooting alone without noise compensation drives the correct rate to 0%, since the marginal law is not preserved.
- Standard OCR metrics were unreliable for artistic fonts and extraneous poster content, so the authors relied on human evaluation.
Limitations
- No systematic evaluation of overshooting's effect on specific aesthetics or on domains beyond image generation.
- Automated OCR tools are unreliable for evaluating artistic text, requiring human evaluation on 100 prompts.
- Higher overshooting strength can introduce over-smoothing artifacts.
Experimental setup
Datasets: DrawTextCreative ChineseDrawText TMDBEval500
Benchmarks: DrawTextCreative ChineseDrawText TMDBEval500
Metrics: OCR accuracy OCR F-measure CLIP score FID correction rate (human evaluation)
Baselines: Euler sampler GlyphControl TextDiffuser Overshooting sampler (without attention modulation)
Builds on: rectified flow Karras stochastic sampling classifier-free guidance
Closest papers by meaning
- HIFA: High-fidelity Text-to-3D Generation with Advanced Diffusion Guidance (ICLR 2024 · similarity 0.88)
- InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · similarity 0.86)
- DreamText: High Fidelity Scene Text Synthesis (CVPR 2025 · similarity 0.86)
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.85)
- Diff-Aid: Inference-time Adaptive Interaction Denoising for Rectified Text-to-Image Generation (NeurIPS 2026 · similarity 0.85)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.85)
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.