acceptodds
Under review as a conference paper at ICLR 2027

From Prompting to Composing: A Spatial Canvas Interface for Poster Generation

Abstract

Text prompting is an indirect interface for poster generation, requiring users to encode inherently two-dimensional composition intent into a one-dimensional sequence of words. We introduce a **Spatial Canvas Interface** that enables users to directly compose generation intent in space through four complementary binding types: semantic, identity, text, and pixel, together with Text Specifications for individual elements and global appearance. Based on this interface, we develop **Compo**, a poster generation model adapted from a pretrained image editing model to understand Spatial Canvas inputs and Text Specifications. Compo supports both direct inference, where users explicitly construct the canvas, and agentic mode, where a high-level request is automatically translated into a planned Spatial Canvas. To train Compo, we develop a scalable pipeline that automatically constructs supervision data for different binding types and their combinations, enabling efficient adaptation without training a specialized poster generator from scratch. We further introduce a benchmark that evaluates adherence to individual binding types and their joint composition. Experiments show that Compo achieves stronger compositional controllability than both general-purpose image generation models and dedicated poster generation systems while maintaining high visual quality. By decoupling intent specification from visual generation, our work shifts poster generation from *prompting* toward *composing*.

Primary area:
generative models
Keywords:
Image Generation, Image Editing, Diffusion Model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yWdeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a Spatial Canvas Interface that combines semantic, identity, text, and pixel bindings with element-level and global text specifications. The authors construct Compo-200K through automatic poster generation and annotation, and train the proposed model Compo. They evaluate the model on a 500-sample benchmark. Experiments show that Compo achieves stronger compositional controllability than both general-purpose image generation models and dedicated poster generation systems while maintaining high visual quality.

Critical strengths

(1) Poster generation involves various distinct types of constraints that should not be conflated. Specifically, preserving the identity of a subject while altering their pose is a fundamentally different requirement from maintaining the shape and position of specific visual elements. Representing these requirements on a unified canvas, while establishing clear links to textual specifications, offers a logically coherent approach with practical utility. Furthermore, this representation can serve as an intermediate construct within an automated planning process.

Critical weaknesses

(1)Compo receives extensive training on the input format (i.e., spatial canvas and text specifications), while the general image editing baselines do not. Several baseline outputs even reproduce the canvas labels. Without a comparison using the same backbone and training data but a different conditioning format, the results show the value of the complete system more clearly than the claimed advantage of composing over prompting.

(2)The benchmark is too close to the training pipeline to establish performance on independently composed inputs. Appendix A.2 uses the same generation process with different seeds, and the references and layouts are derived from rendered target posters. This leaves performance on unrelated user assets and new layouts largely untested.

Questions for authors

(1) What instructions and input conversions were used for the strongest baselines? In particular, how were canvas labels distinguished from text that should appear in the output? This would help determine how much of the reported gap comes from input interpretation.

(2) Does the test split discard identities, styles, or themes, or only generated samples? Are there existing quantitative results on independently assembled canvases with external references? Such evidence would directly affect the assessment of generalization.

4 comments

At first, we sincerely thank you for your valuable time, constructive comments, and positive rating.

W1: Advantages of composing and effectiveness of the Spatial Canvas representation

1. Advantages of composing over prompting. We introduce a controllable poster generation task that jointly supports four constraints: (1) Semantic Binding associates textual concepts with regions; (2) Identity Binding preserves reference identities while following detailed instructions; (3) Text Binding specifies exact text and locations; and (4) Pixel Binding preserves designated visual content. Existing general-purpose text-and-image-to-image models primarily rely on textual instructions, which, even when detailed, may struggle to accurately incorporate diverse visual references and maintain consistency. Meanwhile, dedicated poster generation methods typically employ task-specific conditioning formats that cannot simultaneously support all four binding types. In contrast, our Spatial Canvas Interface unifies heterogeneous control conditions into a single, compositional representation, eliminating the need for separate conditioning mechanisms.

2. Compatibility with general image editing models that are without training. Modern pretrained image editing models commonly use VLMs as visual-semantic encoders and possess strong visual understanding and instruction-following capabilities from large-scale pretraining. Our Spatial Canvas, combined with detailed text specifications and explicit interpretation instructions, naturally aligns with their standard text-and-image-to-image (TI2I) interface.

