Source-linked AI summary
UniReason 1.0: A Unified Reasoning Framework for World Knowledge Aligned Image Generation and Editing
Dianyi Wang, Chaofan Ma, Feng Han, Size Wu, Wei Song, Yibin Wang, Zhixiong Zhang, Tianhang Wang, Siyuan Wang, Zhongyu Wei, Jiaqi Wang
TL;DR
Unified multimodal models struggle with reasoning-intensive image synthesis and commonly separate T2I generation from image editing. UniReason unifies them through world knowledge-enhanced textual reasoning and editing-like visual refinement, achieving advanced performance across generation and editing benchmarks while maintaining broad synthesis capability.
Problem
Existing unified models struggle with complex synthesis requiring implicit world knowledge and typically treat T2I generation and image editing as separate tasks.
Method
UniReason combines world knowledge-enhanced textual reasoning before synthesis with self-reflective, fine-grained editing-like visual refinement after generation.
Results
UniReason achieves advanced performance across T2I and editing benchmarks, with the best overall performance among open-source unified multimodal models and comparable or superior results to several closed-source models.
Takeaways & Limitations
Unifying generation and editing enables world-knowledge-grounded reasoning, iterative visual correction, and strong instruction following and visual fidelity across both task types.
Takeaways & Limitations
Responsible deployment must address misuse risks from generated visual content through system-level safeguards beyond this paper’s scope.
Abstract
from arXiv · showhide
Unified multimodal models often struggle with complex synthesis tasks that demand deep reasoning, and typically treat text-to-image generation and image editing as isolated capabilities rather than interconnected reasoning steps. To address this, we propose UniReason, a unified framework that harmonizes these two tasks through two complementary reasoning paradigms. We incorporate world knowledge-enhanced textual reasoning into generation to infer implicit knowledge, and leverage editing capabilities for fine-grained editing-like visual refinement to further correct visual errors via self-reflection. This approach unifies generation and editing within a shared architecture, mirroring the human cognitive process of planning followed by refinement. We support this framework by systematically constructing a large-scale reasoning-centric dataset (~300k samples) covering five major knowledge domains (e.g., cultural commonsense, physics, etc.) for textual reasoning, alongside an agent-generated corpus for visual refinement. Extensive experiments demonstrate that UniReason achieves advanced performance on reasoning-intensive benchmarks such as WISE, KrisBench and UniREditBench, while maintaining superior general synthesis capabilities.
1 Introduction
UniReason addresses the limits of surface-level reasoning and separate treatment of T2I generation and image editing by unifying world-knowledge reasoning with visual refinement. It supports this framework with dedicated training data, two-stage training, and strong benchmark performance.
- Unified multimodal models struggle with synthesis requiring precise instruction following, world knowledge, and faithful visual output.
- Existing approaches often reason only before generation, limiting their ability to use visual feedback to detect and correct output errors.
- The framework jointly treats T2I generation and image editing because refinement and editing share a reasoning pattern that enables bidirectional capability transfer.
- UniReason combines world knowledge-enhanced textual reasoning before synthesis with fine-grained editing-like visual refinement afterward.
- Training data covers five knowledge categories for textual reasoning and uses an agent pipeline for refinement supervision within a two-stage training strategy.
- Extensive experiments report advanced performance on GenEval, WISE, UniREditBench, and KrisBench.
2 Related Work
Related work develops unified multimodal models and textual reasoning for image synthesis, but existing methods leave implicit world knowledge and generation-editing synergies insufficiently addressed. UniReason targets both gaps within a unified reasoning framework.
- Image generation and editing are related tasks conditioned respectively on textual descriptions and reference images, with Diffusion Transformers and flow matching supporting modern generation frameworks.
- Unified multimodal models combine image understanding and generation through modular alignments or shared language-model architectures.
- Prior image-generation reasoning commonly adapts textual chain-of-thought in a reason-then-generate paradigm.
- UniReason infers implicit world knowledge rather than merely parsing instructions and exploits synergies between T2I generation and image editing.
3 Preliminary
UniReason builds on Bagel’s unified multimodal architecture, which uses separate understanding and generation experts for textual outputs and latent-space image synthesis. Its interleaved reasoning formulation iteratively updates reasoning text and images, treating refinement as editing.
- Bagel uses a Mixture-of-Transformers architecture with a ViT encoder to support unified image understanding and generation.
- The understanding expert generates context-aware text through standard next-token prediction conditioned on multimodal context.
- The generation expert produces semantically aligned images through rectified flow in a VAE latent space, using latent flow-matching loss.
- T2I generation outputs intermediate reasoning tokens and an image, while editing takes an existing image and instruction to produce reasoning text and an edited image.
- Interleaved reasoning iteratively updates images and reasoning text from multimodal context, with each refinement step interpreted as an editing operation conditioned on the reasoning trace.
4 Method
UniReason unifies text-to-image generation and image editing through world knowledge-enhanced textual reasoning followed by editing-like visual refinement. It supports these paradigms with specialized data construction and two-stage training.
- Framework overview: UniReason uses textual reasoning for initial synthesis and editing-like visual refinement for iterative improvement within one multimodal framework.The framework is organized into world knowledge-enhanced textual reasoning and fine-grained editing-like visual refinement phases.
