Facial expression image editing requires fine-grained control to strictly preserve human identity and background while precisely manipulating expression. However, existing editing benchmarks primarily focus on general scenarios, lacking high-quality facial images and corresponding editing instructions. Furthermore, current evaluation metrics exhibit systemic biases in this task, often favoring lazy editing or overfit editing. To bridge these gaps, we propose FED-Bench, a comprehensive benchmark featuring rigorous testing and an accurate evaluation suite. First, we carefully construct a benchmark of 747 triplets through a cascaded and scalable pipeline, each comprising an original image, an editing instruction, and a ground-truth image for precise evaluation. Second, we introduce FED-Score, a cross-granularity evaluation protocol that disentangles assessment into three dimensions: Alignment for verifying instruction following, Fidelity for testing image quality and identity preservation, and Relative Expression Gain for quantifying the magnitude of expression changes, effectively mitigating the aforementioned evaluation biases. Third, we benchmark 18 image editing models, revealing that current approaches struggle to simultaneously achieve high fidelity and accurate expression manipulation, with fine-grained instruction following identified as the primary bottleneck. Finally, leveraging the scalable characteristic of introduced benchmark engine, we provide a 20k+ in-the-wild facial training set and demonstrate its effectiveness by fine-tuning a baseline model that achieves significant performance gains. Our benchmark and related code will be made publicly open soon.
Figures & tables
Figure 1: The overview of FED-Bench. The figure illustrates representative samples from our benchmark covering seven basic emotions.
Figure 2: The overall illustrations of FED-Bench construction pipeline.
Table 3
Table 3: Human alignment test. We report the accuracy between each metric’s preference and the human consensus under a 2AFC protocol. F : computed on the face region; GT : compared with the ground truth image.
Dense Instructions
Simple Instructions
Method
ID
BG
PQ
SC
GTA
REG
Score
Method
ID
BG
PQ
SC
GTA
REG
Score
Qwen-image-edit-plus
.58
17.5
9.7
8.8
5.7
1.18
.469
Qwen-image-edit-plus
.63
16.8
9.8
8.8
5.8
1.13
.492
SeedDream 4.0 [ 29 ]
.62
15.5
9.5
9.1
4.3
1.37
.379
SeedDream 4.0
.69
16.4
9.7
8.0
5.0
1.26
.413
FLUX.2 Pro [ 17 ]
.58
13.9
9.8
9.0
4.0
1.37
.377
FLUX.2 Pro
.58
14.0
9.6
8.8
5.2
1.37
.400
Qwen-image-edit
.45
19.6
9.3
8.9
4.2
1.37
.337
Qwen-image-edit-2511
.44
15.3
9.8
9.4
4.0
1.42
.361
FLUX-Kontext-FED
.68
7.3
9.7
6.2
3.4
0.95
.332
Qwen-image-edit
.43
19.9
9.6
8.7
4.4
1.37
.343
Table 4: Benchmarking results on FED-Bench. Left: Dense instructions (ranked by Dense FED-Score). Right: Simple instructions (ranked by Simple FED-Score). We report raw mean values for sub-metrics and normalized FED-Score. BG ↓ : lower is better. REG: optimal at 1.0. Bold : best; underline : second best.
Figure 3: The qualitative comparison of facial expression editing on our FED-Bench. Each row shows a different editing task with the original image, ground truth target, and results from part of evaluated models. From left to right: original image, ground truth, followed by the evaluation results.
Fine-grained facial expression editing has long been limited by intrinsic semantic overlap. To address this, we construct the Flex Facial Expression (FFE) dataset with continuous affective annotations and establish FFE-Bench to evaluate structural confusion, editing accuracy, linear controllability, and the trade-off between expression editing and identity preservation. We propose PixelSmile, a diffusion framework that disentangles expression semantics via fully symmetric joint training. PixelSmile combines intensity supervision with contrastive learning to produce stronger and more distinguishable expressions, achieving precise and stable linear expression control through textual latent interpolation. Extensive experiments demonstrate that PixelSmile achieves superior disentanglement and robust identity preservation, confirming its effectiveness for continuous, controllable, and fine-grained expression editing, while naturally supporting smooth expression blending.
Recent advances in image editing have enabled models to handle complex instructions with impressive realism. However, existing evaluation frameworks lag behind: current benchmarks suffer from narrow task coverage, while standard metrics fail to adequately capture visual consistency, i.e., the preservation of identity, structure and semantic coherence between edited and original images. To address these limitations, we introduce GEditBench v2, a comprehensive benchmark with 1,200 real-world user queries spanning 23 tasks, including a dedicated open-set category for unconstrained, out-of-distribution editing instructions beyond predefined tasks. Furthermore, we propose PVC-Judge, an open-source pairwise assessment model for visual consistency, trained via two novel region-decoupled preference data synthesis pipelines. Besides, we construct VCReward-Bench using expert-annotated preference pairs to assess the alignment of PVC-Judge with human judgments on visual consistency evaluation. Experiments show that our PVC-Judge achieves state-of-the-art evaluation performance among open-source models and even surpasses GPT-5.1 on average. Finally, by benchmarking 16 frontier editing models, we show that GEditBench v2 enables more human-aligned evaluation, revealing critical limitations of current models, and providing a reliable foundation for advancing precise image editing.
Zhangqi Jiang, Zheng Sun, Xianfang Zeng +7
Nanyang Technological University · StepFun · Southeast University
Fine-grained image editing requires more than producing a visually plausible result: an editor must execute the requested attribute change precisely while leaving everything else intact. However, existing benchmarks leave a critical gap between realism and verifiability: benchmarks built on realistic images typically rely on human or vision--language model judgments, while deterministic evaluation has largely focused on synthetic shape canvases, with application-oriented extensions primarily limited to charts. This makes it difficult to determine precisely how much of a requested edit was executed, where unintended changes occurred, and whether small differences between models reflect genuine editing capability or evaluator uncertainty. To bridge this gap, we present VeriEdit-Bench, a benchmark for fine-grained, instruction-faithful image editing across realistic structured assets with deterministic, four-axis evaluation. Its 1,740 cases are compiled from the source code of 153 Scalable Vector Graphics (SVG) graphics, charts, web interfaces, and presentation slides. Controlled source-code edits preserve the original visual context while yielding exact target images, pixel-level edit masks, and explicit edit specifications, enabling reproducible scoring along four axes: edit fidelity, preservation, localization, and magnitude. Evaluating eleven editors, we find that even the strongest model remains far from full credit; rankings for the same recoloring operation reverse between charts and SVG graphics; and outputs with similar pixel-accuracy profiles can still differ substantially in localization and change magnitude. This decomposition yields graded, verifiable feedback and exposes model-specific capability and failure profiles that holistic scores or evaluator-dependent judgments may obscure.