3. Effectiveness of the Spatial Canvas representation. To further isolate the benefit of our representation from task-specific training, we additionally evaluate ChatGPT Images 2.6 on 50 benchmark cases under two input settings: Detailed Prompt (text-only) and Spatial Canvas + Text Specifications. We also include Compo-LongCat for comparison. The human evaluation results (Elo / Win Rate) are shown below:

ModelsAesthetic QualityInstruction Following
ChatGPT Images 2.6 (w/ Detailed Prompt)1065.6 / 64.5%928.7 / 37.0%
ChatGPT Images 2.6 (w/ Spatial Canvas + Text Specifications)1056.1 / 59.0%1020.4 / 55.5%
Compo-LongCat (Ours)878.4 / 26.5%1051.0 / 57.5%

With the same pretrained model and without task-specific training, introducing the Spatial Canvas substantially improves instruction following (Elo: 928.7 → 1020.4; Win Rate: 37.0% → 55.5%), indicating better adherence to reference and layout constraints. These results provide evidence that our Spatial Canvas representation offers benefits beyond detailed text-only prompting.

4. Comparison using the same backbone and training data but a different conditioning format. To further isolate the benefit of our Spatial Canvas representation, we reimplement CreatiDesign's conditioning format on the LongCat-Image backbone and train it on the same Compo-200K dataset. However, CreatiDesign's conditioning format does not natively support constraints such as our Identity and Semantic Bindings. Therefore, as in our baseline evaluation, we convert the original Compo-200K inputs into CreatiDesign-compatible formats while preserving as much constraint information as possible. Due to time and computational constraints, we cannot reimplement and retrain every heterogeneous conditioning format on LongCat-Image. We therefore select CreatiDesign as a representative case study to examine the impact of different conditioning formats under the same backbone and training dataset.

The results are shown below:

ModelsPixel DINO-IPixel CLIP-IPixel M-DINOPixel SSIMPixel IoUIdentity DINO-IIdentity CLIP-IIdentity M-DINOIdentity CLIP-TIdentity IoUSemantic CLIP-TSemantic IoUText NEDText Sen. Acc.Text IoUOverall IoUAvg. Score
CreatiDesign-LongCat (w/ SFT on Compo-200K)0.6980.7480.6080.7680.8000.4700.5550.3900.1280.4700.1060.4400.7550.6100.7100.6050.550
Compo-LongCat (Ours)0.7350.7760.6420.7930.7810.7950.8910.7570.2610.9040.2590.8280.9700.9400.9440.8850.752
Compo-Qwen (Ours)0.7340.7700.6490.7820.7730.8100.9100.7810.2630.9320.2620.8500.9400.8680.9340.8860.751

Under the same backbone and training dataset, Compo-LongCat substantially outperforms CreatiDesign-LongCat (0.752 vs. 0.550 Avg. Score), particularly on Identity and Semantic Binding metrics. These results further highlight the limitations of task-specific conditioning formats when extending to new types of constraints, and support the flexibility of our unified Spatial Canvas representation.

Q1: Input format conversion when testing baselines

1. General text-and-image-to-image models. For general image editing models, we provide additional instruction interpretation:

task_description = (
    "Generate a poster according to the input image "
    "and the following text prompt. The input image specifies the "
    "layout and reference, while the text prompt defines their detailed generation requirements.\n"

    "Text labels on the input image only serve as "
    "identifiers linking regions to textual descriptions. "

    "Unlabeled assets must preserve original shapes and appearances, allowing only lighting adjustments.\n"

    "Remove all auxiliary bounding boxes "
    "and text labels from the final image."
)

alongside the Spatial Canvas and text specifications to explicitly explain our task and the semantics of different binding types. In particular, these instructions clarify the distinction between canvas labels, which serve as identifiers for associating visual regions with textual specifications, and actual text content that should appear in the generated poster. This additional guidance is intended to help pretrained models correctly interpret our Spatial Canvas without task-specific training.

2. Dedicated poster generation models. For dedicated poster generation methods, we implement deterministic, code-based input conversions that transform our structured annotations into the model-specific conditioning formats, preserving the original constraint information to the extent supported by each model's interface. Importantly, these methods do not directly receive our rendered Spatial Canvas. Instead, they receive the converted annotations in their native input formats. Therefore, the potential confusion between canvas labels and actual poster text does not arise for these dedicated models. Finally, we adapt the inputs of different baselines to their respective interfaces while preserving the intended control constraints as much as possible.

Q2: Separation between the training data and the benchmark, and the effectiveness of existing quantitative results on independently assembled canvas with external references

Our benchmark split is not performed merely at the level of generated poster samples; instead, we deliberately introduce differences at the source data construction stage to reduce overlap between training and evaluation.

