acceptodds
Withdrawn from ICLR 2027

Qwen-Image-Omni: Complex Visual Programs Don't Need External Reasoner

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unified multimodal models (UMMs) support multi-turn understanding and generation of interleaved images and text. Complex visual programming requires finer control at each turn: choosing an operation, identifying its target, and recalling historical images with distinct roles while preserving earlier decisions. We present Qwen-Image-Omni, which connects a pretrained Vision-Language Model (VLM) and Diffusion Transformer (DiT) through structured skill calls and a shared visual history. The VLM uses its internal reasoning to select the next operation, assign reference roles, and expand the goal into an executable visual instruction. The same VLM supplies the DiT's conditioning, and each generated image becomes available for subsequent reasoning and execution. We construct histories that preserve dependencies between single-turn skills, then adapt visual execution and prompt expansion in two stages. Autoregressive supervision targets visual prompt content while system instructions specify call syntax. Under matched training and call budgets, the combined skill-and-reference interface improves complex visual design and historical editing over general instructions with natural-language references. Public benchmarks further demonstrate strong interleaved generation, reasoning-informed editing, and general generation and editing capability.

Primary area:
generative models
Keywords:
Multimodal Generation, Diffusion Transformers, Visual Program Execution

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_ELdPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Qwen-Image-Omni with a pretrained VLM and Diffusion Transformer (DiT) through structured skill calls and a shared history for complex visual programming without external reasoner.

A multi-turn adaptation method is introduced to construct dependency-preserving histories from single-turn skills and trains visual execution and prompt expansion in 2 stages.

Qwen-Image-Omni achieves strong performance on public benchmarks.

Critical strengths

1, The idea of supporting complex visual programs by combining shared visual history, structured natural-language calls, and two-stage adaptation is interesting and makes sense.

2, Qwen-Image-Omni demonstrates strong performance on 4 benchmarks such as UEval, RISEBench, GenEval 2, and GEdit-Bench, outperforming open unified multimodal baselines.

3, Ablation studies are conducted to support the multi-turn capability of Qwen-Image-Omni on ComplexEval by task family.

Critical weaknesses

1, Qwen-Image-Omni seems only for visual-language multi-turn understanding and generation. For Qwen-Image-Omni, I expect it will also support multimodality beyond image and text.

2, Lack the experimental results for inference efficiency comparison. For example, in Table 1, Qwen-Image-Omni is showing performance improvement over baselines such as BAGEL/Emu3.5. It is also good to show latency comparison.

3, Qwen-Image-Omni is doing better than Gemini-2.5-Flash-Image on reasoning-informed editing in Table 2, and image generation and editing in Table 3. The paper did not share why Qwen-Image-Omni outperformed Gemini-2.5-Flash-Image and provided qualitative comparison. Take Fig.4 examples for qualitative comparison.

4, Many details are missing in this paper. It is hard to understand why the performance is improving, such as training data, training cost, and training recipe for reproducing the performance.

5, Some symbols are not fully explained, for example, what is jj and mijm_{ij} in Eq. 1. Why AR and FM have unit coefficients. Will it change if architecture is changing, or batch size is changing.

6, Not much explanation for Table 4, for example, the Coarse interface is doing better than the Full model on the Interleaved task.

7, FM convergence is very fast in Fig. 5. Why is this happening and why do we need to train 10k steps?

Questions for authors

See Weakness. I will change my rating based on the rebuttal.

Reviewer_RMc5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper is generally well organized and easy to follow, addressing a practically relevant problem across multiple complex task settings. However, its contributions lack sufficient novelty and a clear research insight that provides new value to the community. The central claims are not adequately supported by the experimental evidence, and the ablation studies fail to isolate the contributions of individual design components. ADditionally, insufficient methodological and experimental details raise concerns about reproducibility. Overall, the paper falls short of establishing a sufficiently novel, well-supported, and reproducible research contribution.

Critical strengths

1.Clarity of presentation. The paper is generally well organized and easy to follow.

2.Practical relevance and breadth of scope. The paper addresses a practically relevant problem and attempts to evaluate the proposed approach across multiple complex task settings.

Critical weaknesses
  1. Lack of Novelty and Research Insight. The paper addresses a problem that is already well recognized by the community, but it does not provide a sufficiently clear explanation of the underlying insights or empirical observations that motivate the proposed solution. As a result, the work offers limited new knowledge or value to the research community.

  2. Weak Core Claims and Insufficient Experimental Support. The paper's central claims and contributions are relatively weak, and the experimental evidence does not adequately substantiate them. In particular, there is a gap between the paper's core claims and the supporting experimental results. Furthermore, the ablation studies do not sufficiently disentangle the contributions of individual design components, making it difficult to assess the effectiveness of each component and the necessity of the overall design.

  3. Insufficient Methodological and Experimental Details. The paper lacks sufficiently detailed descriptions of the experimental setup and implementation details. This raises concerns about the reproducibility of the results and makes it difficult for readers to independently verify the proposed approach.

Questions for authors

Please refer to "Weaknesses" above.

Reviewer_43TaOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes Qwen-Image-Omni, a unified multimodal model built upon Qwen-Image-2.0 and Qwen3-VL-8B for complex visual programs. Specifically, the model introduces structured skill calls to specify visual transformations, identify reference images, and control content preservation. It also maintains a shared image-text history to support multi-turn generation and editing.

