acceptodds
Withdrawn from ICLR 2027

DecoupleGen: Intent-Aware CoT Layout Planning and Controllable Text Rendering for Video Covers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Video covers are often a viewer's first encounter with a video, yet their typography must be composed over a content-bearing frame whose people, objects, actions, and native screen content remain recognizable. We study this progressive design problem with DecoupleGen, which connects frame selection, intent-aware layout planning, and background-conditioned text rendering through an explicit text–quadrilateral–color interface. The frame selector identifies a relevant and visually usable background. The planner first produces a natural-language rationale for title hierarchy, placement, and contrast, then predicts structured text regions and quantized colors. A diffusion renderer converts that plan into rasterized spatial and color conditions and uses localized OCR-feature supervision to strengthen character fidelity. We introduce Cover-100K, containing 99,781 aligned examples of recovered clean backgrounds, completed covers, strings, geometry, colors, and generated design rationales. On Cover-600 with supplied clean backgrounds, DecoupleGen improves text accuracy and distributional image quality over released editors, while human raters favor its planned layouts over PosterLlama. Component studies support the contributions of cover-domain supervision and structured rendering. Together, the dataset and structured interface provide a practical path from video content and design intent to readable covers.

Primary area:
generative models
Keywords:
Diffusion model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xB5dOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DecoupleGen to address the problem of video cover generation. The authors construct the Cover-100K dataset, which contains 99,781 pairs of aligned samples covering a wide range of information, including clean backgrounds, complete covers, text content, geometric layouts, color information, and the design rationale behind the generated images. Experimental results demonstrate that DecoupleGen achieves superior generation performance. User experiments further indicate that, compared to other methods, users prefer the results generated by DecoupleGen.

Critical strengths
  1. This paper constructs the Cover-100K video cover dataset, which is organized around a curated collection of high-quality video covers and provides feature information across multiple dimensions.

  2. The authors propose the DecoupleGen method, which is based on intent-aware CoT layout planning. Experimental results validate the effectiveness of this method.

Critical weaknesses
  1. There are issues regarding the fairness of the comparative experiments. Section 5.2 notes that "none were fine-tuned on Cover-100K"; this implies that the method proposed in the paper utilizes Cover-100K for supervised fine-tuning, whereas some baseline models were not fine-tuned on the same dataset. We recommend conducting comparative experiments under a unified training setup.

  2. The experimental analysis lacks depth. The current analysis mainly focuses on the improvement of indicators such as FID and the presentation of qualitative results. The paper does not conduct in-depth analysis of failure cases. Therefore, it is currently difficult to determine in which specific scenarios each module is effective and the reasons behind the performance improvement. The existing ablation experiments mainly demonstrate the effectiveness of each module, but the analysis of its underlying mechanism is still insufficient.

  3. There is a lack of generalization experiments across different backbone networks. The current method is mainly verified through Qwen-Image-Edit. It is recommended to further test the combination of DecoupleGen with other generation models as the backbone network to verify whether the performance improvement is independent of the model and to demonstrate stronger generalization ability.

Questions for authors
  1. There are issues regarding the fairness of the comparative experiments. Section 5.2 notes that "none were fine-tuned on Cover-100K"; this implies that the method proposed in the paper utilizes Cover-100K for supervised fine-tuning, whereas some baseline models were not fine-tuned on the same dataset. We recommend conducting comparative experiments under a unified training setup.

  2. The experimental analysis lacks depth. The current analysis mainly focuses on the improvement of indicators such as FID and the presentation of qualitative results. The paper does not conduct in-depth analysis of failure cases. Therefore, it is currently difficult to determine in which specific scenarios each module is effective and the reasons behind the performance improvement. The existing ablation experiments mainly demonstrate the effectiveness of each module, but the analysis of its underlying mechanism is still insufficient.

  3. There is a lack of generalization experiments across different backbone networks. The current method is mainly verified through Qwen-Image-Edit. It is recommended to further test the combination of DecoupleGen with other generation models as the backbone network to verify whether the performance improvement is independent of the model and to demonstrate stronger generalization ability.

Reviewer_1SghOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DecoupleGen creates video covers while preserving the selected frame. It first generates a design rationale and a structured layout with text regions and quantized colors, then uses a diffusion renderer with spatial and color controls and localized OCR-feature supervision. The authors introduce Cover-100K with 99,781 aligned examples and evaluate on Cover-600. Results show better text accuracy and image quality than several released editors, while human raters prefer its planned layouts to PosterLlama. Ablations suggest that cover-specific fine-tuning and explicit rendering controls both contribute to the gains.

Critical strengths
  1. Separating layout planning from rendering gives the system a clear and inspectable way to control text, position, and color.
  2. The evaluation includes OCR metrics, image quality measures, human ratings, and component ablations. The results provide useful evidence that the proposed controls improve cover generation.
Critical weaknesses
  1. The main Cover-600 evaluation uses supplied clean backgrounds, so it does not validate the full pipeline, especially frame selection on raw videos.
  2. The paper provides limited evidence of generalization to video sources and cover styles outside the collected data. A more diverse external evaluation would strengthen the claims.
Questions for authors
  1. Can the authors report an end-to-end evaluation on raw videos that includes frame selection, layout planning, and rendering?
Reviewer_NZ9jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The goal of this paper is to generate a plausible video cover from a video with design intent. A three-stage framework, i.e., DecoupleGen, is proposed with frame selection, intent-aware layout planning, and background-conditioned text rendering. A new dataset, Cover-100K, is further constructed with 99,781 aligned examples. Experiments show the effectiveness of the proposed method.

Critical strengths

The idea of decoupling video-cover design into frame selection, layout planning, and controllable rendering is a reasonable and practical solution. The constructed large-scale dataset could be useful for future research in the community once released.

Critical weaknesses
  1. The technical contribution is relatively limited. The proposed DecoupleGen appears to be a straightforward combination of several existing tasks, including video thumbnail selection, layout planning, and rendering. The core modules adopted in the framework are mainly built upon existing off-the-shelf foundation models (Qwen3-VL-8B, Qwen-Image-Edit).
  2. While frame selection is a core stage of the progressive video-to-cover problem, it is not involved in the DecoupleGen framework or evaluated, as claimed in the abstract.
  3. Unlike the proposed method, none of the compared baselines are fine-tuned with the proposed dataset, making the evaluations unfair and the reported results less convincing.
  4. The experiments are not thorough enough to prove the superior performance of the proposed method. Since DecoupleGen contains three stages, the performance of each stage should be evaluated separately against existing specialized works.
  5. The quantitative evaluation metrics only contain three scores, lacking standard layout metrics and image rendering metrics used by existing works.
  6. Since the proposed method involves several stages, a detailed analysis of computational cost and running time is necessary.
Questions for authors
  1. How does the proposed frame selection process quantitatively compare against existing video thumbnail selection baselines?
  2. How do the authors ensure that the generated layout is aligned with design intent without suffering from MLLM bias?

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

Related papers

Discussion (0)

Sign in to comment.