cs.CVApr 3, 2026
SaveExpressEdit: Fast Editing of Stylized Facial Expressions with Diffusion Models in Photoshop
Organizations: University of California, Santa Barbara
Abstract
Facial expressions of characters are a vital component of visual storytelling. While current AI image editing models hold promise for assisting artists in the task of stylized expression editing, these models introduce global noise and pixel drift into the edited image, preventing the integration of these models into professional image editing software and workflows. To bridge this gap, we introduce ExpressEdit, a fully open-source Photoshop plugin that is free from common artifacts of proprietary image editing models and robustly synergizes with native Photoshop operations such as Liquify. ExpressEdit seamlessly edits an expression within 3 seconds on a single consumer-grade GPU, significantly faster than popular proprietary models. Moreover, to support the generation of diverse expressions according to different narrative needs, we compile a comprehensive expression database of 135 expression tags enriched with example stories and images designed for retrieval-augmented generation. We open source the code and dataset to facilitate future research and artistic exploration.
Figures & tables
Figure 2: ExpressEdit consists of two consecutive pipelines for a user-friendly yet professional editing experience. The prompt generation pipeline takes in a story paragraph and uses a VLM to retrieve relevant expression tags from a multi-modal expression database we curate. The relevant expression tags are inserted into a prompt, which is used in the image editing pipeline. The image editing pipeline starts by the user applying coarse transformations (such as Liquify) and casual selections on the original image, taking at most a few seconds of manual effort. Then, combined with the prompt, ExpressEdit robustly generates high-quality expressions based on the inputs.
Figure 3: Creative short stories are included in ExpressEdit to assist retrieval-augmented generation (RAG). The prompt template and an example story are shown here.
Figure 4: Baseline methods introduce destructive noise into the original image after each editing step, whereas the pixel changes from ExpressEdit are non-destructive, and the minor artifacts are easily repaired. Please see Section 3.2 for details.
| Step | Prompt for Baseline Models | Prompt for ExpressEdit (Prefix and Suffix Omitted) |
|---|---|---|
| 1 | Change the color of her left eye to green | green eye |
| 2 | Change the color of her right eye to blue | blue eye |
| 3 | Make her smile | green eye, blue eye, smile |
| 4 | Change the color of her bow to pink | pink bow |
| 5 | Make her bangs blunt | green eye, blue eye, blunt bangs |
| 6 | Add a yellow star sticker under her right eye | blue eye, yellow star sticker |
Table 1: To iteratively edit one image, ExpressEdit accepts succinct but informative prompts. For ExpressEdit , the prompt suffix and prefix ( Figure 2(b) ) are kept fixed, and the tags can be updated minimally to reflect any change on the character. If the selection potentially covers more than one facial element, additional descripions can be added to stabilize the results. For example, when editing the bangs to be blunt in Step 5, the selection may touch the eyes, so adding eye color tags helps prevent the eye color from changing. Notably, no description of composition is needed. For each step, the region of interest is indicated by native Photoshop operations, instead of by text. The prompt for ExpressEdit also does not need to strictly follow a set of existing tags, but the tag format empirically leads to better results.
Figure 5: The selection version of Nano Banana Pro in Photoshop and the naive inpainting both create visible artifacts around the selected region. Nano Banana Pro leaves artifacts around earlobes and the chin, making it impossible to contain the destructive noise via selecting. Naive inpainting also leaves artifacts on the right side of the neck and on the braid, necessitating the use of the SPICE backend.
Figure 6: ExpressEdit can generate detailed expressions even with challenging, high-resolution inputs, whereas Nano Banana 2 Pro generates lower quality faces, despite its support for 2K image generation. Nano Banana 2 Pro generates a face with blurry contours, creates visible artifacts around the red string, lowers the saturation, and arbitrarily changes the shape of the face.
Figure 7: ExpressEdit precisely follows the instruction (reducing the diameter of the irises by 50%) with a simple transformation in Photoshop as a hint to the diffusion model. To reduce the iris size, the user only needs to select the irises, transform their size, and leave the holes from transformation in a white color. There is no need to manually re-draw the shadows on the eyeball, as ExpressEdit automatically fixes the gap. While given a clear instruction specifying the percentage of size change, all baseline models fail on this task. ExpressEdit , however, correctly generates the result, even without the numeric number or expression tags in the prompt. Only the prompt prefix for the character and the prompt suffix for the style were used.
Figure 8: ExpressEdit conveniently fixes artifacts from manual editing. The first row shows the results from the Liquify tool in Photoshop, and the second row shows the fixed results. Liquify supports high editing flexibility at the notorious cost of long manual editing time. A casual use of the tool introduces heavy deformations on the white bow or on the iris, but the deformation can be quickly fixed.
Figure 9: Besides expressions, ExpressEdit can fix other details on a character. The correct design of the character includes 4 bows with interleaving blue and yellow colors [ 72 ] . Hinted by simple sketches, ExpressEdit fixes the design in a few seconds.
| Method | Latency (s) |
|---|---|
| FLUX.2 [max] | 49.94 13.39 |
| GPT | 46.01 11.74 |
| Grok | 7.11 0.50 |
| Nano Banana 2 | |
| Fast (without Reasoning) | 23.18 3.92 |
| Pro (with Reasoning) | 41.92 22.08 |
Table 2: Besides its high quality, ExpressEdit also achieves lowest inference latency using a single consumer-grade GPU. The latencies are evaluated on an 1024 1024 image. This table shows mean standard deviation over 10 editing runs. The latency was measured from clicking Generate to the image appearing. Due to the open source nature of ExpressEdit , we were able to test it under a stable environment without the interference of other resource-intensive software, resulting in small standard deviations. An additional overhead of 1 to 2 seconds should be expected if the user uses slower devices, or if the user uses other software for painting assistance at the same time. Still, ExpressEdit is the fastest, and it generates the cleanest results ( Section 3.2 ) without any API cost.
Figure 10: Using a speed-up (lightning) LoRA dramatically cuts the latency to 2.18 seconds, with negligible impact on the level of details. ExpressEdit with a lighting LoRA achieves 46% lower latency, at the cost of small artifacts on eyelashes.
More in this topic
Image Editing
Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.
See all 315 papers