1. Difference between training and benchmark data. As described in Section 3.1 (Poster Creation, Page 4) and Appendix A.1 (Text Asset Collection, Page 16), target posters are constructed from predefined text asset repositories, including content and style repositories. Importantly, the repositories used for training and evaluation only have a limited overlap, with additional object categories and styles in the benchmark that are absent from the training repositories. Furthermore, we employ different random seeds for benchmark construction and introduce different object count distributions. Variations in object categories, object counts, and random sampling naturally produce different spatial arrangements and layout combinations. Together, these measures help reduce distributional overlap, ensuring that our benchmark is not simply constructed by regenerating training samples with different random seeds.

2. Generalization to novel references and layouts. According to the difference mentioned above, our benchmark includes newly constructed Spatial Canvas with reference assets and layout combinations that differ from those encountered during training. Therefore, the quantitative results in Table 1 already provide evidence of Compo's generalization to new content, styles, and automatically assembled layouts, rather than merely measuring performance on previously seen poster configurations.

Furthermore, Appendix Figures 8-9 (Page 22-23) present qualitative examples using external visual assets outside our training data, demonstrating Compo's ability to handle unseen references in real-world poster design.

W2: Independence and difference between the benchmark and the training data

We would like to clarify that we intentionally adopt the same construction pipeline for the training and benchmark sets to ensure a consistent input format between training and inference. Specifically, both sets employ the same Spatial Canvas structure, incorporating Pixel Binding, Text Binding, Semantic Binding, and Identity Binding. However, sharing the same construction pipeline does not imply that the two sets contain overlapping data. They differ in three important aspects:

1. Data composition: Each training sample consists of a paired Spatial Canvas and its corresponding poster image, whereas each benchmark sample contains only a Spatial Canvas, from which the model is required to generate a poster image independently.

2. Spatial Canvas content: As described in Appendix Section A.2 (Pages 16-17), although both sets share the same Spatial Canvas structure, the actual content of their components differs. Specifically, the benchmark set uses different semantic categories for Semantic Binding, different text content and numbers of text boxes for Text Binding, and different reference images for Pixel Binding. Since these elements jointly determine the layout and visual content of the generated posters, such differences lead to distinct poster layouts and content, thereby avoiding content overlap between the training and benchmark sets.

3. Poster styles: Beyond differences in layout and content, the benchmark set also includes poster styles that are absent from the training set, further evaluating the model's ability to generalize to unseen visual styles.

Therefore, the shared construction pipeline ensures consistent input representations between training and inference, while the differences in data composition, Spatial Canvas content, and poster styles enable a fair and comprehensive evaluation of the model's generalization capabilities on unseen data.

Dear Reviewer yWde,

Thank you again for your valuable time, constructive feedback, and positive rating. With only one day remaining in the discussion period, we would like to kindly follow up on our responses.

Following your suggestions, we have further clarified the advantages and effectiveness of the composition nature of our proposed Spatial Canvas representation, and conducted additional comparisons using the same backbone and training data but different conditioning formats. We have also clarified the separation between the training data and evaluation benchmark, and discussed the existing quantitative and qualitative evidence of generalization to novel layouts and external references.

We would greatly appreciate any further feedback or comments you may have on our responses.

Best regards,
Authors

Reviewer_6yACOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to address the limitations of text-based prompting for poster generation. The authors propose a Spatial Canvas Interface that integrates four types of bindings: semantic, identity, text, and pixel, enabling fine-grained control over poster composition. Based on this interface, the authors develop Compo, which adapts pretrained image editing models through supervised fine-tuning and reinforcement learning. Moreover, an automatic data generation pipeline is introduced to construct the Compo-200K dataset. The proposed method supports both direct and agentic poster generation. Experimental results demonstrate that Compo achieves better compositional controllability than existing image generation and poster generation methods.

Critical strengths
  • The authors propose a Spatial Canvas Interface that unifies four complementary binding types, enabling explicit control over object semantics, visual identities, text rendering, and spatial layouts. This design provides a more intuitive and flexible way to specify poster compositions than conventional text-based prompting.
  • To avoid expensive manual annotations, the authors develop an automatic data generation pipeline that constructs approximately 200K training samples. Furthermore, a two-stage training strategy combining supervised fine-tuning and reinforcement learning is introduced to improve text accuracy, instance consistency, and layout adherence.
  • Extensive experiments demonstrate that the proposed method outperforms existing general-purpose image generation models and dedicated poster generation methods on most evaluation metrics. The ablation studies further verify the effectiveness of the proposed reinforcement learning rewards.
Critical weaknesses
  • Poster quality is not only about correct layout and text. Visual appeal, clear focus, and overall design are also important. However, the evaluation mainly focuses on constraint accuracy. In some examples, such as Figure 5, CreatiDesign produces visually appealing results despite errors in text rendering. Also, it is unclear whether the synthetic training data was filtered for aesthetic quality.

