多模態生成
WithEveryone Generates Group Portraits of Up to Ten People by Binding Identities to Layouts, but Weights Remain Unavailable
A team from Tencent Hunyuan and Fudan University has introduced WithEveryone, which first binds each reference person to a structured layout and then supervises generation using a regionalized identity loss. In the authors’ identity-disjoint evaluation, it improves subject coverage and reduces copy-and-paste artifacts, but the license of the existing base model prevents the team from releasing the research-version weights.
The challenge of multi-person reference-image generation is not merely whether each face looks right. It also involves ensuring that every identity appears in the correct position and determining how to match multiple predicted faces one-to-one with the reference subjects during training. [WithEveryone](https://arxiv.org/abs/2608.20336) divides the problem into planning and rendering stages: each reference person is first converted into an addressable identity token. The model then predicts an identity–layout plan containing person regions, face regions, and pose information, and uses that plan as a visual condition to generate group portraits of up to ten people.
The core training method, Layout-Grounded ID Loss, avoids matching face embeddings to multiple predicted regions after the fact. Instead, it uses annotated face regions to apply the loss for a specified identity directly at the target location. Another technique, ID Representation Forcing, requires the model to predict each identity representation individually before it begins synthesizing the image, reducing the chance of dropping subjects during long-sequence generation. This explicit binding also allows layout planning, identity preservation, and final rendering to be debugged separately.
In the authors’ evaluation, where subject identities did not overlap with the training set, face similarity in the target scenario increased from GPT-Image-2’s 0.462 to 0.499, while the copy-and-paste artifact metric fell from 0.169 to 0.055. The system covered 97.3% of specified identities, with a duplicate-subject rate of 2.8%. These results suggest that the main bottleneck in multi-person generation may be addressable planning and supervision alignment, rather than simply increasing face-encoder capacity.
However, the [public repository](https://github.com/doby-xu/WithEveryone) currently contains only a description of the method, with no executable code or model weights. The team says the research version was built on a base model whose license prohibits checkpoint redistribution and that it is retraining a releasable version. As a result, speed, memory requirements, failure cases, and evaluation implementation cannot yet be independently verified. Face similarity also does not guarantee correctness in pose, hands, occlusion, or cultural context. The key questions going forward are whether the open version can retain the paper’s results and whether layout control remains stable beyond ten people.