Source-linked AI summary
Unified Thinker: A General Reasoning Modular Core for Image Generation
Sashuai Zhou, Qiang Zhou, Jijin Hu, Hanqing Yang, Yue Cao, Junpeng Ma, Yinchao Ma, Jun Song, Tiezheng Ge, Cheng Yu, Bo Zheng, Zhou Zhao
TL;DR
Open-source image models remain challenged by logic-intensive instructions, and text-space plans may not translate into executable images. Unified Thinker separates a trainable planner from the image generator and grounds its plans with reinforcement learning using final-image feedback. It improves reasoning and instruction-following outcomes across generation and editing settings and multiple generator backbones.
Problem
Open-source image models still struggle with logic-intensive or implicit directives, and external text-space plans can mismatch generator capabilities.
Source: §1 · p. 1Method
Unified Thinker separates a trainable MLLM Thinker that creates structured plans from the image Generator, then uses supervised training and generator-in-the-loop reinforcement learning with final-image constraint rewards.
Source: §1 · p. 2Results
Unified Thinker substantially improves instruction following and constraint satisfaction across four generation and editing settings, with gains across multiple generator backbones.
Source: §1 · p. 2Takeaways & Limitations
The reported gains are especially pronounced on reasoning-intensive requests, supporting the value of separating reasoning from rendering.
Source: §6 · p. 8The approach depends on the quality and coverage of its intermediate representation, training data, and automatic RL rewards, which can bias results and limit generalization beyond evaluated benchmarks.
Source: §7 · p. 9
Abstract
from arXiv · showhide
Despite impressive progress in high-fidelity image synthesis, generative models still struggle with logic-intensive instruction following, exposing a persistent reasoning--execution gap. Meanwhile, closed-source systems (e.g., Nano Banana) have demonstrated strong reasoning-driven image generation, highlighting a substantial gap to current open-source models. We argue that closing this gap requires not merely better visual generators, but executable reasoning: decomposing high-level intents into grounded, verifiable plans that directly steer the generative process. To this end, we propose Unified Thinker, a task-agnostic reasoning architecture for general image generation, designed as a unified planning core that can plug into diverse generators and workflows. Unified Thinker decouples a dedicated Thinker from the image Generator, enabling modular upgrades of reasoning without retraining the entire generative model. We further introduce a two-stage training paradigm: we first build a structured planning interface for the Thinker, then apply reinforcement learning to ground its policy in pixel-level feedback, encouraging plans that optimize visual correctness over textual plausibility. Extensive experiments on text-to-image generation and image editing show that Unified Thinker substantially improves image reasoning and generation quality.
1 Introduction
Open-source image models still struggle with logic-intensive instructions, while existing reasoning approaches trade off modularity, visual fidelity, and reliable execution. Unified Thinker separates planning from image synthesis and trains plans against visual outcomes, improving performance across tasks and generator backbones.
1 Introduction
- Open-source models struggle with logic-intensive or implicit directives, while built-in reasoning can compromise modularity and visual fidelity and external plans may not match generator capabilities.Iterative planning in external-planner methods can also increase compute.Source: §1 · p. 1
- The Thinker converts instructions into hierarchical plans with an intent summary, explicit constraints, and ordered sub-goals that the Generator uses as conditioning.The modular design is intended to transfer across text-to-image and editing tasks and work with different generator backbones.Source: §1 · p. 2
- Supervised fine-tuning on HieraReason-40K establishes plan quality, then generator-in-the-loop reinforcement learning rewards final-image constraint satisfaction.The resulting pixel-level feedback encourages plans that are executable by the Generator, not merely semantically plausible.Source: §1 · p. 2
- Unified Thinker improves instruction following and constraint satisfaction across four image-generation and editing settings, with gains across multiple generator backbones.The evaluations cover text-to-image reasoning, reasoning-based editing, general text-to-image generation, and general image editing.Source: §1 · pp. 2–3
- The paper presents a decoupled reasoning-generation framework and an end-to-end training pipeline spanning hierarchical reasoning data construction and execution-led reinforcement learning.Its stated contributions also include verifying cross-model portability of the reasoning core.Source: §1 · p. 3
2 Related Work
Prior image-generation work includes diffusion and unified multimodal models, instruction-guided editing, and several strategies for adding reasoning. Unified Thinker is presented as a standalone decoupled thinker intended to transfer across generation and editing tasks.
2.1 Foundational Generative Models
- Diffusion models, unified multimodal systems such as Bagel and OmniGen, and instruction-guided editing methods form the foundational landscape for image generation.The cited examples include latent diffusion, joint text-image token modeling, and editing through instruction-guided manipulation.Source: §2.1 · p. 3
2.2 Reasoning for Image Generation
- Reasoning approaches structure prompts into intermediate steps or layouts, reason about intent and constraints before drawing, or refine images through reflection and correction.Examples include latent-intent inference by R-Genie and reflective revision by Reflect-DiT and EditThinker.Source: §2.2 · p. 3
- Unlike these approaches, Unified Thinker is a standalone decoupled module intended to provide reusable reasoning across image-generation and image-editing tasks.The paper frames it as transferable across diverse tasks rather than tied to one generation workflow.Source: §2.2 · p. 3
3 Data Construction
HieraReason-40K is a general-purpose corpus pairing image-generation or editing instructions with structured reasoning traces that end in enhanced prompts. Its traces make implicit requirements explicit and convert resolved visual targets into generator-ready specifications.
