FlowMimic Turns Image Editing Data into Video Editing Data

Intro

Video editing models are often limited less by architecture than by data. Building curated video editing datasets usually requires mask annotations, synthesized pairs from image-to-video pipelines or ControlNet-like guidance, and extra filtering with vision-language models. That process is costly, slow, and still does not cover as many editing tasks as image editing datasets do. FlowMimic proposes a practical answer: use image editing samples as a base and convert them into video editing training data in a mask-free way.

Key ideas

• Online sample generation from image editing data: a pixel-pair temporal warped flow field is used to produce corresponding video editing samples in real time.
• No mask dependency: the method avoids the labor-heavy annotation step that usually slows down video editing dataset construction.
• One model for image and video: the paper treats images as a special case of video and introduces two losses, modality mimic generation loss and modality mimic editing loss, to align the two modalities through mutual imitation.
• Language-guided localization: instruction-based editing requires understanding the prompt, finding the relevant region in the reference content, and changing only that region. The paper aims to internalize this capability instead of relying on external helpers or explicit mask inputs.

Why it matters

The main contribution of FlowMimic is not just a new training trick. It is a broader way to think about multimodal editing: image editing data can serve as a transferable source for video editing, and the gap between the two can be reduced through shared learning objectives. If this approach scales well, it could lower the cost of collecting video editing data and broaden the range of editable tasks.

It also points to an important design choice for future systems. Rather than adding more external control modules, some capabilities may be better learned inside the model itself, especially region grounding and localized modification for instruction-based editing. That makes the framework interesting for both generation and editing research.

Based on the public abstract, FlowMimic appears to be a methodological proposal rather than a product-style release. Its real significance will depend on the strength of the reported experiments, but the direction is clear: make video editing data easier to obtain, and bring image and video editing closer together.

Source: Hugging Face Daily Papers