WithEveryone: Unified Planning and Identity Grounding for Group Image Generation
Hengyuan Xu, Qixun Wang, Yiji Cheng, Miles Yang, Zhao Zhong, Wei Cheng, Xingjun Ma, Yu-gang Jiang
cs.CV
2026-08-21
WithEveryone generates 5–10-person photos with layout-grounded ID loss. Sim(Tgt) is 0.499 vs 0.462 for GPT-Image 2; copy-paste falls to 0.055; coverage hits 97.3%.
Identity-preserving generation is usable for one or two people and falls apart at five to ten specified faces in one image. Each reference must bind to a distinct person and place; the identity loss must also assign noisy predicted faces to those references. Open methods mostly stop at three to five people; Nano Banana Pro caps at seven. Signals dilute, faces duplicate or merge, and limbs and spacing collapse together.
Prior identity losses recover correspondence with Hungarian matching on face embeddings. Early in training, at high noise, predicted faces look alike, so the assignment is nearly random and each mismatch pulls two identities toward each other. The supervision that should sharpen people cancels itself.
WithEveryone is a unified multimodal model: text and structured plans are autoregressive, image latents use flow matching. The sequence is causal.
Each selected reference yields a 512-d ArcFace vector, projected into one ID token, so references are an addressed set rather than an unordered pool. A single token is easy to ignore in a 10–20K-token interleaved context, so ID Representation Forcing asks the model to predict, before synthesis, a representation aligned with each target ArcFace. Those hidden states stay available to later image tokens.
Spatial relations go through Layout CoT. In a fixed order the model emits person indices, face boxes, body boxes, pose keypoints, and identity–layout bindings on a 2,002-token coordinate vocabulary. A deterministic renderer turns the plan into a condition image. Correspondence is read from annotated boxes, not guessed from generated pixels. Layout-Grounded ID Loss crops predicted and target images at the same regions and runs ArcFace. The pairing is correct by construction, so five to ten people can be supervised independently. Training uses about 400K group images; the 210-example benchmark is identity-disjoint.
The main table averages ArcFace, FaceNet, and AdaFace.
| Method | Sim(Tgt)↑ | Sim(Ref)↑ | Copy-Paste↓ | Coverage↑ | Dup↓ |
| WithEveryone | 0.499 | 0.540 | 0.055 | 0.973 | 0.028 |
| GPT-Image 2 | 0.462 | 0.583 | 0.169 | 0.905 | 0.075 |
| Seedream 5.0 Pro | 0.436 | 0.522 | 0.114 | 0.913 | 0.065 |
| Nano Banana 2 | 0.451 | 0.480 | 0.045 | 0.884 | 0.099 |
| WithAnyone | 0.405 | 0.483 | 0.096 | 0.957 | 0.045 |
Target-context similarity is highest. Reference similarity trails GPT-Image 2, and copy-paste is much lower, so the score is less consistent with pasted reference faces. Coverage is 97.3% with a 2.8% duplicate rate. Open general-purpose models mostly stay below 0.42 coverage. By group size, ArcFace falls from 0.629 at five people to 0.571 at ten; GPT-Image 2 falls from 0.593 to 0.496.
LG-ID Loss alone lifts Sim(Ref) from 0.339 to 0.506 and Sim(Tgt) from 0.304 to 0.435, the largest single-component gain. The full model scores 0.461 Sim(Tgt) at 1K and 0.499 at the 2K main protocol. Plan IoU is 0.773; swapping in ground-truth layout still helps, so leftover error sits more on planning than on executing a plan.
Group photos, event posters, and multi-character shots cannot be assembled from single-person draws. Identity supervision moves from matching faces in the generated image to using layout annotations as addresses. If the product need is five-plus identity preservation, annotated-box supervision is the piece to add, not another pile of reference tokens. This is a group-scene engineering result on 210 images, not a new general text-to-image SOTA.
210 examples, and only ten at the ten-person end, so per-size curves are trends. Identity metrics inherit detector and recognizer demographic bias. Systems differ in maximum resolution and in whether they can plan a layout; ID-Patch and WithAnyone were given ground-truth layouts. When the prompt under-specifies pose, many layouts are valid and a single reference layout can punish a reasonable plan. Higher identity fidelity also raises the risk of placing a real person in a scene they never joined.