How could identity grounding for group image generation change group image synthesis over the next two years?

What exactly is identity grounding, and why does it matter?

Identity grounding means the model doesn't just know 'there are three people'—it knows which reference face goes to which person and where that person stands in the image. Without it, text-to-image models often produce duplicate faces, merge two identities into one, or miscount the number of people. A 2025 study called this an 'identity crisis' and showed that even state-of-the-art models struggle with multi-human scenes [2]. Identity grounding solves this by explicitly linking each reference identity to a specific region or token in the image, so the model can't mix them up.

The 2026 WithEveryone framework is the clearest example: it injects each selected identity as an 'addressed token' and predicts a structured identity-layout plan before rendering the image [1]. This explicit planning step is what allows it to handle up to ten reference identities at once—a scale that was previously unreliable. The key innovation is a training objective that uses annotated face regions to supervise the intended identities directly, avoiding the unstable embedding-based face matching that often fails with multiple faces [1].

How much better will group images actually get?

The numbers from the strongest study are encouraging but not magical. WithEveryone improved face similarity from 0.462 to 0.499—a modest 8% gain—but that's on a benchmark that already had decent similarity [1]. The bigger win was in reducing copy-paste artifacts: from 0.169 to 0.055, a 67% drop [1]. That means the model is far less likely to just clone the same face multiple times, which is a common failure in group generation. It also covered 97.3% of requested identities with only a 2.8% duplicate rate, so you're much more likely to get every person you asked for, and each one distinct [1].

Another 2025 study, DisCo, took a different route—reinforcement learning with diversity constraints—and achieved 98.6% unique-face accuracy on a dedicated test set [2]. That's near-perfect in terms of avoiding duplicate faces, though it doesn't guarantee identity fidelity (that is, matching the reference faces). The two approaches are complementary: WithEveryone focuses on binding identities to locations, while DisCo focuses on making faces distinct. Together, they suggest that the next two years will bring models that can both keep identities straight and avoid the 'clone army' effect.

What's the catch? Are there limits?

The main catch is that these results come from controlled benchmarks, not real-world chaos. WithEveryone's benchmark was 'identity-disjoint,' meaning the test identities were not seen during training—a good test of generalization, but still a curated setting [1]. DisCo's test set is also synthetic, and it doesn't require real data, which is a plus for scalability but may not capture all real-world variations [2]. So while the improvements are real, they may not translate perfectly to every user's photo album.

Another limit is scale: WithEveryone handles up to ten identities, but real group photos can have dozens of people. The paper doesn't show what happens beyond ten, and the authors note that training-time identity losses become unstable with many faces [1]. UniPortrait, a 2024 framework, also supports multiple IDs but focuses on editability and layout diversity rather than extreme scale [3]. So over the next two years, expect reliable group generation for small to medium groups (up to ten), but don't expect flawless results for a wedding party of fifty just yet.

About These Sources

This answer is built on 3 studies (all preprints) — published from 2024 to 2026, 3 from 2024 or later — selected as the most relevant from 3 studies that passed quality screening, drawn from 33 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions