# SOAR: Self-Correction for Optimal Alignment and Refinement in Diffusion Models

arXiv: https://arxiv.org/abs/2604.12617 · HTML: https://mlsanity.com/paper/2604.12617/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- Placed in Diffusion and flow policies for robotics, one of 14 directions in Diffusion models.
- Builds on 28 of its 31 references that are in the map.
- First posted on arXiv 14 Apr 2026.

## Placed in

- Diffusion models / [Diffusion and flow policies for robotics](https://mlsanity.com/diffusion/diffusion-and-flow-policies-for-robotics/)
- LLM alignment / RLHF / preference optimisation / [Diffusion and multimodal model alignment](https://mlsanity.com/alignment/diffusion-and-multimodal-model-alignment/)

## Builds on (most cited first, 28 of 28)

- [Denoising Diffusion Probabilistic Models](https://mlsanity.com/paper/2006.11239/) (2020)
- [High-Resolution Image Synthesis with Latent Diffusion Models](https://mlsanity.com/paper/2112.10752/) (2021)
- [Score-Based Generative Modeling through Stochastic Differential Equations](https://mlsanity.com/paper/2011.13456/) (2020)
- [Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding](https://mlsanity.com/paper/2205.11487/) (2022)
- [Scalable Diffusion Models with Transformers](https://mlsanity.com/paper/2212.09748/) (2022)
- [Flow Matching for Generative Modeling](https://mlsanity.com/paper/2210.02747/) (2022)
- [SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis](https://mlsanity.com/paper/2307.01952/) (2023)
- [Scaling Rectified Flow Transformers for High-Resolution Image Synthesis](https://mlsanity.com/paper/2403.03206/) (2024)
- [Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow](https://mlsanity.com/paper/2209.03003/) (2022)
- [HunyuanVideo: A Systematic Framework For Large Video Generative Models](https://mlsanity.com/paper/2412.03603/) (2024)
- [ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation](https://mlsanity.com/paper/2304.05977/) (2023)
- [Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis](https://mlsanity.com/paper/2306.09341/) (2023)
- [Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks](https://mlsanity.com/paper/1506.03099/) (2015)
- [Diffusion Model Alignment Using Direct Preference Optimization](https://mlsanity.com/paper/2311.12908/) (2023)
- [Flow-GRPO: Training Flow Matching Models via Online RL](https://mlsanity.com/paper/2505.05470/) (2025)
- [Human Preference Score: Better Aligning Text-to-Image Models with Human Preference](https://mlsanity.com/paper/2303.14420/) (2023)
- [DiffusionNFT: Online Diffusion Reinforcement with Forward Process](https://mlsanity.com/paper/2509.16117/) (2025)
- [Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs](https://mlsanity.com/paper/2401.11708/) (2024)
- [Optimizing Prompts for Text-to-Image Generation](https://mlsanity.com/paper/2212.09611/) (2022)
- [T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT](https://mlsanity.com/paper/2505.00703/) (2025)
- [Self-correcting LLM-controlled Diffusion Models](https://mlsanity.com/paper/2311.16090/) (2023)
- [Elucidating the Exposure Bias in Diffusion Models](https://mlsanity.com/paper/2308.15321/) (2023)
- [Input Perturbation Reduces Exposure Bias in Diffusion Models](https://mlsanity.com/paper/2301.11706/) (2023)
- [Improving Text-to-Image Consistency via Automatic Prompt Optimization](https://mlsanity.com/paper/2403.17804/) (2024)
- [SimpleAR: Pushing the Frontier of Autoregressive Visual Generation through Pretraining, SFT, and RL](https://mlsanity.com/paper/2504.11455/) (2025)
- [Dynamic Prompt Optimizing for Text-to-Image Generation](https://mlsanity.com/paper/2404.04095/) (2024)
- [PromptEnhancer: A Simple Approach to Enhance Text-to-Image Models via Chain-of-Thought Prompt Rewriting](https://mlsanity.com/paper/2509.04545/) (2025)
- [RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning](https://mlsanity.com/paper/2505.17540/) (2025)

## Built on by (most cited first, 6 of 6)

- [D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models](https://mlsanity.com/paper/2605.05204/) (2026)
- [Continuous-Time Distribution Matching for Few-Step Diffusion Distillation](https://mlsanity.com/paper/2605.06376/) (2026)
- [dots.tts Technical Report](https://mlsanity.com/paper/2606.07080/) (2026)
- [ReNFT: Repairing Mode Collapse in Reward Post-Training via Internal Probability-Mass Recalibration](https://mlsanity.com/paper/2609.00061/) (2026)
- [TRACE-GS: On-Policy Trajectory Distillation with Privileged Geometric Conditioning for Sparse-View 3DGS Restoration](https://mlsanity.com/paper/2608.10286/) (2026)
- [ReCurveflow: A Flow Matching Framework that Learns Curved Reaction Trajectories to Predict Transition State Geometries](https://mlsanity.com/paper/2608.20869/) (2026)

## Abstract (arXiv)

The post-training pipeline for diffusion models currently has two stages: supervised fine-tuning (SFT) on curated data and reinforcement learning (RL) with reward models. A fundamental gap separates them. SFT optimizes the denoiser only on ground-truth states sampled from the forward noising process; once inference deviates from these ideal states, subsequent denoising relies on out-of-distribution generalization rather than learned correction, exhibiting the same exposure bias that afflicts autoregressive models, but accumulated along the denoising trajectory instead of the token sequence. RL can in principle address this mismatch, yet its terminal reward signal is sparse, suffers from credit-assignment difficulty, and risks reward hacking. We propose SOAR (Self-Correction for Optimal Alignment and Refinement), a bias-correction post-training method that fills this gap. Starting from a real sample, SOAR performs a single stop-gradient rollout with the current model, re-noises the resulting off-trajectory state, and supervises the model to steer back toward the original clean target. The method is on-policy, reward-free, and provides dense per-timestep supervision with no credit-assignment problem. On SD3.5-Medium, SOAR improves GenEval from 0.70 to 0.78 and OCR from 0.64 to 0.67 over SFT, while simultaneously raising all model-based preference scores. In controlled reward-specific experiments, SOAR surpasses Flow-GRPO in final metric value on both aesthetic and text-image alignment tasks, despite having no access to a reward model. Since SOAR's base loss subsumes the standard SFT objective, it can directly replace SFT as a stronger first post-training stage after pretraining, while remaining fully compatible with subsequent RL alignment.