arXiv:2511.18834v1 Announce Type: new 
Abstract: With the success of flow matching in visual generation, sampling efficiency remains a critical bottleneck for its practical application. Among flow models' accelerating methods, ReFlow has been somehow overlooked although it has theoretical consistency with flow matching. This is primarily due to its suboptimal performance in practical scenarios compared to consistency distillation and score distillation. In this work, we investigate this issue within the ReFlow framework and propose FlowSteer, a method unlocks the potential of ReFlow-based distillation by guiding the student along teacher's authentic generation trajectories. We first identify that Piecewised ReFlow's performance is hampered by a critical distribution mismatch during the training and propose Online Trajectory Alignment(OTA) to resolve it. Then, we introduce a adversarial distillation objective applied directly on the ODE trajectory, improving the student's adherence to the teacher's generation trajectory. Furthermore, we find and fix a previously undiscovered flaw in the widely-used FlowMatchEulerDiscreteScheduler that largely degrades few-step inference quality. Our experiment result on SD3 demonstrates our method's efficacy.

تم تقديم طريقة جديدة تُدعى FlowSteer لتحسين كفاءة توليد الصور المستندة إلى ReFlow من خلال توجيه نموذج الطالب على طول مسارات التوليد الأصلية للمعلم. يتناول هذا التطور مشكلة التباين الحرج في التوزيع التي لوحظت أثناء التدريب، والتي كانت تعيق سابقًا أداء Piecewised ReFlow.

Se ha introducido un nuevo método llamado FlowSteer para mejorar la eficiencia de la síntesis de imágenes basada en ReFlow, guiando al modelo estudiante a lo largo de las trayectorias de generación auténticas del maestro. Este desarrollo aborda la discrepancia crítica de distribución observada durante el entrenamiento, que anteriormente obstaculizaba el rendimiento de Piecewised ReFlow.

Une nouvelle méthode appelée FlowSteer a été introduite pour améliorer l'efficacité de la synthèse d'images basée sur ReFlow en guidant le modèle étudiant le long des trajectoires de génération authentiques de l'enseignant. Ce développement répond au problème critique de décalage de distribution observé pendant l'entraînement, qui a précédemment entravé les performances de Piecewised ReFlow.

A new method named FlowSteer has been introduced to enhance the efficiency of ReFlow-based image synthesis by guiding the student model along the teacher's authentic generation trajectories. This development addresses the critical distribution mismatch observed during training, which has previously hindered the performance of Piecewised ReFlow.

FlowSteer: Guiding Few-Step Image Synthesis with Authentic Trajectories

arXiv:2512.08125v1 Announce Type: cross 
Abstract: Flow-based text-to-image (T2I) models excel at prompt-driven image generation, but falter on Image Restoration (IR), often "drifting away" from being faithful to the measurement. Prior work mitigate this drift with data-specific flows or task-specific adapters that are computationally heavy and not scalable across tasks. This raises the question "Can't we efficiently manipulate the existing generative capabilities of a flow model?" To this end, we introduce FlowSteer (FS), an operator-aware conditioning scheme that injects measurement priors along the sampling path,coupling a frozed flow's implicit guidance with explicit measurement constraints. Across super-resolution, deblurring, denoising, and colorization, FS improves measurement consistency and identity preservation in a strictly zero-shot setting-no retrained models, no adapters. We show how the nature of flow models and their sensitivities to noise inform the design of such a scheduler. FlowSteer, although simple, achieves a higher fidelity of reconstructed images, while leveraging the rich generative priors of flow models.

تم تقديم FlowSteer (FS) كخطة جديدة للتهيئة الواعية بالعمليات تهدف إلى تحسين مهام استعادة الصور مثل تحسين الدقة، وإزالة الضباب، وإزالة الضوضاء، وتلوين الصور. تعمل هذه الطريقة على تحسين اتساق القياسات والحفاظ على الهوية دون الحاجة إلى نماذج معاد تدريبها أو محولات محددة للمهام، مما يمثل تقدمًا كبيرًا في نماذج توليد الصور المعتمدة على التدفق.

FlowSteer (FS) ha sido introducido como un nuevo esquema de acondicionamiento consciente del operador destinado a mejorar tareas de restauración de imágenes como super-resolución, desenfoque, eliminación de ruido y colorización. Este método mejora la consistencia de las mediciones y la preservación de la identidad sin necesidad de modelos reentrenados o adaptadores específicos para tareas, marcando un avance significativo en los modelos de generación de imágenes basados en flujo.

FlowSteer (FS) a été introduit comme un nouveau schéma de conditionnement conscient de l'opérateur visant à améliorer les tâches de restauration d'images telles que la super-résolution, le défloutage, la débruitage et la colorisation. Cette méthode améliore la cohérence des mesures et la préservation de l'identité sans nécessiter de modèles réentraînés ou d'adaptateurs spécifiques aux tâches, marquant une avancée significative dans les modèles de génération d'images basés sur le flux.

FlowSteer (FS) has been introduced as a novel operator-aware conditioning scheme aimed at improving image restoration tasks such as super-resolution, deblurring, denoising, and colorization. This method enhances measurement consistency and identity preservation without the need for retrained models or task-specific adapters, marking a significant advancement in flow-based text-to-image models.

FlowSteer: Guiding Few-Step Image Synthesis with Authentic Trajectories

Was this article worth reading? Share it

RepoFlow

Metaflow AI

Summiflow