FlowMimic: Mask-free Visual Editing and Generation with Pixel-pair Warped Flow Fields

Forum topic · 小凯 · 2026-07-22

Summary

FlowMimic (arXiv:2607.18227, cs.CV) by Dingyun Zhang, Lixue Gong, and Wei Liu explores unifying video and image generation and editing within a single model. The paper addresses the scarcity of video editing training data, since existing pipelines rely on labor-intensive curation involving object mask annotation, I2V models with ControlNet-style guidance (which introduces erroneous pairs), and VLM-based quality filtering, limiting task diversity. The authors propose a pixel-pair temporal warped flow field that converts image editing samples into corresponding video editing samples in real time, and show that models can learn video editing across multiple task levels using only such data. Treating the image modality as a special case of video, they design modality-mimic generation and editing losses that align the capabilities and output distributions of both modalities through mutual imitation.

Overview

Field: Computer Vision (cs.CV) Authors: Dingyun Zhang, Lixue Gong, Wei Liu Published: 2026-07-20 arXiv: 2607.18227

Key Points

FlowMimic integrates video and image modalities — both generation and editing — within a single model, following the mainstream direction of visual research.
Current approaches to collecting video editing data are labor-intensive and time-consuming, relying on:
Object mask annotation.
I2V models with ControlNet-style guidance, which introduce erroneous paired synthesis.
VLM-based quality filtering or refinement.
These pipelines have limited task scalability, so video editing task diversity falls far below what is available for image editing models.
The authors develop a pixel-pair temporal warped flow field that can generate corresponding video editing samples from image editing samples in real time.
Experiments across multiple levels of video editing tasks show that models can learn video editing using only such generated data — no masks or curated video editing data required.
The image modality is treated as a special form of the video modality.
Based on this, the paper designs modality-mimic generation loss and modality-mimic editing loss, which align the capabilities and output distributions of the two modalities through mutual imitation.

Original Abstract (Condensed)

> We develop a pixel-pair temporal warped flow field for real-time video editing data generation from image editing samples. A modality mimic generation and editing loss aligns capabilities between image and video modalities through mutual imitation.

Links: arXiv:2607.18227

*Auto-collected on 2026-07-22.*