  • 2.Table 2 shows that the reward components have different impacts. Removing the Layout Reward only reduces the Avg. Score from 0.752 to 0.750. Its additional benefit seems limited and deserves further discussion.

  • 3.The Avg. Score averages 15 metrics, but different binding types have different numbers of metrics. This may give some binding types more influence on the final score. Also, several evaluation metrics are used as RL rewards. A more balanced evaluation across binding types would be helpful.

  • The proposed model mainly adapts existing pretrained image editing models using LoRA and an existing reinforcement learning algorithm. The technical novelty beyond the Spatial Canvas Interface could be further clarified.

  • The evaluation benchmark is generated using a pipeline similar to that of the training dataset. Experiments on independent real-world poster datasets would better demonstrate the generalization capability.

  • Lack of human evaluation. The experiments mainly rely on automatic metrics, which cannot fully reflect the aesthetics, readability, and overall design quality of generated posters.

  • The automatic data generation pipeline relies on segmentation and OCR models, which may introduce inaccurate annotations. The quality of these annotations is not quantitatively assessed.

Questions for authors
  1. How do the authors ensure the quality of the synthetic dataset? Could low-quality outputs from the source image generation models affect Compo's performance?

  2. Why does the Layout Reward provide only a small improvement?

  3. Could the authors show two real-world poster design examples (e.g., animated and live-action movie posters) using text-only requests in Agentic Mode? A small user study on visual appeal, design quality, and theme clarity would also be helpful.

5 comments

At first, we sincerely thank you for your valuable time and constructive comments.

W1: Aesthetic quality of evaluation samples and training data

1. Aesthetic evaluation. To provide a more comprehensive comparison, we conducted an additional human evaluation on 100 randomly selected samples from our 500-sample benchmark, focusing on aesthetic quality, including visual appeal, composition, and overall design coherence. The results (Elo / Win Rate) are shown below.

To reduce the workload of human evaluation, we only compare Compo-LongCat with representative baselines:

ModelsAesthetic Quality
HiDream-O1-Image935.1 / 51.2%
PosterMaker591.3 / 8.0%
PosterOmni853.4 / 31.8%
CreatiDesign1259.9 / 71.3%
Compo-LongCat (Ours)1360.2 / 87.8%

Compo achieves the highest aesthetic quality among all evaluated methods, demonstrating that improved fine-grained controllability does not come at the expense of visual quality.

2. Aesthetic filtering of training data. During the construction of Compo-200K, we incorporated aesthetic filtering using CLIP-Aesthetics score to remove visually unsatisfactory posters. This helps ensure that our training data provides supervision for both constraint satisfaction and aesthetic quality, rather than controllability alone.

W2, Q2: Effectiveness of the layout reward

Our model already achieves strong layout adherence after SFT, leaving limited room for further improvement through RL. For example, as shown in Table 2, the IoU for Identity Binding increases from 0.902 (w/o Layout Reward) to 0.904 (Full Reward), which is consistent with the reviewer's observation.

Nevertheless, layout controllability is a central objective of our framework. We therefore retain the layout reward to explicitly reinforce this capability during RL, despite its limited additional gains after SFT. Since we observed no negative impact from its inclusion, we consider it valuable for maintaining a comprehensive reward formulation covering all key aspects of controllable poster generation.

W3: Balanced Avg. Score and independent human evaluation

1. Recomputed Avg. Score. To ensure a more balanced evaluation, we recalculated the Avg. Score by first averaging the metrics within each binding type, then averaging across the four binding types. This ensures equal contributions from each binding, regardless of the number of associated metrics. The revised results are shown below:

Compo-LongCat (Ours)Compo-Qwen (Ours)Step1X-EditGLM-ImageBAGELChronoEdit-14BFLUX.2-klein-9BJoyAI-Image-EditFireRed-Image-Edit-1.1Boogu-Image-0.1-Edit
0.7400.7380.2710.3100.2040.3210.2940.3370.3020.343
HiDream-O1-ImageSenseNova-U1DreamOmni2OmniGen2LongCat-Image-EditQwen-Image-Edit-2511PosterMakerPosterOmniCreatiDesign
0.4460.1480.2630.2280.3710.3700.4660.3820.396

Both Compo-LongCat (0.740) and Compo-Qwen (0.738) continue to substantially outperform all baselines, confirming the robustness of our performance advantages under this balanced aggregation.

