An arXiv preprint reports that WithEveryone, a system designed to generate one coherent group image from five to ten reference identities, performed better than GPT-Image 2 on automated measures of identity preservation in the paper’s main benchmark. The reported comparison also found fewer measured copy-paste artifacts and duplicate identities.
WithEveryone’s target-context identity-similarity score was 0.499, compared with 0.462 for GPT-Image 2. Its copy-paste artifact score was 0.055 versus 0.169. The model covered 97.3% of the reference identities and had a 2.8% duplicate rate.
The gap was consistent across paired examples: WithEveryone had the higher target-context score on 73% of cases. A 20,000-resample bootstrap interval put the estimated advantage between 0.027 and 0.048, while a Wilcoxon test gave p = 3.8 × 10−11.
What the system tries to coordinate
WithEveryone puts identity and composition into the same generation pipeline. It loads selected references as identity tokens, predicts a structured multi-person layout linking people to positions, renders that layout as a visual condition and uses the shared context for flow-based image generation.
The benchmark contained 210 real group-image examples with identities disjoint from training. It was split into 60 examples with five references, 50 with six, 40 with seven, 30 with eight, 20 with nine and 10 with ten. Each target had the same number of detected faces as references and no duplicate identities.
For the main comparison, the paper averaged results from ArcFace, FaceNet and AdaFace—automated face-recognition encoders—and matched generated faces to the references. WithEveryone was evaluated at 2K resolution in that comparison and at 1K in other reported tests; ground-truth layouts were supplied to two comparison systems.
Performance held up, but not perfectly
In an ArcFace scaling analysis, WithEveryone ranked first at every group size. Its similarity score fell from 0.629 with five references to 0.571 with ten; GPT-Image 2 fell from 0.593 to 0.496, and UMO from 0.412 to 0.330. Because the largest group-size strata were the smallest, these per-size figures are trends rather than precise estimates.
In the layout ablation, the model-predicted layout variant had reference-identity similarity of 0.364, versus 0.339 for the default; its Count score was 0.828 versus 0.771, and coverage was 0.813 versus 0.741. A ground-truth layout reached 0.412 on reference similarity and 0.367 on target-context similarity, but it used information from the target image and served as an oracle.
In an identity-supervision ablation, the version with LG-ID Loss—computed from corresponding predicted and target face crops using face boxes and landmarks—had reference similarity of 0.506 versus 0.339 for the default and target-context similarity of 0.435 versus 0.304. The Full configuration reached 0.555 and 0.461, but it changed several components and the training corpus at once, so the extra gain cannot be assigned to LG-ID Loss alone.
The model’s own layout plan recorded Plan IoU of 0.773, while the full-model training analysis ended at 0.814. Since target-derived layouts still produced better identity and coverage results, the authors interpret the gap as indicating that plan prediction remains an important source of error.
A narrow test of a difficult problem
The study’s conclusions are limited to one benchmark of 210 examples, and the largest groups have the fewest examples. Its identity scores also depend on face detectors and recognizers whose accuracy may vary across demographic groups.
Cross-model comparisons have other built-in differences: WithEveryone used 2K in the main comparison and 1K elsewhere, while ground-truth layouts were supplied to ID-Patch and WithAnyone. The Full ablation also changed multiple components and the training corpus together, making it impossible to isolate the contribution of one ingredient from that configuration alone.
The paper is a version 1 arXiv preprint dated 20 August 2026. Its automated results do not establish how people would judge identity fidelity, nor do they show that identity-conditioned generation is safe, consensual or resistant to impersonation misuse.
Paper data and sources
Original title: WithEveryone: Unified Planning and Identity Grounding for Group Image Generation
Authors: Hengyuan Xu, Qixun Wang, Yiji Cheng et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text