Resolving the Identity Crisis in Text-to-Image Generation

1. Problem Statement: The Identity Crisis

Current State-of-the-Art (SOTA) text-to-image models (e.g., diffusion and flow-matching models) suffer from a persistent failure mode when generating scenes with multiple humans, termed the "Identity Crisis."

• Symptoms: When prompted to generate a group of people, models often produce:
  • Duplicate Faces: Multiple individuals in a single image share identical or near-identical facial features.
  • Identity Merging: Distinct identities are collapsed into a single archetype.
  • Miscounting: The model fails to generate the requested number of people.
  • Cross-Sample Repetition: Even when generating multiple samples for the same prompt, the model recycles the same set of identities across different images rather than creating diverse groups.
• Impact: This limits the utility of synthetic data for applications requiring realistic group dynamics, such as storytelling, social simulation, and personalized media. Existing methods focus on aesthetic quality and prompt adherence but lack mechanisms to enforce identity diversity both within an image (intra-image) and across a batch of generated samples (cross-sample).

2. Methodology: DISCO (Reinforcement with Diversity Constraints)

The authors propose DISCO, a reinforcement learning (RL) framework designed to fine-tune flow-matching text-to-image models. It utilizes Group-Relative Policy Optimization (GRPO) to optimize non-differentiable objectives without requiring real-world annotated data.

A. Core Framework

• Base Model: Flow-matching models (e.g., FLUX-Dev, Krea-Dev).
• Training Strategy: The model generates a group of $M$ trajectories (images) for a single prompt. The GRPO algorithm computes advantages based on the relative performance of these samples within the group, rather than against a fixed baseline.
• Curriculum Learning: To handle the increasing complexity of multi-human scenes, DISCO employs a single-stage curriculum. It starts training on simple prompts (2–4 people) and gradually anneals to a uniform distribution covering the full range (up to $N_{max}$ people). This stabilizes training, particularly for specialist models.

B. Compositional Reward Function

The core innovation is a multi-component reward signal $R$ that guides the RL optimization. For a generated image $x_i$ within a group $G$:
$$R = \alpha r^d_{img} + \beta r^d_{grp} + \gamma r^c_{img} + \zeta r^q_{img}$$

1. Intra-Image Diversity ($r^d_{img}$):
   • Goal: Penalize duplicate faces within a single image.
   • Mechanism: Uses RetinaFace for detection and ArcFace for embedding. It calculates the maximum cosine similarity between any pair of faces in the image. The reward is $1 - \max(s_{ij})$. If fewer than 2 faces are detected, a neutral reward is given.
2. Group-wise Diversity ($r^d_{grp}$): The Novel Contribution
   • Goal: Prevent the model from simply shifting duplicates from one image to another across the batch (global identity collapse).
   • Mechanism: A counterfactual "leave-one-out" statistic. For each image $i$ in the group, the system calculates the average pairwise similarity of the entire group ($S_G$) and the group without image $i$ ($S_{G-i}$).
   • Reward: If removing image $i$ increases the overall group diversity (i.e., $S_{G-i} > S_G$), image $i$ is rewarded. This explicitly penalizes samples that contribute to global repetition.
3. Count Accuracy ($r^c_{img}$):
   • Goal: Enforce the exact number of people requested in the prompt.
   • Mechanism: Binary reward (1 if detected count equals target, 0 otherwise). This prevents "reward hacking" where the model generates fewer people to avoid the difficulty of creating diverse faces.
4. Quality & Alignment ($r^q_{img}$):
   • Goal: Maintain visual fidelity and prompt adherence.
   • Mechanism: Uses HPSv3 (Human Preference Score) normalized to $[0, 1]$. This ensures the model does not sacrifice image quality or prompt following (e.g., clothing, hairstyle) to maximize diversity.

3. Key Contributions

1. Identification of the Identity Crisis: Formalized the problem of intra-image and cross-sample identity collapse in multi-human generation.
2. Group-Relative Counterfactual Reward: Discovered that optimizing intra-image diversity alone leads to "global diversity collapse" (duplicates migrate across samples). The proposed group-wise reward explicitly penalizes this behavior.
3. Reward Hacking Mitigation: Identified that diversity optimization alone causes the model to generate fewer people or unnatural "grid" arrangements. The combination of count control and HPS quality rewards effectively regularizes these failure modes.
4. Annotation-Free Scalability: The method requires no real-world human data or manual annotations, relying solely on synthetic prompts and automated metrics (face detectors/embeddings).
5. Generalizability: Demonstrated that optimizing for facial diversity implicitly improves diversity in other attributes (clothing, body type) and preserves general compositional capabilities (object counting, spatial reasoning).

4. Experimental Results

The authors evaluated DISCO on the DiverseHumans Testset (1,200 prompts, 2–7 people) and MultiHuman-TestBench (MHTB).

• Quantitative Performance:
  • Unique Face Accuracy (UFA): DISCO (Flux) achieved 98.6%, significantly outperforming the baseline Flux (48.2%) and proprietary models like GPT-Image-1 (85.1%).
  • Global Identity Spread (GIS): DISCO achieved 98.3%, compared to 50.5% for the baseline and 89.8% for GPT-Image-1.
  • Count Accuracy: Improved from 70.8% (baseline) to 92.4%.
  • Quality: HPS scores remained competitive (33.4 vs. 31.7 baseline), proving quality was not sacrificed.
• Scalability: Performance remained high even as the number of people increased to 6–7, whereas baseline models degraded sharply.
• Ablation Studies: Confirmed that removing the group-wise reward leads to global collapse, and removing count control leads to under-generation. The curriculum learning was critical for specialist models (Krea) but less so for generalists (Flux).

5. Significance and Conclusion

DISCO represents a paradigm shift in text-to-image generation by treating identity diversity as a first-class optimization objective rather than a side effect.

• Technical Impact: It establishes that RL with group-relative rewards can solve complex, non-differentiable constraints (like "no duplicates in a batch") that supervised fine-tuning cannot.
• Practical Application: It enables the generation of large-scale, diverse synthetic datasets for training other AI models (e.g., for social interaction research or personalized media) without the cost of manual data collection.
• Future Directions: The authors note that this framework can be extended to video generation (spatiotemporal identity consistency) and personalized generation, with follow-up work already exploring these avenues.

In summary, DISCO resolves the "Identity Crisis" by teaching models that uniqueness is a global constraint, not just a local one, resulting in the most diverse and accurate multi-human image generation to date.