Unified Text-Image Generation with Weakness-Targeted Post-Training

TL;DR

The paper addresses the challenge of fully unifying text and image generation in a single inference, eliminating explicit modality-switching. By introducing reward-weighted regression (RWR) and a weakness-targeted MMGW dataset, it trains a unified model to autonomously transition from textual reasoning to visual synthesis, achieving improvements across four T2I benchmarks. The key contributions include identifying QwenVQAScore as an effective offline reward, demonstrating the superiority of weakness-targeted data over broad or benchmark-aligned prompts, and showing that multimodal training yields strong gains in knowledge-intensive tasks while exposing persistent text-rendering bottlenecks. This work provides practical guidance for designing data-efficient, reward-driven multimodal fine-tuning and highlights the trade-offs between joint reasoning and pure image generation in real-world applications.

Abstract

Unified multimodal generation architectures that jointly produce text and images have recently emerged as a promising direction for text-to-image (T2I) synthesis. However, many existing systems rely on explicit modality switching, generating reasoning text before switching manually to image generation. This separate, sequential inference process limits cross-modal coupling and prohibits automatic multimodal generation. This work explores post-training to achieve fully unified text-image generation, where models autonomously transition from textual reasoning to visual synthesis within a single inference process. We examine the impact of joint text-image generation on T2I performance and the relative importance of each modality during post-training. We additionally explore different post-training data strategies, showing that a targeted dataset addressing specific limitations achieves superior results compared to broad image-caption corpora or benchmark-aligned data. Using offline, reward-weighted post-training with fully self-generated synthetic data, our approach enables improvements in multimodal image generation across four diverse T2I benchmarks, demonstrating the effectiveness of reward-weighting both modalities and strategically designed post-training data.

Figures (6)

• Figure 1: Sample generations before and after our post-training. Our post-training enables successful generations for challenging, previously failed prompts as well as fully automatic joint text-image generation.
• Figure 2: The Multi-Modal Generative Weaknesses (MMGW) Dataset includes five semantic categories that reliably induce generation failures in unified multimodal models. Here we show one representative prompt and failure image for each category.
• Figure 3: Reward distributions across all samples. QwenVQAScore exhibits a distinct bimodal distribution, effectively discriminating between low-quality and high-quality samples. In contrast, all other reward functions display unimodal distributions with limited variance, resulting in minimal discriminative power across samples.
• Figure 4: Reward distributions for individual prompts show that QwenVQAScore effectively scores generations and produces intra-prompt rewards that distinguish good from bad samples; more prompt-level reward distributions are shown in Appendix \ref{['app:reward_dists']}. Other reward functions do not correlate with generation quality.
• Figure 5: Sample generations for a OneIG-Bench Text category prompt. BAGEL Image-Only produces the clearest text, consistent with the stronger text accuracy of image-only generation methods seen in Table \ref{['tab:oneigbench_modality']}. Our Multimodal RWR post-training helps remedy some of the text rendering failures of BAGEL Multimodal.