2. Independent human evaluation. To address the potential overlap between evaluation metrics and RL rewards, we additionally conducted independent human evaluations on 100 randomly selected benchmark samples, assessing Aesthetic Quality and Instruction Following.

The results (Elo / Win Rate) are shown below. To reduce the workload of human evaluation, we only compare Compo-LongCat with the representative baselines:

ModelsAesthetic QualityInstruction Following
HiDream-O1-Image935.1 / 51.2%938.1 / 51.5%
PosterMaker591.3 / 8.0%862.4 / 34.1%
PosterOmni853.4 / 31.8%938.8 / 37.0%
CreatiDesign1259.9 / 71.3%860.3 / 33.4%
Compo-LongCat (Ours)1360.2 / 87.8%1400.4 / 94.0%

Compo-LongCat achieves the highest Elo scores (1360.2 / 1400.4) and win rates (87.8% / 94.0%) in both dimensions among the five evaluated methods. These results provide complementary evidence that our improvements extend beyond automatic metrics to human-perceived quality and task satisfaction.

W4: Technical novelty

Our primary novelty lies in the unified formulation of multimodal reference- and layout-driven poster generation using the proposed Spatial Canvas Interface, rather than new architectures or RL algorithms. Our contributions also include a complete pipeline for dataset construction, SFT/RL training, and evaluation.

1. Unified formulation of multiple control conditions. Existing layout- or reference-driven generation methods typically focus on individual control conditions or support limited reference constraints. In contrast, Compo unifies four complementary bindings: (1) Semantic Binding associates textual concepts with target regions; (2) Identity Binding preserves reference identities while allowing pose, action, or appearance modifications; (3) Text Binding specifies exact text content and locations; and (4) Pixel Binding preserves designated visual content. Together, these enable flexible, fine-grained control within a single generation task.

2. Architecture-agnostic integration with pretrained image editing models. Existing methods often introduce task-specific modules (e.g., CreatiDesign's Layout Encoder), potentially requiring architectural adaptations for additional control conditions. In contrast, our Spatial Canvas Interface encodes spatial layouts and heterogeneous reference constraints into a single input image, accompanied by textual specifications. This reformulates multi-condition controllable generation as a standard text-and-image-to-image (TI2I) task, directly compatible with pretrained image editing models without architectural modifications. Consequently, our framework can leverage different base models through standard SFT and RL, offering a simple and extensible approach to diverse control conditions.

W5: Benchmark is independent from the training data. Real-world or synthetic ground-truth posters do not affect evaluation scores.

Our evaluation does not directly depend on the visual appearance or quality of target posters, whether synthetic or real-world.

During benchmark construction, we first generate target posters and then derive their annotations, Spatial Canvas, and text prompts through reverse annotation. During evaluation, Compo and all baselines receive the same Spatial Canvas and text prompt (or adapted inputs for dedicated methods) to generate their own posters. Importantly, evaluation metrics are computed from the generated posters and their ground-truth annotations, rather than image-level comparisons with the original target posters. Thus, the synthetic nature of target posters does not directly affect evaluation scores through similarity to synthetic ground truth.

Furthermore, Appendix Figure 8 (Page 22) presents qualitative examples using real-world themes and visual assets, providing additional evidence of Compo's ability to generalize beyond synthetic training content.

W6: Independent human evaluation on multiple aspects

We conducted additional human evaluation on 100 randomly selected samples from our 500-sample benchmark, covering four dimensions: (1) Aesthetic Quality, assessing visual appeal; (2) Readability, evaluating text legibility and clarity; (3) Instruction Following, assessing adherence to task requirements; and (4) Overall Design Quality, evaluating composition, visual coherence, and overall design effectiveness.

The results (Elo / Win Rate) are shown below. To reduce the workload of human evaluation, we only compare Compo-LongCat with the representative baselines:

ModelsAesthetic QualityReadabilityInstruction FollowingOverall
HiDream-O1-Image935.1 / 51.2%848.0 / 39.4%938.1 / 51.5%943.0 / 53.9%
PosterMaker591.3 / 8.0%590.7 / 6.1%862.4 / 34.1%806.7 / 26.9%
PosterOmni853.4 / 31.8%976.2 / 46.9%938.8 / 37.0%998.0 / 42.1%
CreatiDesign1259.9 / 71.3%1261.5 / 73.3%860.3 / 33.4%863.8 / 34.6%
Compo-LongCat (Ours)1360.2 / 87.8%1323.6 / 84.4%1400.4 / 94.0%1388.5 / 92.5%

Compo achieves the best performance across all four dimensions, demonstrating that its fine-grained controllability is accompanied by superior human-perceived visual quality, readability, and overall design effectiveness. These independent human evaluations provide complementary evidence beyond automatic metrics and RL rewards, further supporting the effectiveness of our evaluation protocol.

