---
base_model: black-forest-labs/FLUX.1-dev
library_name: peft
license: other
tags:
  - lora
  - flux
  - text-to-image
  - diffusers
  - peft
  - reinforcement-learning
  - face-diversity
  - multi-human
language:
  - en
---

# Resolving the Identity Crisis in Text-to-Image Generation

**DisCO** (Diversity Constraints via GRPO) is a LoRA adapter for [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) that dramatically improves facial diversity and identity distinctness in multi-human image generation.

[![Paper](https://img.shields.io/badge/arXiv-2510.01399-b31b1b.svg)](https://arxiv.org/abs/2510.01399)
[![Project Page](https://img.shields.io/badge/Project-Page-blue)](https://qualcomm-ai-research.github.io/disco/)
[![GitHub](https://img.shields.io/badge/GitHub-Code-black?logo=github)](https://github.com/Qualcomm-AI-research/disco)
[![CVPR 2026](https://img.shields.io/badge/CVPR-2026-purple)](https://arxiv.org/abs/2510.01399)

![DisCO Training Method](https://raw.githubusercontent.com/Qualcomm-AI-research/disco/main/assets/disco_main.jpg)

---

## Abstract

> Text-to-image models tend to generate duplicate faces, merge identities, or miscount people in multi-human scenes. DisCO addresses this by fine-tuning flow-matching models via reinforcement learning to optimize facial diversity both within individual images and across sample batches. The method employs a composite reward function that addresses facial similarity penalties, deters identity repetition, ensures accurate person counting, and preserves image quality. Testing on our evaluation benchmark demonstrates superior performance, achieving approximately 98.6% unique-face accuracy while also outperforming both open-source and proprietary competitors. Notably, the approach requires no external training data, making it a scalable solution for generating images containing multiple distinct individuals.

![DisCO scoresheet](https://raw.githubusercontent.com/Qualcomm-AI-research/disco/main/assets/disco_scoresheet.jpg)

---

## Model Details

| Field | Value |
|---|---|
| Base model | `black-forest-labs/FLUX.1-dev` |
| Adapter type | LoRA (PEFT) |
| LoRA rank (r) | 64 |
| LoRA alpha | 128 |
| Dropout | 0.0 |
| Weight init | Gaussian |
| Target modules | All attention projections + feed-forward layers (12 modules) |
| PEFT version | 0.17.0 |

**Target modules:** `attn.to_q`, `attn.to_k`, `attn.to_v`, `attn.to_out.0`, `attn.to_add_out`, `attn.add_q_proj`, `attn.add_k_proj`, `attn.add_v_proj`, `ff.net.0.proj`, `ff.net.2`, `ff_context.net.0.proj`, `ff_context.net.2`

---

## Notes

> **Disclaimer:** The model weights released here are not identical to the model reported in the paper. They are provided for research and demonstration purposes only. Performance may vary from the results reported in the paper.

---

## Usage

### Option 1: DisCO inference repo (recommended)

Clone the repo — weights are pulled automatically from HuggingFace on first run:

```bash
git clone https://github.com/Qualcomm-AI-research/disco.git
cd disco
```

Run inference (no manual weight download needed):

```bash
# Side-by-side base vs DisCO
python inference.py --prompt "A stunning close-up of Six people on a campus walkway, clear faces visible, fine detail, lifelike rendering, diversity in ethnicity." --compare

# DisCO only
python inference.py --prompt "A stunning close-up of Six people on a campus walkway, clear faces visible, fine detail, lifelike rendering, diversity in ethnicity."
```

If you prefer to use locally downloaded weights (e.g. from [GitHub Releases](https://github.com/Qualcomm-AI-research/disco/releases)), place them at `loras/disco/adapter_model.safetensors` — the code will use the local copy automatically.

Or launch the Gradio demo:

```bash
python app.py
# open http://localhost:7864
```

### Option 2: Directly with PEFT + Diffusers

```python
import torch
from diffusers import FluxPipeline
from peft import PeftModel

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.bfloat16,
).to("cuda")

pipe.transformer = PeftModel.from_pretrained(
    pipe.transformer,
    "Qualcomm-AI-Research/disco",
)
pipe.transformer = pipe.transformer.merge_and_unload()

image = pipe(
    "A stunning close-up of Six people on a campus walkway, clear faces visible, "
    "fine detail, lifelike rendering, diversity in ethnicity.",
    height=1024,
    width=1024,
    num_inference_steps=28,
    guidance_scale=3.5,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]

image.save("disco_output.png")
```

### Recommended prompts

```
Two people on a shallow beach, diverse faces, clear faces visible, realistic lighting
Four people in a city plaza, midday, diverse faces, clear faces visible, high fidelity
Six people on a campus walkway, diverse faces, clear faces visible, lifelike rendering
```

---

## Hardware Requirements

- **GPU:** NVIDIA A100 40 GB (or equivalent, ≥ 40 GB VRAM)
- **CUDA:** 12.4
- **Python:** 3.11
- **PyTorch:** 2.6.0

---

## Training Details

DisCO is trained with Flow-GRPO, a flow-matching adaptation of Group Relative Policy Optimization (GRPO):

- **Base model:** FLUX.1-dev (flow-matching transformer)
- **Training algorithm:** Flow-GRPO (RL via composite reward signal)
- **Reward components:**
  - Face similarity penalty (ArcFace embeddings, intra-image)
  - Batch-level identity diversity (cross-sample)
  - Person count accuracy (headcount matching prompt)
  - HPS v3 image quality score
- **No external training data** — rewards are computed fully at inference time
- **Training regime:** bf16 mixed precision, 7 GPUs

---

## Citation

```bibtex
@InProceedings{Borse_2026_CVPR,
  author    = {Borse, Shubhankar and Farhadzadeh, Farzad and Hayat, Munawar and Porikli, Fatih},
  title     = {Resolving the Identity Crisis in Text-to-Image Generation},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2026},
  pages     = {36703--36712},
}
```

---

## License

These weights are derived from [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) and are subject to the [FLUX.1 Non-Commercial License](https://github.com/Qualcomm-AI-research/disco/blob/main/LICENSE-FLUX1-dev.txt). Any use of these weights must comply with that license.

Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.