🦋 FlowMimic: Mask-free Visual Editing and Generation with Pixel-pair Warped Flow Field for Online Video Editing Data Generation and Modality Mimicry
## ✨ Information
- 🌳 **`2026/08/26`**: The content of Sec. C.9 has been supplemented, and Sec. C.11 has been included in [main.pdf](https://huggingface.co/datasets/FlowMimic/Uncompressed/blob/main/main.pdf).
- 🍂 **`2026/08/04`**: Sec. C.10 has been included in [main.pdf](https://huggingface.co/datasets/FlowMimic/Uncompressed/blob/main/main.pdf), which contains the details of the Jensen–Shannon (JS) divergence and Hellinger distance described in this research.
- ☘️ **`2026/07/21`**: [main.pdf](https://huggingface.co/datasets/FlowMimic/Uncompressed/blob/main/main.pdf) in this dataset repository is the uncompressed file of the corresponding [arXiv file](https://arxiv.org/pdf/2607.18227). Thank you for reading, and have a nice day 🌷
## 🍃 Brief Introduction
Please refer to the paper for details.
We abstract the essence of video editing learning as the temporal maintaining of pixel-level editing effect, propose and model a temporal point-flow warped field to scalably generate video editing data in real time from image editing samples alone, demonstrate the effectiveness of this pixel-aligned paradigm towards learning video editing across diverse-level editing tasks, and reveal its promising extensibility to high-level video tasks such as segmentation-related tasks, low-level video tasks such as video deblurring, and controllable video generation such as condition maps to video.
We posit that a model trained concurrently on both image and video modalities for generation and editing tasks should yield relatively aligned output distributions across the two modalities, as the image modality is essentially a special case of a video modality. Accordingly, we design a modality mimic generation loss and a modality mimic editing loss.
The capacity for language-based visual editing fundamentally requires a model to comprehend the instruction and the reference visual inputs, localize the corresponding region in the reference visual input, and modify it exclusively. Rather than relying on fine-tuning additional modules such as VLMs with connector adaptation or assisted mask sequence inputs, we endow the model with this internalized ability through introducing sense-related tasks and corresponding editing-region-aware latent-level and attention-level losses.
Viewed from the perspective of on-policy self-distillation (OPSD), the mimicry procedure underlying the modality mimicry paradigm, which we explored last April, essentially constitutes a form of OPSD for flow matching models. The editing and generation results of modality mimicry demonstrate that OPSD of flow matching models is indeed feasible. The mimicry methods described in this research—for instance, the modality mimic generation feature score loss or feature loss—can also be applied to the OPSD of image or video flow matching models, such as Z-image and Stable Diffusion 3.5 Medium.
## 🎞️ Video Editing Results
More results of various video editing tasks are demonstrated in the paper—all achieved by the same 1.3B model.
Change Word ("PEACE")Non-rigid Element Replacement and Add Visual Effect
Change Hairstyle ("French short hair in light purple")Video Relighting ("the gentle light of the setting sun")
Temporal Generalization (121 Inference Frames: 3× above the Training Average)
Instruction: "Put pale purple crystalline magical hats on the two men."
Instruction: "Change the lighting to resemble the sunlight of an autumn afternoon."
Instruction: "Add a light purple scarf to the woman."
To conserve inference time, we present results with 121 inference frames for temporal generalization in this paper. We also demonstrate the video editing results at 149 inference frames here, i.e., 4× the training average.
Instruction: "Change the blond woman to a clay blond woman."
Since only four mid-range GPUs were available for this study after the model had been trained, to reduce inference time, the number of inference frames was set to 25 for the video editing experiments except the temporal generalization experiment. Here we present some editing results produced by the model using the original video frame counts. For inference frame counts exceeding GPU memory capacity, we either used the first 121 frames or presented results from 25 frames; the model's editing performance did not degrade with variations in the inference frame count distribution. It should be noted that, in the multi-reference video editing results, the second column corresponds to the reference images; for clarity, we display them throughout the entire MP4 video.
Non-rigid Element Replacement ("UFO")Rigid Element Replacement ("basketball")
Material Modification ("porcelain woman")Color Alteration ("fuchsia")
Multi-reference Material AlterationVideo Virtual Try-on
Change Color Tone ("nostalgic sepia brown")SAM3-like Video Referring Expression Segmentation ("red double-decker bus")
Unseen Task during Training - Subtitle RemovalUnseen Task during Training - Multi-reference Change Hairstyle
Video Facial BeautificationVideo Facial Beautification and Acne Removal
Video Background Replacement
## 🌿 Citation
If you find this research useful for your work, or you would like to build upon the findings or contents described herein, the following BibTeX could be used for citation:
```bibtex
@article{zhang2026flowmimic,
title = {FlowMimic: Mask-free Visual Editing and Generation with Pixel-pair Warped Flow Field for Online Video Editing Data Generation and Modality Mimicry},
author = {Zhang, Dingyun and Gong, Lixue and Liu, Wei},
journal = {arXiv preprint arXiv:2607.18227},
year = {2026}
}
```