Critical strengths
  1. The proposed framework supports multi-turn image editing and allows previously generated results to be reused in subsequent editing tasks.

  2. Extensive experiments are conducted across multiple benchmarks.

Critical weaknesses
  1. Several key concepts and design choices are not clearly explained. For example, what exactly are the structured skills in this paper?

  2. The overall performance is still limited compared with strong proprietary models such as GPT and Gemini. Moreover, since the proposed model is built upon Qwen-Image-2.0, the authors should provide a direct comparison with Qwen-Image-2.0 on each applicable task to better demonstrate the effectiveness of the proposed framework.

  3. The paper lacks sufficient qualitative results and failure case analysis. In addition, in the Long-context Reference example in Fig. 1, the final editing result does not fully follow the pose instruction specified in R3.

  4. The training data construction is not sufficiently described, particularly for the prompt-only supervision stage. More details about the data sources, annotation process, etc., should be provided.

Questions for authors

See weaknesses.

Reviewer_pqe3OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Qwen-Image-Omni, a unified image generation and editing system built on Qwen-Image-2.0. It uses the existing VLM to produce structured skill calls specifying operation types, historical image references, and expanded visual instructions. The authors construct multi-turn training histories from single-turn generation and editing data, followed by two-stage adaptation with prompt-only autoregressive supervision and flow matching. Experiments cover public generation and editing benchmarks and a custom ComplexEval benchmark for multi-step visual tasks.

Critical strengths

S1. A practical interface for multi-step visual tasks. The combination of skill tags, explicit reference IDs, and expanded visual instructions provides a useful way to organize image generation and editing over a shared history. Reusing the existing VLM also avoids deploying a separate reasoning model. S2. Meaningful controlled comparisons. Table 4 compares Full and Coarse under matched training and call budgets. Full improves visual-design success from 0.62 to 0.70 and long-history editing from 0.29 to 0.41. Prompt-only AR also substantially outperforms full-sequence AR on visual design (0.70 versus 0.43). S3. Broad evaluation and a dedicated multi-turn protocol. The experiments cover several public generation and editing benchmarks. ComplexEval further evaluates multi-step execution and historical reference use, with paired random seeds and blinded human assessment.

Critical weaknesses

W1. The methodological novelty needs clearer positioning. Several closely related works are missing from the discussion. Think-Then-Generate (ICML 2026) already uses an LLM-based encoder for internal reasoning and prompt expansion. Re-Align (CVPR 2026) explicitly separates semantic guidance from reference association. A more direct comparison would help establish what the proposed visual-call interface adds beyond these methods. W2. Unclear PE configurations and unquantified deployment benefits. A central motivation is to eliminate the separately deployed Prompt Enhancer (PE) used by Qwen-Image-2.0. However, the PE or equivalent prompt-rewriting configurations of the baselines in Tables 1–3 are not consistently specified. More importantly, the paper does not directly compare Qwen-Image-Omni with the official Qwen-Image-2.0 + PE pipeline in terms of both generation quality and practical serving costs, such as end-to-end latency, throughput, and GPU memory requirements. Given the additional two-stage training involved, the distinction between one-time training overhead and recurring inference savings would also be useful. Without this comparison, the practical benefit of removing the external PE remains difficult to assess. W3. Limited evidence for learned visual decision-making. Section 3.3 applies autoregressive supervision only to visual prompt tokens, leaving skill selection, reference IDs, and reasoning tokens without direct AR supervision. Although FM gradients also reach the VLM, the experiments do not clearly distinguish improvements in operation selection from improvements in image execution. An analysis of generated skill calls and reference-selection accuracy would help clarify the role of internal reasoning. W4. The contribution of dependency-preserving history mixing is not directly established. Section 3.2 constructs multi-turn histories from 10M generation and 5M editing examples, supplemented by coherent sequences from other sources. However, the data composition and construction of reference-role instructions are insufficiently detailed. The existing ablations also share the same history-mixing strategy, making its specific contribution difficult to assess. W5. Limited analysis of long-range reference retrieval. Historical reference reuse is a central motivation, yet Table 4 mainly reports aggregate success rates for single- and multi-reference editing. It remains unclear how performance changes with reference distance, history length, and the number of competing images. A breakdown of the existing ComplexEval results along these dimensions would provide more direct evidence for the proposed history-addressing mechanism.

Questions for authors

Q1. Could the authors clarify how the proposed method differs from Think-Then-Generate and Re-Align, particularly in internal reasoning, reference association, and historical conditioning? This would help identify the paper's specific methodological contribution. Q2. Could the authors clarify which baselines use PE, CoT, or equivalent prompt rewriting? For the official Qwen-Image-2.0 + PE pipeline, could they provide a focused comparison of generation quality and deployment costs (e.g., latency, throughput, and GPU memory) against Qwen-Image-Omni? Q3. How reliably does the VLM select skills and historical references without direct AR supervision on those tokens? Could the authors provide an analysis of generated calls and reference-selection performance, including how it varies with reference distance or history length?

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

Related papers

Discussion (0)

Sign in to comment.