Interleaving Reasoning for Better Text-to-Image Generation

1 INTRODUCTION

• Unified multimodal models still lag systems such as GPT-4o in instruction-following for image generation and preservation of visual details.The stated challenge is that images may be broadly semantically correct while lacking visual quality and fine-grained fidelity, including textures, shadows, and delicate structures.Source: §1 · pp. 3–4
• Interleaving Reasoning Generation (IRG) first uses text-based thinking to guide an image, then reflects on that image to refine quality and details while preserving its established subject matter.The authors distinguish this from prior reflection approaches that focus on major semantic or structural corrections, describing IRG as an end-to-end approach focused on finer details and overall visual quality.Source: §1 · p. 4
• IRGL-300K organizes six learning modes around initial thinking and generation, textual reflection, and faithful implementation of refinements in a subsequent image.Training first strengthens thinking and reflection, then uses full thinking–image trajectories to tune the IRG pipeline.Source: §1 · p. 4
• IRG achieves state-of-the-art performance, with absolute gains of 5–10 points across GenEval, WISE, TIIF, GenAI-Bench, and OneIG-EN.The authors also report improved visual quality and fine-grained fidelity.Source: §1 · p. 4

2.2.1 OVERVIEW

• IRGL-300K is organized into six decomposed learning modes spanning text-based thinking and the complete high-quality image-generation pipeline.The pipeline overview also describes a dedicated CFG condition design for improved-image generation.Source: Figure 3 · p. 5

2.1 PRELIMINARIES AND NOTATIONS

• IRG encodes the initial image into ViT and VAE features, then uses interleaved text–image representations for progressive generation and cross-modal information fusion.The method limits its study to a single refinement iteration, producing a two-turn process.Source: §2.1 · p. 5, §2.2.1 · p. 5

2.2.2 INTERLEAVING REASONING GENERATION LEARNING

• The six IRGL modes separately train initial and improving thinking, as well as image generation conditioned on the associated reasoning.They form two objectives: learning text-based thinking and learning the high-quality image-generation pipeline under reasoning supervision.Source: §2.2.2 · pp. 6–7
• IRGL uses two training stages: first train all six tasks to strengthen reasoning while retaining full trajectories, then tune the complete pipeline on full thinking–image trajectories.The authors note that full IRG data is difficult to obtain at scale, motivating intermediate objectives that train text-based thinking without exposing it to low-quality image data.Source: §2.2.2 · p. 7

2.2.3 INTERLEAVING REASONING DATA CONSTRUCTION

• Training data pairs prompts and images with generated initial reasoning, and uses base-model images paired with improved images and stage-level reflection text for refinement learning.For the improving full-learning data, GPT-4o produces the improved image from the prompt and base-model initial image; an MLLM generates the stage-level improving reasoning.Source: §2.2.3 · pp. 7–8

2.2.4 INFERENCE STRATEGY

• At inference, CFG separately compares conditions with and without initial-image information and with and without reflection text, using guidance scales of 2.0.The authors state that this customized strategy helps maintain visual quality and fidelity and benefits generation stability.Source: §2.2.4 · p. 8

3.1 EXPERIMENT SETTINGS

• The experiments use six learning modes, including four tasks with 50K OSP1024-286k samples each, and train for 2K then 30K steps.Initial Full Learning uses BLIP3o-60k; Improving Full Learning uses approximately 30K samples distilled from GPT-4o.Source: §3.1 · pp. 9–10

3.2 MAIN RESULTS

• IRG scores 0.85 on GenEval and 0.77 on WISE, leading unified models on compositional generation and knowledge-grounded generation, respectively.On GenEval, FLUX.1-dev scores 0.82; on WISE, FLUX.1-dev scores 0.50.Source: §3.2 · p. 11
• IRG achieves TIIF overall scores of 76.00 (short) and 73.77 (long), and the best GenAI-Bench Overall score of 0.84.TIIF is evaluated with QwenVL2.5-72B; the benchmarks assess complex instruction following and compositional text-to-visual generation.Source: Table 3 · p. 10, Table 4 · p. 10, §3.2 · p. 12
• IRG scores 0.415 overall on OneIG-EN, ranking second to GPT-4o at 0.533 and achieving the best alignment and style scores.The overall score averages five dimensions.Source: Table 5 · p. 11, §3.2 · p. 12

3.3 ANALYSIS OF EXPERIMENTAL RESULTS

• Ablations show the largest reported gains with IRG plus six decomposed learning modes, reaching WISE 0.77, TIIF 76.00/73.77, and GenAI-Bench 0.84.Adding high-quality image training to self-CoT raises WISE from 0.70 to 0.73 and TIIF from 68.06/68.78 to 70.69/69.85; adding IRG improves WISE and GenAI-Bench by 0.03 each.Source: Table 6 · p. 11, §3.3 · p. 12
• The full IRG pipeline receives higher average ranking agreement from multiple automatic raters than the first-step-only variant, 63.3% versus 36.7%.The evaluators include Qwen2.5-VL-72B, GPT-4o, and UnifiedReward.Source: §3.3 · p. 14
• Compared with BAGEL and BAGEL w/ self-CoT, IRG’s reflection step improves visual quality and fine-grained fidelity, including textures and poorly rendered details.The visualization comparison uses examples from WISE and GenAI-Bench at 1024×1024 resolution.Source: Figure 4 · p. 13, §3.3 · p. 14
• Remaining failures include over-smoothing repetitive textures, text-rendering drift under dense constraints, and local edits that perturb crowded-scene layouts.These issues are most pronounced when the improving step introduces many simultaneous edits; conservative editing may improve stability but cap gains.Source: §3.3 · p. 14

4.1 UNIFIED MULTIMODAL UNDERSTANDING AND GENERATION MODELS

• Unified multimodal models are grouped as autoregressive, additional-diffusion, and unified integrated-transformer approaches.IRG uses BAGEL, a unified integrated transformer pretrained on large-scale interleaved text–image data.Source: §4.1 · p. 14
• IRG is based on BAGEL, but its framework is described as applicable in principle to all three model categories because they handle interleaved inputs and outputs.Source: §4.1 · p. 14

4.2 REASONING MODELS

• Prior T2I studies explore text-based reasoning, while IRG integrates interleaving reasoning and distinguishes itself from reflection-based methods by targeting visual quality, fine-grained fidelity, and aesthetic appeal alongside semantic correctness.Source: §4.2 · pp. 14–15

• IRG generates images through a text–image–text–image process, using text reasoning to guide an initial image and reflection to produce a refined image.Source: §5 · p. 15
• Extensive experiments on mainstream benchmarks demonstrate significant improvements in generation performance.Source: §5 · p. 15
• IRG places strong emphasis on enhancing visual quality and fine-grained details.Source: §5 · p. 15