acceptodds
Withdrawn from ICLR 2027

WithEveryone: Unified Planning and Identity Grounding for Group Image Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Identity-preserving image generation becomes increasingly unreliable when a scene must contain many specified people. Beyond retaining each identity, the model must bind every reference to a distinct person and location, while training-time identity losses must establish correspondence among several noisy predicted faces. We introduce WithEveryone, a unified framework for generating group images up to ten reference identities. WithEveryone injects each selected identity as an addressed token, predicts a structured identity–layout plan, and renders the plan as a visual condition. Its key objective, Layout-Grounded ID Loss, uses annotated face regions to supervise the intended identities directly, avoiding unstable embedding-based face matching; ID Representation Forcing additionally trains a prediction for each identity before image synthesis. On an identity-disjoint benchmark, WithEveryone achieves the highest target-context identity similarity, improving Sim(Tgt) from 0.462 for GPT-Image 2 to 0.499, while reducing copy-paste artifacts from 0.169 to 0.055. It further covers 97.3% of the requested identities with a duplicate rate of only 2.8%. These results show that explicit identity–layout grounding enables identity-preserving generation to scale to larger groups without relying on direct reference-face copying.

Primary area:
generative models
Keywords:
AIGC, ID-preserving generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jYkyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on group image generation involving 5–10 individuals, rather than personalized generation for a single person or a small number of individuals. By integrating ID tokens, Representation Forcing, identity-aware Layout, CoT, and LG-ID Loss, the proposed method achieves the best Sim(Tgt) while maintaining a lower Copy-Paste score. Furthermore, the experimental results demonstrate the effectiveness of the proposed LG-ID Loss.

Critical strengths
  1. The paper addresses a practically valuable task. Personalized image generation involving multiple individuals, particularly five or more, remains relatively underexplored in existing research.
  2. The paper is well-written, clearly structured, and easy to follow.
  3. The main experiments include comparisons with both open-source and proprietary models, making the evaluation comprehensive and well-designed. The ablation studies are also thorough and effectively demonstrate the effectiveness of the proposed LG-ID Loss.
Critical weaknesses
  1. Celebrities can already be specified through text prompts, and GPT-Image 2 can generate reasonably good results. What are the advantages of the proposed method in this scenario? Is the method limited to generating celebrities, or does the benchmark also include ordinary individuals?
  2. When an identity is specified through a reference image, can other identities still be specified through text? For example, can the method generate a photo of <ID> and Elon Musk, where one person is specified by an ID reference image and Elon Musk is specified through text?
  3. How is the aesthetic quality of the generated images after identity customization? Are there any quantitative metrics to evaluate this aspect?
  4. Which model is used to predict human keypoints? How reasonable and accurate are the predicted human structures, and has this been validated? Can the model generate anatomically plausible human bodies using only bounding boxes? A visualization-based ablation study might help clarify this.
Questions for authors
  1. Some implementation details are missing. Equation (3) does not explain what L_ID and L_NTP represent or provide their mathematical formulations. Similarly, the formulation of LG-ID Loss appears to be missing.
  2. What is the demographic distribution of the benchmark constructed from 210 real-world groups? For example, does it consist entirely of celebrities, or does it also include ordinary individuals?
  3. The paper does not clearly specify the backbone model or codebase used in the proposed method. Is it built upon WithAnyone? If so, are the main modifications the introduction of LG-ID Loss and Layout CoT?
  4. What is the difference between Sim(Ref) and Sim(Tgt)?
Reviewer_YEamOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper developed a method WithEveryone to generate group images of 5–10 people by binding each identity to a layout plan, improving similarity while eliminating copy-paste artifacts.

Critical strengths
  1. The paper addresses a meaningful problem: generating coherent images containing five to ten specified identities. Integrating identity conditioning, layout planning, and synthesis is well motivated.
  2. The reported results are promising: Sim(Tgt) improves from 0.462 to 0.499 over GPT-Image 2, with higher coverage and lower Copy-Paste scores. Identity-disjoint evaluation and paired statistical analysis strengthen the empirical evidence.
Critical weaknesses
  1. The novelty of LG-ID remains unclear. WithAnyone already uses target-image landmarks to align generated and target faces before computing an ArcFace identity loss (§5.1, Eqs. 4–6). The manuscript needs to establish what LG-ID adds beyond extending this supervision to more identities and integrating it with planning.