3 Data Construction
- HieraReason-40K combines four sources spanning text-to-image generation, general editing, reasoning image generation, and reasoning image editing.Examples pair an instruction, optionally with reference images, with a structured reasoning trace ending in an enhanced prompt.Source: §3 · p. 3
- The construction pipeline combines seed knowledge and user requests to produce structured reasoning traces and executable enhanced prompts.The figure describes these components as the data-construction flow.Source: §3 · p. 3, Figure 3 · p. 4
- Each trace identifies task type and intent, resolves implicit requirements into a concrete visual target, then converts that target into an enhanced prompt.For editing, the prompt describes only intended changes and assumes unspecified content carries over from the reference image.Source: §3 · p. 4
- Gemini3-Pro generates initial traces, which are normalized and filtered or rewritten for format, edit-only compliance, visual specificity, and consistency.Task-general system prompts are also designed to cover diverse generation and editing scenarios.Source: §3 · p. 4
4 Framework and Training
Unified Thinker separates structured planning from image synthesis: a trainable Thinker produces executable visual specifications for a diffusion-based Generator. Training first jointly supervises reasoning and generation, then uses dual-phase reinforcement learning to improve planning and execution.
4 Framework and Training
- The framework decouples a trainable Thinker, which produces structured reasoning and an executable visual specification, from a diffusion-based Generator conditioned on its outputs.The same think-then-execute design is intended for image generation and editing.Source: §4 · p. 4
4.1 Joint Supervised Fine-Tuning
- Joint supervised fine-tuning trains synchronized understanding and generation views, using language-modeling loss for the Thinker and diffusion denoising loss for the Generator.The two learning signals are combined with a weighted sum; λ balances them.Source: §4.1 · pp. 4–5
4.2 Dual-Phase Reinforcement Learning
- Dual-phase reinforcement learning first rewards Thinker plans by relative image-quality performance with a fixed Generator, then improves Generator execution by crediting denoising trajectories that yield better images.The Thinker samples multiple reasoning paths for the same request; generation-oriented RL uses distinct rollouts.Source: §4.2 · pp. 5–6
5 Experiments
Unified Thinker improves reasoning-intensive image editing and generation, with reported gains also extending to general generation and editing. Ablations examine training stages, Thinker backbones, and transfer to another generator.
5.1 Experimental Setup
- Evaluations span four generation and editing settings; Qwen-Image-Edit executes plans, while Qwen3-VL-30B supplies visual-correctness and logical-consistency rewards.Training uses HieraReason-40K for supervised fine-tuning and 4K sampled instances for reinforcement learning.Source: §5 · p. 6, §5.1 · p. 6
5.2 Main Results
- Unified Thinker improves reasoning-heavy editing on RISEBench and leads open-source models on WiseBench, narrowing the gap to GPT-4o.RISEBench gains are especially large for temporal and spatial reasoning; WiseBench gains cover most domains.Source: §5.2 · p. 6, Table 1 · p. 7, Table 4 · p. 8
- On PRISM, overall quality improves mainly through aesthetics while preserving alignment; on GEditBench, gains are modest but consistent across reported metrics.The supplied captions identify PRISM alignment, aesthetics, and average scores, and GEditBench metrics on its English split.Source: §5.2 · p. 6, Table 3 · p. 7, Table 2 · p. 7
5.3 Ablation Study
- Adding the Thinker improves reasoning benchmarks but slightly hurts GEdit low-level editing; joint fine-tuning alleviates the mismatch, and two-stage Dual-RL yields consistent gains across benchmarks.The ablation progressively adds the Thinker, joint fine-tuning, and Dual-RL to the Qwen-Image-Edit baseline.Source: §5.3 · pp. 6–7, Table 5 · p. 8
- A stronger Thinker generally improves reasoning-oriented performance, though the 7B variant can be preferred for PRISM aesthetics; Thinker also gives stable gains when transferred to BAGEL.With either tested Thinker backbone, reasoning and consistency gains over Qwen-Image-Edit are reported while visual quality remains comparable.Source: §5.3 · p. 8, Table 6 · p. 8
6 Conclusion
Unified Thinker separates reasoning and planning from image synthesis through a structured intermediate representation. Its two-stage training and experiments support gains over open-source baselines, especially for reasoning-intensive requests.
6 Conclusion
- The decoupled Thinker–Generator framework provides diffusion models with transferable reasoning and planning.The Thinker converts text-to-image and editing requests into a structured executable representation, leaving the Generator to synthesize the image.Source: §6 · p. 8
- A roughly 40K cold-start corpus and a two-stage reinforcement-learning pipeline enhance planning and execution.The stated goal is to improve both planning and execution through the training pipeline.Source: §6 · p. 8
- Experiments report consistent gains over strong open-source baselines, especially on reasoning-intensive requests.Source: §6 · p. 8
7 Limitations
The approach has constraints in its training and reward dependencies, transfer across diffusion backends, difficult edits, and inference cost.
