RealGen: Photorealistic Text-to-Image Generation via Detector-Guided Rewards

1. Introduction

• Current powerful models can still produce artifacts such as overly smooth skin and oily facial sheens, while human-preference scores may conflate realism with aesthetic appeal.The introduction notes that candid, realistic images may score poorly when they lack aesthetic appeal.Source: §1 · p. 2
• RealGen uses semantic- and feature-level synthetic-image detectors as rewards, guiding generation to evade detection and reduce artifacts.The approach treats images with fewer AI artifacts as more realistic and uses both visible-artifact analysis and deep-feature analysis.Source: §1 · p. 2
• The framework uses an LLM to enrich user prompts and a diffusion model to generate images, optimizing both stages with detector-scored GRPO.The LLM is trained with the generator fixed; the diffusion model is then optimized while keeping the text input consistent.Source: §1 · p. 2
• RealBench evaluates photorealism with automated Detector-Scoring and Arena-Scoring protocols.Arena-Scoring uses pairwise comparisons against other model outputs or real images, with at least 3,000 random pairings per model.Source: §1 · pp. 2–3
• RealGen significantly outperforms general and specialized text-to-image models in realism, detail, and aesthetics.The comparison includes models such as GPT-Image-1, Qwen-Image, and FLUX-Krea.Source: §1 · p. 3

2.1. Image Generation Models

• Powerful text-to-image models can generate precise objects and complex text, but realistic image generation remains challenging, including oily-looking faces.RealGen focuses on improving image realism and detail.Source: §2.1 · p. 3

2.2. Reinforcement Learning and Human Preferences in Image Generation

• Preference-based reinforcement learning can align outputs with human aesthetics, but preference rewards may bias colors and fail to reward realistic images that lack aesthetic appeal.The cited examples include HPSv2 favoring red-toned images and PickScore preferring purple ones; RealGen instead uses detection models without human-preference data.Source: §2.2 · p. 3

2.3. Synthesis detection for Image generation

• Synthetic-image detectors have advanced, but prior detector-feedback methods mainly use post-processing, whereas RealGen applies reinforcement learning to optimize the generative model itself.The related work describes detectors based on pretrained vision models and MLLMs, including methods that can explain detected artifacts.Source: §2.3 · p. 3

3.1. Model architecture

• RealGen's LLM expands the user's initial instruction into a longer description, which the diffusion model uses to denoise and decode the final image.The system uses Qwen-3 4B and FLUX.1-dev with fine-tuned LoRA layers; both components receive SFT before GRPO optimization.Source: §3.1 · pp. 3–4

3.2. Detection Reward

• The detector reward combines semantic and feature-level artifact signals with text-image alignment, normalizes reward dimensions within a batch, and fuses them into one advantage.Forensic-Chat assesses visible content such as smooth, greasy skin; OmniAID targets frequency artifacts and abnormal noise; Long-CLIP supplies the alignment reward.Source: §3.2 · p. 4

3.3. Post-training: Detection-GRPO

• With the diffusion model frozen, GRPO trains the LLM to rewrite prompts; it learns to add scenic details, imperfections, and auxiliary phrases such as “shot on iPhone.”The frozen diffusion model generates images from sampled rewrites, and their detector reward guides the LLM update.Source: §3.3 · p. 5
• In the second stage, GRPO optimizes the diffusion model while the LLM is frozen, using full denoising inference for evaluation and random consecutive-step exploration.This design addresses noisy or blurry outcomes from limited-step or full-trajectory RL exploration, which can mislead detectors; some short prompts bypass the LLM to support varying prompt lengths.Source: §3.3 · p. 5

4.1. Dataset overview

• RealBench contains 1,000 real-world images and corresponding captions across seven categories, with a larger share devoted to the challenging, common portrait category.Unlike benchmarks focused on instruction following or human preference, RealBench focuses exclusively on photorealism.Source: §4.1 · p. 5

4.2. Detector-Based Realism Quantification

• Detector-Scoring treats an image's probability of being judged “real” as its photorealism score, using semantic- and feature-level detectors.Evaluators include Forensic-Chat, OmniAID, GPT-5, and the reward-independent feature detector Effort.Source: §4.2 · pp. 5–6

4.3. Arena-Style Preference Evaluation

• Arena-Scoring uses GPT-5 to choose the more realistic of two semantically similar images for the same prompt, then calculates each model's win rate.Each model undergoes at least 3,000 random pairwise battles against other models' outputs and real-world images.Source: §4.3 · p. 6

5.1. Experimental Setup

• Experiments use Qwen3-4B-Instruct and FLUX.1.dev, with comparisons on RealBench and the HPDv2 Photo subset against general and realism-specialized models.Training uses the real-image subset of HPD v3; RL runs for approximately 230 steps, and baselines use official default settings.Source: §5.1 · pp. 6–7

5.2. Main Results

• RealGen outperforms leading T2I models across multiple key photorealism metrics, including specialized methods Flux-Krea and SRPO.The reported advantage holds with or without LLM prompt rewriting, and strong performance on held-out GPT-5 and Effort evaluators is offered as evidence of generalizable realism.Source: §5.2 · p. 7
• RealGen approaches a 50% win rate against real images and has the highest overall win rate in model-versus-model arena comparisons.Eight of 13 competing models score below 30% against real images.Source: §5.2 · p. 7, Figure 6 · p. 7
• In qualitative comparisons, RealGen images show superior texture and detail, while competing models exhibit oily textures, smooth skin, or an unnatural warm cast.The comparison includes FLUX-dev, Bagel, Qwen-Image, and GPT-Image-1.Source: §5.2 · p. 7, Figure 5 · p. 7
• On the HPDv2 Photo subset, RealGen achieves high detector-based realism scores and ranks among the top in human aesthetic preference.The authors report this as evidence of enhanced photorealism alongside maintained aesthetic quality.Source: §5.2 · p. 8, Table 2 · p. 8

5.3. Ablation experiments

• Progressive ablations report that LLM prompt optimization improves realism and detail, with diffusion-model optimization further improving image quality.The LLM enriches prompts with realistic descriptions; diffusion optimization further enhances the realism of people and glass.Source: §5.3 · p. 8, Figure 7 · p. 8, Figure 8 · p. 8
• Detector-Score outperforms human-preference rewards on multiple held-out metrics, while PickScore and HPSv2.1 tend toward cartoonish or artistic styles and oily portrait textures.The qualitative comparison reports more realistic results from Detector-Score.Source: §5.3 · p. 8, Table 3 · p. 8, Figure 9 · p. 8

• RealGen uses semantic- and feature-level detectors as rewards with GRPO to optimize the generation pipeline for image realism and detail.The conclusion frames this as a “Detection for Generation” paradigm.Source: §6 · p. 9
• RealBench provides human-free photorealism evaluation through Detector-Scoring and Arena-Scoring.Source: §6 · p. 9
• The stated aim is to address the photorealism gap in current text-to-image models and encourage further progress in photorealistic image synthesis.Source: §6 · p. 9