acceptodds
Under review as a conference paper at ICLR 2027

A Unified Reasoning Framework for World Knowledge-Aligned Image Generation and Editing

Abstract

Unified multimodal models often struggle with complex synthesis tasks that demand deep reasoning with world knowledge, 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 a dual reasoning paradigm. We formulate generation as world knowledge-enhanced planning to inject implicit constraints, and leverage editing capabilities for fine-grained visual refinement to further correct visual errors via self-reflection. This approach unifies generation and editing within a shared representation, 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 planning, alongside an agent-generated corpus for visual self-correction. Extensive experiments demonstrate that UniReason achieves advanced performance on reasoning-intensive benchmarks such as WISE and KrisBench, while maintaining superior general synthesis capabilities on GenEval and ImgEdit.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Unified Multimodal Models,Unified Multimodal Reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Ym1aOpenReview ↗
Rating
4(+1)
Confidence
3(−1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes UniReason, a unified multimodal reasoning framework for image generation and editing. The method first employs world knowledge-enhanced textual reasoning to infer implicit semantic constraints and guide initial image synthesis. It then uses visual self-reflection to identify generation errors and performs fine-grained corrections through image editing.

The authors construct approximately 300K reasoning and refinement training samples and adopt a two-stage training strategy. Experiments on WISE, KrisBench, UniREditBench, and general image generation and editing benchmarks demonstrate competitive performance across multiple tasks.

Critical strengths

Strong empirical performance. UniReason achieves leading or competitive results across multiple knowledge-intensive image generation and editing benchmarks, with particularly strong performance on UniREditBench and GenEval. The model also maintains strong capabilities on general image generation and editing tasks. The extensive evaluations and ablation studies further support the effectiveness of the proposed framework.

Critical weaknesses

Insufficient evidence for the synergy between generation and editing. One of the paper's central motivations is that image generation and editing share substantial reasoning capabilities and can mutually benefit from joint training. However, the proposed approach introduces large-scale additional training data and relies on multiple powerful teacher models to construct reasoning and refinement samples. Although the paper includes several ablation studies, these primarily demonstrate the effectiveness of textual reasoning and visual refinement components. A systematic comparison between separate and joint training under matched data scales and training budgets is still missing. Therefore, the current results do not sufficiently establish the claimed synergy between the two tasks.

Questions for authors

The authors argue in the Introduction that image generation and editing have inherent synergies and can mutually reinforce each other within a unified interleaved reasoning framework. Could the authors compare T2I-only, Editing-only, and Joint Training under matched data scales and training budgets, or evaluate a pipeline combining separately trained generation and editing models? Such comparisons would help establish whether joint training provides additional benefits beyond simply combining the two capabilities.

If the authors make a convincing effort to address this concern, I would consider raising my rating.

Reviewer_WvFBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces UniReason, a unified multimodal framework designed to enhance both text-to-image (T2I) generation and image editing by integrating deep reasoning with world knowledge. UniReason employs World Knowledge-Enhanced Textual Reasoning to infer implicit constraints from abstract user instructions before generating an initial image. And utilizes Fine-grained Editing-like Visual Refinement, where the model performs self-reflection on the generated draft and applies targeted visual corrections, treating post-generation critique as an image editing task to create a synergistic loop between generation and editing. Built upon the BAGEL architecture and trained via a two-stage fine-tuning strategy, UniReason achieves state-of-the-art performance among open-source models on knowledge-intensive benchmarks like WISE, KrisBench, and UniREditBench, while maintaining superior general synthesis capabilities on standard evaluations such as GenEval and ImgEdit. The paper is clearly written and easy to follow; however, I have significant concerns regarding its novelty and the actual effectiveness of the proposed method.