W7: Annotation quality of the proposed dataset

To quantitatively assess annotation quality, we randomly sampled 100 training samples from Compo-200K dataset and manually verified both OCR and segmentation results. We adopt an image-level success criterion: a sample is considered correctly annotated only if all text is accurately recognized and all target objects are successfully segmented. Under this criterion, our pipeline achieves a 93% annotation success rate, supporting its reliability.

Furthermore, as described in Section 3.1 (Page 4) and Appendix Section A.1 (Page 16), we deliberately excluded overly uncommon object categories, actions, and poses when constructing our text asset repository, reducing the difficulty of automatic segmentation and annotation.

Q1: Quality of the synthetic dataset and its effectiveness for training

1. Quality of source image generation models. For poster creation, we employ strong open-weight models like Qwen-Image-2512, offering competitive instruction following, text rendering, and visual quality. Similar approaches have been adopted in peer-reviewed studies: PosterOmni (CVPR 2026) uses Qwen-Image for ground-truth poster dataset creation, while CreatiDesign (ICLR 2026) employs FLUX.1-dev for target poster background generation. These precedents support the feasibility of using pretrained models for synthetic poster data construction.

2. Quality filtering and validation. Beyond selecting capable source models, we apply CLIP-Aesthetics to filter visually unsatisfactory samples. Furthermore, we conduct a manual validation of annotation quality. Please see W7 for details.

3. Effectiveness of training data. The quantitative results in Table 1 (Page 7), the qualitative comparisons in Figure 5 (Page 8) and Appendix Figures 6-8 (Page 20-22), and the independent human evaluations in W6 demonstrate Compo's strong performance after training on our synthetic dataset, providing empirical evidence for the effectiveness and quality of our data construction pipeline.

Q3: Additional visualizations of real-world poster design

1. Real-world poster design examples. To demonstrate Compo's practical applicability, we visualize two additional live-action and animated movie poster examples generated from text-only requests using Agentic Mode. The manuscript PDF has been updated, please take a look at Appendix Section D (Page 18).

As presented in Appendix Section D.1 (Page 18) and Appendix Figure 9 (Page 23), Compo generates visually coherent posters for live-action and animated movies, demonstrating its generalization capability and practical applicability.

2. User Study. Please See W6 for details.

Reviewer_xGZJOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Compo, which uses a spatial canvas to unify semantic, identity, text, and pixel constraints for joint control over visual elements and text layout in poster generation. The authors construct approximately 200K synthetic training examples and adapt image editing models through supervised fine-tuning and DiffusionNFT-based reinforcement learning, supporting both direct canvas input and agentic planning. On a newly constructed 500-example benchmark, the method achieves strong results in text accuracy, instance preservation, and layout adherence.

Critical strengths
  1. A complete pipeline with practical value. The spatial canvas, automatic data construction, and binding-aware reward optimization form a coherent pipeline for jointly controlling multiple text regions and visual elements. This provides a more explicit way to specify complex poster designs than text descriptions alone.
  2. Compo-LongCat achieves an average score of 0.752 on the proposed benchmark, compared with 0.504 for the strongest evaluated general-purpose baseline; RL further improves sentence accuracy from 0.886 to 0.940. These results support the practical value of the pipeline and its post-training for text and layout control.
Critical weaknesses
  1. The advantage over strong alternatives is not sufficiently established. The main comparisons focus on open-weight models and specialized poster generation methods, omitting recent leading commercial models and detailed-prompt baselines. It therefore remains unclear whether the additional canvas construction and agentic workflow provide benefits that existing strong models cannot readily achieve.
  2. Insufficient independent validation of data and outcomes. Target posters and annotations are largely constructed automatically, while RL rewards and evaluation heavily share recognition tools and metrics, without sufficient human validation of data quality or perceptual outcomes. This leaves uncertainty about whether the reported scores reflect actual design quality and how reliably the method handles failure-prone cases.
Questions for authors
  1. Have the authors compared against recent leading commercial models, particularly a baseline in which the same agent produces a detailed text prompt for direct poster generation? Could this comparison establish the independent benefit of the spatial canvas over text-only planning?
  2. Given that the target posters are generated using open-weight models, have the authors evaluated the quality–cost trade-off of using leading closed-source models to generate or filter the data? What targeted data cleaning procedures were applied?
  3. Are there independent human evaluations showing that improvements in automatic metrics correspond to better text rendering, layout, and aesthetics? Was reward hacking observed during RL—where rewards increased without corresponding improvements in actual quality—and, if so, how was it detected and mitigated?
  4. What are the main limitations when handling dense, long-form text, conflicting constraints, or failures to obtain reference assets? Please provide representative failure cases of the proposed method and clarify its applicability limits.
  5. Compared with generation systems such as Instagram-v4 and two-stage pipelines that use models such as GPT-6 to generate code followed by rendering, what trade-offs does Compo offer in text accuracy, editability, and visual integration? Is there evidence from comparable tasks identifying the scenarios in which it is advantageous?
5 comments