  2. The complete system’s gains are insufficiently isolated. In Table 2, LG-ID alone achieves 0.435 Sim(Tgt), versus 0.461 for the complete model. However, the latter simultaneously adds identity tokens, Representation Forcing, Layout CoT, and additional layout data. Although acknowledged by the authors, this confounding limits conclusions about the individual contributions and complementarity of these components.

Questions for authors

Identity-position binding is not directly validated. The described evaluation measures identity similarity through global matching, geometric agreement through Plan IoU, and plan quality through RLS. These separate measurements may not detect correct identities generated at exchanged locations. This leaves a central grounding claim insufficiently supported.

Reviewer_7CJvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

My understanding of the paper is as follows:

  • The paper tackles the problem of generating 5-10 people in the same scene(multi-human gen).
  • The architecture used is a BAGEL-style unified architecture with reasoning and diffusion in the same transformer.
  • The paper proposes a benchmark of 210 real images with 5-10 identities.
  • The training data is ~400K samples along with layout reasoning.
  • The paper proposes Identity-Aware CoT, which means the same model with context of input identities reasons the spatial plan and also denoises the image.
  • The training uses NTP loss for reasoning+spatial layouts.
  • Two loss terms are proposed by the authors, Representation-Forcing(predict the ArcFace embeddings of each face) as well as Layout Grounded ID loss (Arcface of the x0 pred image in locations where the model reasoned about the faces).
Critical strengths

The strengths of this work are as follows:

Motivation: The whole system is well motivated. Multi-Human generation is a difficult problem and of interest to the community. The benchmark is necessary as current benchmarks don't measure more than 5 faces. The joint architecture which reasons spatial plan and generates the image is intuitive and connected to the motivation.

Clarity/Presentation: The paper reads well. Figures 2,3,4 are self-explanatory and well annotated. The method is presented in a way that is easy to understand.

Unified Setting: The unified architecture setting is developed well. Based on the results on the proposed benchmark, it seems to work very well and also seems like the correct architecture choice. Additionally, the Representation Forcing loss is highly intuitive and novel. There is high scope of usability for this system moving forward.

Critical weaknesses

[W1] Visual Results: The results in Figure 5 for WithEveryone seems sub-par. For instance, the first image has some identity blending artifacts (also seems like the faces are unrecognizable) and the second row image has distorted text. It seems like the quality of generation is poor. This is also contradicting with the images in Figure 9 which are mostly selfie images. In this case, could the authors clarify if the method is general (can generate any people doing any actions) or only meant for portrait-style or selfie style images. If only meant for selfie style images, this needs to be explicitly mentioned in text.

[W2] Weak Evaluation: The evals focus on face similarity, but metrics such as DINO and CLIP are sub-par for measuring quality of generation, realism and also prompt-action alignment. For measuring this axis, metrics such as HPSv3<1> are much better and can complement CLIP and DINO. Additionally, it seems that there is no report of action categories or person interactions in the benchmark. If the prompts don't contain actions for people, the authors are suggested to evaluate their model on other benchmarks which explicitly probe for and also measure action alignment<2>.

[W3] Coverage: Apart from coverage of person action/interaction prompts, the benchmark does not report any age, gender, ethnicity of it's human subjects. This is crucial to mitigate biases in our evaluations towards certain demographics of faces. If the data is highly skewed, authors are suggested to report this.

[W4] LG-ID Formulation: The LG-ID formulation was motivated as it was deemed better than traditional Hungarian matching of identities due to two reasons, (a)Sensitivity of matching identities in a noisy setting and (b)Increased sensitivity as the number of faces increase. However, there is no ablation study to justify this formulation and if it actually works better than hungarian matching as claimed in motivation. Additionally, a recent work<3> used hungarian matching of spatial centroids instead of ID embeddings themselves, which alleviates the sensitivity of at high noise. They also argue that using spatial crops(similar to witheveryone) directly could lead to facial distortions as this is a hard constraint.

The main reason this is a critical weakness is that this specific contribution of LG-ID loss in it's current form does not justify its motivation.

<1> https://arxiv.org/abs/2508.03789

<2> https://arxiv.org/abs/2506.20879

<3> https://arxiv.org/abs/2511.22690

Questions for authors

[Q1] Is there a bias to selfie style images in the results? Are other natural scenes not being generated well?

[Q2] Have the authors tried to use a quality loss or prompt alignment loss term for the generated image, which might help improve quality in W1?

[Q3] What would happen if the model generates an imperfect layout?

[Q4] Spatial reasoning would be poor at the start of training. In this case, would LG-ID loss break?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.