7 Limitations
- Generalization beyond evaluated benchmarks may be limited by the quality and coverage of the intermediate representation, training data, and automatic RL rewards.These dependencies can also introduce bias.Source: §7 · p. 9
- Generator-agnostic design does not make executability fully invariant across different diffusion backends.Source: §7 · p. 9
- Fine-grained geometric changes, strict locality, and precise text rendering remain challenging, while planning adds inference latency and compute cost.Source: §7 · p. 9
A.1 Joint Supervised Fine-Tuning
Joint supervised fine-tuning uses two Qwen-VL instruction models with LoRA and a mixed HieraReason-40K dataset. Training settings include specified optimization and image-resizing configurations.
A.1 Joint Supervised Fine-Tuning
- Joint supervised fine-tuning covers Qwen2.5-VL-7B-Instruct and Qwen3-VL-8B-Instruct, using LoRA with r=8 applied to all modules.Source: §A.1 · p. 12
- The mixed HieraReason-40K instruction data is trained for five epochs with batch size 4 per device, eight gradient-accumulation steps, and learning rate 4 × 10−5.The setup uses a cosine schedule with 10% warmup and λ = 0.5.Source: §A.1 · p. 12
- Images are resized to a 512-pixel short side while preserving aspect ratio.Source: §A.1 · p. 12
A.2 Dual-Phase Reinforcement Learning
The Thinker is optimized with GRPO using image-generation rollouts and reward-based updates. The appendix reports the rollout, sampling, and optimization settings.
A.2 Dual-Phase Reinforcement Learning
- Training uses GRPO on 64 GPUs, with 16 prompts per rollout batch and 24 candidates generated per prompt.Sequence expansion is enabled, with top-k sampling at 100 and temperature 0.99.Source: §A.2 · p. 12
- Each iteration performs one update epoch with value, reward, and advantage clipping thresholds of 0.5, 10, and 10, respectively.The procedure uses KL regularization of 0.01 against a reference model and no advantage whitening.Source: §A.2 · p. 12
- The Thinker is initialized from Qwen2.5-VL-7B-Instruct/Qwen3-VL-8B-Instruct, while Qwen3-VL-30B-A3B-Instruct judges rewards and Qwen-Image-Edit performs editing.The Thinker is trained in BF16 at a learning rate of 1 × 10^-6; editing uses 10 sampling steps.Source: §A.2 · p. 12
A.3 Training Evaluation
Training evaluation tracks mean reward and per-step rollout generation time. The VLM reward design uses different scoring rubrics for image editing and text-to-image generation.
A.3 Training Evaluation
- Mean reward increases steadily during training, indicating consistent improvement of the learned policy.Figure 5 plots mean reward; Figure 6 reports per-step rollout generation time across training.Source: §A.3 · p. 12, Figure 5 · p. 12, Figure 6 · p. 12
A.4 Reward Model and Design
- For image editing, the VLM judge scores appearance consistency, reasoning/alignment, and visual plausibility on a 1–5 scale, then aggregates them into one reward.The judge receives the pre- and post-edit images, edit instruction, and a reference description of the intended outcome.Source: §A.4 · p. 12
- For text-to-image training, the judge scores consistency, realism, and aesthetic quality from 0 to 2, and their mean is the final reward.The synthesized image is evaluated from the prompt and, when applicable, the answer field extracted from the model output; the final reward is in [0, 2].Source: §A.4 · p. 12
A.5 Evaluated Comparative Models
The appendix describes a comparative-model roster, a prompt format that separates reasoning from visual instructions, and the construction and structure of HieraReason-40K. The dataset supplies structured traces and enhanced prompts for image-generation and editing tasks.
A.5 Evaluated Comparative Models
- The comparative evaluation roster includes multimodal systems and image-generation or editing models, spanning closed-source and open-source approaches.The listed examples include Gemini-2.5-Flash-Image (Nano Banana), GPT-Image-1, BAGEL, Qwen-Image/Edit, FLUX.1, and other models.Source: §A.5 · pp. 12–13
C Details of HieraReason-40K
- HieraReason-40K combines four source datasets, sampling 10K examples from each and distilling them into a unified structured format with Gemini.Examples pair instructions, optionally with reference images, with reasoning traces and enhanced prompts for the generator.Source: Table 7 · p. 13, C Details of HieraReason-40K · pp. 13–14
B System Prompt
- The system prompt distinguishes full-scene descriptions for T2I from change-only instructions for I2I and confines reasoning to <think> while <answer> contains the concrete visual result.It also forbids restating unchanged content to reduce edit drift and covers generation, local edits, combine/transform, and solve/draw scenarios.Source: B System Prompt · p. 13
C Details of HieraReason-40K
- The trace procedure identifies task type and intent, makes implicit requirements explicit through reasoning, and ends with a concrete visual result and enhanced prompt.For I2I, the answer describes only the requested change; for T2I, it describes the scene from scratch.Source: C Details of HieraReason-40K · p. 14