At first, we sincerely thank you for your valuable time, constructive comments, and positive rating.

W1, Q1: Comparisons with leading commercial models and the independent benefit of the Spatial Canvas

1. Comparison with leading commercial models. To further assess our approach, we additionally compare Compo-LongCat with ChatGPT Images 2.6 on 50 benchmark cases through human evaluation, focusing on Aesthetic Quality and Instruction Following.

The results (Elo / Win Rate) are shown below:

ModelsAesthetic QualityInstruction Following
ChatGPT Images 2.6 (w/ Detailed Prompt)1065.6 / 64.5%928.7 / 37.0%
ChatGPT Images 2.6 (w/ Spatial Canvas + Text Specifications)1056.1 / 59.0%1020.4 / 55.5%
Compo-LongCat (Ours)878.4 / 26.5%1051.0 / 57.5%

Compo-LongCat achieves slightly better Instruction Following than ChatGPT Images 2.6, demonstrating competitive controllability against a leading commercial model. Meanwhile, ChatGPT Images 2.6 achieves substantially better Aesthetic Quality while maintaining strong instruction-following capabilities. We believe this gap is primarily related to differences in the aesthetic quality and diversity of training data, where commercial models likely benefit from higher-quality training resources.

2. Effectiveness of the Spatial Canvas Interface. To isolate the benefit of our representation, we further compare ChatGPT Images 2.6 under two input settings: Detailed Prompt (text-only) and Spatial Canvas + Text Specifications. Introducing the Spatial Canvas substantially improves instruction following (Elo: 928.7 → 1020.4; Win Rate: 37.0% → 55.5%), indicating more accurate adherence to reference and layout constraints. This provides direct evidence that our Spatial Canvas representation offers benefits beyond detailed text-only instructions, even for strong commercial models.

W2, Q3: Evaluation-reward overlap and independent human evaluation

1. Evaluation metrics and RL reward design. We established our benchmark and evaluation metrics before introducing RL, specifically to capture task-relevant aspects of human preference. For example, Sentence Accuracy (Sen. Acc.) measures whether an entire poster is free of text rendering errors, directly reflecting a critical aspect of poster quality. Using evaluation-aligned metrics as RL rewards is common practice in both LLM and diffusion model post-training. For instance, DeepSeek-R1[1] employs rule-based accuracy rewards for mathematics and coding, while Flow-GRPO[2], DanceGRPO[3], and DiffusionNFT[4] optimize task-specific objectives using rewards such as GenEval, OCR accuracy, HPSv2.1, and PickScore. Following this practice, we adopt selected evaluation-aligned metrics as RL rewards to further improve constraint satisfaction. We believe such overlap is reasonable when the metrics meaningfully reflect the desired task outcomes.

To validate whether improvements in automatic metrics translate into perceptual improvements, we conducted two human evaluations on 100 randomly selected samples from our 500-sample benchmark, assessing Aesthetic Quality and Instruction Following. These evaluations are independent of the automatic metrics and recognition tools used for RL rewards.

The results (Elo / Win Rate) are shown below. To reduce the workload of human evaluation, we only compare Compo-LongCat with representative baselines:

ModelsAesthetic QualityInstruction Following
HiDream-O1-Image935.1 / 51.2%938.1 / 51.5%
PosterMaker591.3 / 8.0%862.4 / 34.1%
PosterOmni853.4 / 31.8%938.8 / 37.0%
CreatiDesign1259.9 / 71.3%860.3 / 33.4%
Compo-LongCat (Ours)1360.2 / 87.8%1400.4 / 94.0%

Our method achieves the best results in both dimensions, providing independent human evidence supporting its improvements beyond automatic metrics.