- World knowledge-enhanced textual reasoning: World knowledge-enhanced reasoning expands underspecified prompts by inferring commonsense, cultural, spatial, temporal, and scientific knowledge.The resulting guidance supplies structured details for image synthesis beyond surface-level instruction reorganization.
- Data preparation: The textual-reasoning data covers five knowledge categories and combines manually seeded, expanded, chain-of-thought-annotated, and rendered image samples.Editing data includes original images, editing instructions, and desired outcomes from UniREdit-Data-100K.
- Fine-grained visual refinement: Visual refinement reassesses an initial image against prior reasoning, identifies inconsistencies or missing details, and applies targeted corrections.An agent pipeline uses generation, verification, refinement, and comparison to construct supervision for both generation and editing.
- Training strategy: Two-stage supervised fine-tuning first strengthens foundational image synthesis, then jointly trains understanding and generation branches on interleaved reasoning data.The second stage supervises textual reasoning and image outputs, with text and image losses balanced by scalar weights.
5 Experiments
Experiments evaluate UniReason on knowledge-intensive and general generation and editing benchmarks, along with ablations of training, reasoning, and refinement. Results show strong open-source performance, broad knowledge coverage, retained general ability, and consistent component-wise gains.
- Experimental setup: The evaluation uses WISE for knowledge-intensive T2I generation and KrisBench and UniREditBench for knowledge-intensive image editing.These benchmarks assess world knowledge reasoning, semantic alignment, and refinement capabilities.
- Main results: UniReason achieves the best overall performance among open-source unified multimodal models across knowledge-intensive image generation and editing tasks.It is comparable to Seedream 4.0 and GPT-4o on T2I generation, surpasses Gemini 2.0 on KrisBench, and outperforms Seedream 4.0 on UniREditBench.
- Main results: UniReason shows broad world-knowledge coverage, with highest performance in cultural commonsense, spatial reasoning, physics, and chemistry categories.The model also performs strongly across diverse knowledge categories on KrisBench and UniREditBench.
- General ability retention: UniReason remains highly competitive on general image generation and editing benchmarks while improving knowledge-enhanced reasoning.It surpasses leading systems on GenEval, performs best among reasoning models on DPGBench, and is strongest among reasoning-capable models on ImgEdit and GEdit-EN.
- Ablation study: Consistent improvement occurs across all benchmarks as two-stage training, textual reasoning, and visual refinement are progressively added.World knowledge-enhanced reasoning yields a +0.21 improvement on WISE, while refinement further improves overall performance.
- Correlation of editing and refinement: Refinement gains increase monotonically with ImgEdit scores across three knowledge-intensive benchmarks.The trend supports jointly training image editing and T2I generation because insufficient editing capacity can limit reasoning-guided refinement.
6 Conclusion
UniReason unifies text-to-image generation and image editing through world knowledge-enhanced textual reasoning and editing-like visual refinement. Extensive benchmark experiments report advanced performance across both tasks.
- UniReason harmonizes text-to-image generation and image editing by exploiting their shared structural synergies.
- World knowledge-enhanced textual reasoning infers implicit commonsense and physical laws, while editing-like visual refinement enables iterative reflection and correction.
- High-quality datasets spanning five knowledge categories and a two-stage training strategy support the unified reasoning framework.
- UniReason demonstrates superior instruction following and visual fidelity, with advanced performance across multiple T2I and editing benchmarks.
7 Impact Statement
The work targets improved reasoning and alignment in image generation and editing models while acknowledging potential misuse risks. Responsible deployment requires safeguards beyond the paper’s scope.
- The work focuses on improving reasoning and alignment in image generation and editing models.
- Generated visual content may introduce misuse risks, requiring system-level safeguards and responsible deployment practices beyond this paper’s scope.
A.1 Compared Baselines
The compared baselines include closed-source models and open-source unified multimodal systems spanning autoregressive, discrete diffusion, textual-reasoning, coordinate-based, and interleaved-reasoning approaches.
- Closed-source baselines include GPT-4o, Gemini-2.0, and Seedream4.0.
- Open-source unified multimodal baselines include Harmon, TokenFlow-XL, Janus-Pro, Lumina-DiMOO, MMaDA, and Show-o.
- Naive-reasoning baselines primarily decompose abstract instructions into explicit semantic components, while GoT adds coordinate-based spatial guidance.
- IRG and UniCoT represent another baseline family using interleaved reasoning mechanisms to reorganize semantics during generation.
A.2 Data Preparation Details
UniReason’s data preparation combines world knowledge-enhanced reasoning supervision with agent-generated visual refinement data. Evaluations cover general T2I generation, image editing, and qualitative reasoning-intensive cases.
- Data Preparation Details: The two-phase data pipeline integrates world knowledge-enhanced textual reasoning with fine-grained editing-like visual refinement.
- Phase I: World Knowledge–Enhanced Reasoning Data Construction: Phase I constructs reasoning-grounded samples across five world-knowledge categories and supplements editing triplets with generated reasoning and quality filtering.
- Phase II: Fine-grained Editing-like Visual Refinement Data Construction: Phase II uses an agent pipeline in which generation, verification, refinement, and comparison produce iterative supervision for interleaved reasoning.
- Evaluation: General evaluations cover GenEval and DPG for T2I generation and ImgEdit and GEdit-EN for image editing.
- Qualitative Results: Qualitative results show strong handling of maze navigation, temporal evolution, viewpoint transformations, and fine-grained corrections to faces, text, and hand gestures.