Critical strengths
  1. The paper proposes UniReason, which elegantly unifies text-to-image (T2I) generation and image editing within a single architecture. By treating post-generation refinement as structurally analogous to image editing, it successfully exploits the inherent synergies between these two tasks rather than handling them in isolation.
  2. The authors systematically construct a large-scale, high-quality dataset (~300K samples) covering five diverse world knowledge domains (cultural commonsense, natural science, spatial, temporal, and logical reasoning). The use of an agent-based pipeline (generator, verifier, refinement teacher, and judge) ensures robust supervision for interleaved reasoning and visual self-correction.
  3. The paper is well-structured, clearly written, and easy to follow, making the proposed dual reasoning paradigm (planning followed by refinement) highly accessible.
Critical weaknesses
  1. The core methodology combining Chain-of-Thought (CoT) textual reasoning with iterative visual self-correction—is largely an intuitive integration of existing techniques (e.g., IRG, UniCoT). The paper lacks a fundamental architectural or algorithmic breakthrough, making the technical contribution feel more like a sophisticated engineering combination than a novel scientific discovery.
  2. The paper predominantly highlights successful examples and aggregate benchmark scores but lacks a deep, transparent analysis of failure cases. Understanding where the world knowledge-enhanced reasoning or the visual refinement loop breaks down (e.g., hallucinations during reflection, inability to correct complex structural errors) is critical for assessing the true boundaries of the method.
  3. The paper lacks a clear explanation of the relationship between the model's performance gains, the constructed reasoning dataset, and the two-stage training strategy. It remains ambiguous how much of the overall improvement is driven by the proposed UniReason architectural framework itself versus the high-quality data and specific training recipe. A more detailed analysis or ablation study is needed to clarify the respective roles and individual contributions of the reasoning data and the training strategy in the entire performance enhancement process.
Questions for authors
  1. Could you clarify the relationship between the model's performance improvements, the constructed reasoning dataset (~300K samples), and the two-stage training strategy? Specifically, how much of the overall gain is attributable to the proposed UniReason architectural framework itself versus the high-quality data and the specific training recipe? It would be helpful to provide a detailed ablation study that isolates the respective roles and individual contributions of the reasoning data and the training strategy in the entire performance enhancement process.
  2. The core idea of interleaving textual reasoning with visual refinement appears structurally similar to existing methods like IRG and UniCoT. Beyond unifying T2I generation and image editing, what is the fundamental algorithmic or architectural novelty of UniReason? Could you explicitly articulate why this combination constitutes a non-trivial scientific contribution rather than an intuitive engineering integration of prior techniques?
  3. The evaluation focuses on five specific world knowledge categories (cultural commonsense, natural science, spatial, temporal, and logical reasoning) that were explicitly covered during the data construction phase. Is there a risk that the model has overfitted to these predefined domains? Do you have any results or analysis demonstrating the model's zero-shot generalization ability when facing out-of-distribution reasoning tasks or unseen knowledge scenarios?
1 comment

Dear Reviewer WvFB ,

I hope this message finds you well. As the discussion period is nearing its end with less than one days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your responses are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Reviewer_UqdSOpenReview ↗
Rating
4(+1)
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes UniReason, a unified model for text-to-image generation and image editing that combines world-knowledge-enhanced textual planning with reflection-guided visual refinement. It constructs reasoning and refinement supervision across multiple knowledge domains and trains the framework with two-stage supervised fine-tuning. The approach is evaluated on knowledge-intensive and general generation and editing benchmarks.

Critical strengths

The paper makes a useful connection between knowledge-grounded generation and image editing as two stages of a shared reasoning process. The formulation and the ablation provide direct evidence that textual reasoning and refinement are complementary: the full model improves over both reasoning-only and refinement-only variants on WISE, KrisBench, and UniREditBench. This is the paper’s most consequential contribution, because it moves refinement from a post-hoc wrapper into the training target of a unified model.

The empirical results are broad and consistently favorable across the intended capabilities. Tables show strong results on knowledge-intensive generation and editing, while additional reasoning training does not substantially sacrifice general image generation or editing ability. In particular, the combination of a 0.78 WISE score, 68.23 KrisBench score, 70.06 UniREditBench score, and 0.90 GenEval score supports the claim that the method is not narrowly optimized for one benchmark.

Critical weaknesses

The central empirical comparisons are not sufficiently specified to establish how robust the reported gains are.

Section 5.1 provides optimizer schedules and aggregate dataset sizes, but does not state the number of training or generation seeds, report uncertainty or run-to-run variation, describe the image sampling and refinement settings used at evaluation, or clearly distinguish which baseline numbers were reproduced under a common protocol versus transcribed from prior reports. This matters because several headline margins are modest, for example on WISE relative to strong baselines in Table 1, and the claimed advantage combines training changes with a different test-time reasoning-and-refinement procedure.

The paper should provide an explicit evaluation protocol for every benchmark, including decoding/sampling parameters, number of refinement calls, seed aggregation, evaluator settings where relevant, and confidence intervals or multi-seed results for UniReason. At minimum, reproducing the closest baselines under the same inference budget would make the comparisons much more convincing.

Questions for authors

were UniReason results averaged over multiple generation and training seeds? Please specify the seed count, sampling settings, number of generated candidates per prompt, and whether all baseline results were re-evaluated under the same benchmark and inference protocol.

How do the authors prevent contamination or near-duplicate overlap between the constructed world-knowledge/refinement data and WISE, KrisBench, and UniREditBench? This is particularly important because the data construction uses Wikipedia-derived seed prompts, UniREdit-Data-100K, and LLM-generated expansions

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.