[1] Guo, Daya, et al. "Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning." arXiv preprint arXiv:2501.12948 (2025).

[2] Liu, Jie, et al. "Flow-grpo: Training flow matching models via online rl." Advances in neural information processing systems 38 (2026): 40783-40818.

[3] Xue, Zeyue, et al. "Dancegrpo: Unleashing grpo on visual generation." arXiv preprint arXiv:2505.07818 (2025).

[4] Zheng, Kaiwen, et al. "Diffusionnft: Online diffusion reinforcement with forward process." International Conference on Learning Representations. Vol. 2026. 2026.

2. Reliability of automatic dataset construction pipeline. Our automated data construction and annotation pipeline was designed for scalability and reliability. We deliberately avoided overly complex or uncommon object categories, identities, and actions/poses, reducing the difficulty of automatic segmentation, reposing, and relighting. To quantitatively verify its reliability, we randomly sampled 100 training samples from Compo-200K dataset and manually verified both OCR and segmentation results. We adopt an image-level success criterion: a sample is considered correctly annotated only if all text is accurately recognized and all target objects are successfully segmented. Under this criterion, our pipeline achieves a 93% annotation success rate, supporting its reliability.

3. Reward hacking during RL. We observed no obvious signs of reward hacking. To further investigate this, we independently evaluated the SFT model (Compo-LongCat-Base) and its RL-enhanced counterpart (Compo-LongCat) on 100 randomly selected benchmark samples, focusing on Instruction Following, which is the core ability of our proposed models. The results (Elo / Win Rate) are shown below.

ModelsInstruction Following
Compo-LongCat-Base (SFT)930.4 / 30.5%
Compo-LongCat (RL)1069.6 / 69.5%

The RL model substantially outperforms its SFT counterpart in human evaluation, suggesting that RL improvements extend beyond automatic reward metrics to human-perceived task performance.

Q2: Quality-cost trade-off of using open-source or leading closed-source models to construct the dataset

Our dataset construction pipeline consists of two stages: poster creation (ground-truth creation) and annotation generation. For poster creation, we open-weight models like Qwen-Image-2512, which offer strong instruction-following, text-rendering, and aesthetic capabilities.

While leading closed-source models (e.g., ChatGPT Images 2.6 and Nano Banana 2) may offer higher quality, their API costs pose challenges for large-scale data construction. Based on public Standard API pricing, generating 200K posters at 1024×1024 resolution would cost approximately $10,536 with ChatGPT Images 2.6 ($0.05268/image, High Quality) or $13,440 with Nano Banana 2 ($0.0672/image, 1K). Even with the discounted Batch API, Nano Banana 2 would cost approximately $6,720 for image outputs alone. These estimates exclude input tokens, retries, and data filtering overhead. In contrast, our open-weight models run on existing GPU infrastructure, avoiding per-image API charges and substantially reducing marginal generation costs.

To further assess this trade-off, we conducted an additional human evaluation on 50 poster samples, comparing closed-source models against Qwen-Image-2512. The results (Elo / Win Rate) are shown below:

ModelsOverall Quality
ChatGPT Images 2.61077.0 / 68.0%
Nano Banana 2981.5 / 41.5%
Qwen-Image-2512941.4 / 40.5%

Although closed-source models achieve higher quality, Qwen-Image-2512 remains competitive, particularly against Nano Banana 2, while avoiding substantial API costs. These results support our choice of open-weight models as a practical quality-cost trade-off for large-scale dataset construction.

Q4: Failure cases and applicability limitations

To illustrate Compo's limitations, we visualize two additional representative failure cases involving long-form text rendering and conflicting constraints. The manuscript PDF has been updated, please take a look at Appendix Section D (Page 18).

As shown in Appendix Section D.2 (Page 18) and Appendix Figure 10 (Page 23), in Failure Case 1, given the text "Ready for Every Trip\nBuy at Good Prices\nSuitable for ICLR 2027 Travel\nIn Apr. 2027, USA", Compo incorrectly inserts an extra "o" before "ICLR". This may result from the predominance of short slogans rather than long-form text in our training data. In Failure Case 2, Compo struggles to jointly satisfy multiple interdependent constraints on character placement, pose, and interaction with the bicycle, resulting in incorrect spatial relationships and an implausible visual composition.

These cases clarify Compo's applicability limits: while effective for poster generation involving relatively short text, multimodal visual references, and well-defined spatial constraints, its reliability may decrease with long-form text or complex, potentially conflicting constraints that require precise spatial interactions between visual elements. These remain important directions for future improvement.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.