acceptodds
Withdrawn from ICLR 2027

CreatCtrl: Controllable Creative Graphic Design Generation from References

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Creativity is essential in graphic design. Despite recent advances in graphic design generation, controlling the creativity of generated designs remains unexplored. In this paper, we investigate the problem of creativity control for graphic design generation, and propose a model, CreatCtrl, which learns to elevate the creativity of reference designs in a controllable manner. Our model takes as input a textual design intention, a reference design in raster format, and a continuous creativity parameter, and generates more creative, layered graphic designs. Our model enables users to control the creativity degree of the generated designs by adjusting the creativity parameter in a continuous way, which allows them to explore a wide range of creative variants of the reference to find design inspiration. To train our model, we propose a dataset construction method to synthesize training tuples, each consisting of a reference design, its creative variant, and a parameter representing the creativity gap between them, a reliability weight of the creativity parameter. Our experiments demonstrate that our method outperforms baselines in design quality and creativity controllability, showing the effectiveness and potential of our creativity‑controllable generation paradigm for real‑world design applications.

Primary area:
generative models
Keywords:
Graphic Design Generation

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

This work propose reference-based graphic design generation with a continuous creativity control parameter. For the given inputs (e.g., reference design, a textual design intention, etc), the devised method generate a layered design while exhibiting a controllable degree of creativity. In particular, authors fine-tunes VLM with a regression head on 6k designs with human creativity ratings. This scoring model is used to rank large scale dataset and labeled dataset is used to train the generator. The evaluation includes distributional design metrics, text alignment, an Multimodal LLM evaluator, and a study with 10 professional graphic designers. The method obtains reasonable quantitative and qualitative results. Although the evidence for the central claim of meaningful, generalizable creativity control is substantially weaker than the evidence for producing appealing design variations.

Critical strengths

(1) The problem is useful and challenging.

  • Helping designers explore variations around a reference could be valuable in practical design workflows. Balancing novelty, communication goals, and reference relevance is a difficult problem, and an effective AI tool for this purpose would have clear potential benefits.
Critical weaknesses

(1) The creativity scorer lacks sufficient validation, and the necessity of fine-tuning is not established.

  • The entire pipeline depends on a scalar creativity score learned from human ratings, but Section 3.1 does not explain who supplied the 6,000 designs' labels, how many designers participated, how many ratings each design received, what rubric they followed, or how disagreements were aggregated. The 10 professional designers are explicitly described for the later user study. It is unclear whether they also supplied the scoring labels. Without this information, it is difficult to assess whether the scorer captures broadly meaningful design creativity or the preferences of a particular annotator group.

  • The reported MAE of 0.50 on 600 held-out designs is not compared with a prompted pretrained VLM, a frozen VLM with a trained regression head, or a simple score-prediction baseline. A strong pretrained VLM might already predict these labels reasonably well, but the paper does not test this. Consequently, the results do not establish either the need to fine-tune the VLM or the benefit of the additional training. Human rating variability is also needed to interpret the MAE.

  • The reported model-human consistency of 0.71 is similarly difficult to interpret without scorer baselines, human-human agreement for this specific task, sample counts, and a precise definition of consistency. Moreover, the triplets are drawn from widely separated score strata, making this a relatively coarse ordering test. Success on these triplets does not establish reliable distinctions between nearby creativity levels, which are important for continuous control.

(2) The paper does not adequately distinguish creativity from visual complexity, style, or attractiveness.

  • The paper treats creativity as a scalar attribute but does not provide a sufficiently explicit operational definition or annotation rubric. For graphic design, originality must be considered together with the communication goal and the appropriateness of the solution. A minimalist poster can be highly creative, while a visually elaborate poster can be conventional or ineffective.

  • Figure 2 raises the concern that colorful, abstract, or visually elaborate designs may receive higher scores. Some variations in Figure 3 also change illustration style and color. These examples do not prove that the score model relies on such shortcuts, and the figure also includes restrained designs. However, the paper provides no controlled analysis that rules out this explanation. For example, it could compare designs with matched content and layout while varying saturation or ornamentation, and evaluate creative minimalist designs against visually complex but conventional ones.

  • There is also a conceptual gap between an absolute score difference, s(Y) - s(X), and creativity relative to a reference. Two existing designs with different absolute ratings need not form a meaningful creative transformation. Retrieving a target using visual embedding similarity does not itself establish that the target preserves the reference's brief or represents an original reinterpretation of it.

(3) Required disclosure and citation presentation need correction.

  • I could not find an AI use statement in the supplied manuscript. ICLR 2027 requires a dedicated AI use disclosure section. The authors should add the required statement describing relevant AI use in the research and manuscript preparation.

  • Minor: The citations are also difficult to read. Many appear as unparenthesized author-year strings embedded in the prose, especially in long citation lists and statements such as “Qwen3-VL-30B Bai et al. (2025).” Parenthetical citations (e.g., "\citep") where appropriate would improve readability.

Questions for authors

(1) Who provided the creativity ratings for the 6,000 Awwwards designs? Please report the number and backgrounds of annotators, ratings per design, annotation rubric, aggregation procedure, and inter-rater agreement. Were these annotators independent of the later evaluation participants?

(2) How is creativity distinguished from visual complexity, saturation, ornamentation, stylistic change, and aesthetic quality? What evidence shows that the score model handles creative minimalist designs and reference-relative originality appropriately?

(3) What permissions, licenses, or other basis support the collection and use of Awwwards and Freepik designs, including model training, publication of examples, and any proposed redistribution? Please also clarify provenance and usage rights for retrieved image and SVG assets.

(4) Was the need for IRB or equivalent ethics review assessed for both annotation and the designer study? Please report approval, exemption, or the applicable determination that review was not required, together with consent, compensation, and participant-data handling procedures.

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

This paper presents a method for generating new graphic design with a reference design and a creativity scale. The proposed method basically include a training data generation process and a layout generation process. The experiment results suggest that the proposed method outperform other similar methods.

Critical strengths
  • The proposed task, which create design variation is important and could potentially enhance the efficiency of design.
  • The proposed "training data generation" pipeline could be useful for other design task.
Critical weaknesses
  • While the task is important and the resulting method might be useful for design application, I am not that confident about the research contribution this work can make.
    • Specifically, the overall work fit into "collect/create a tailored data from LLM, train a tailored model using these data, and then compete with other general purpose method" work.
    • Because of this, the main method part imho is the combination of existing works. And the overall paper is like a manual of how to combine these technique, but not about which part is not discovered by previous methods.
  • Furthermore, because of the previous point, I believe the comparison is not fail to other methods as well.
    • For example, I believe the image editing Kontext mostly used to do is not about graphic design editing.
    • And it is not intuitive or does not make sense why "edit strength" can be seen as creativity? (across the comparison between TIE and LLM? If this is the case, that means we can simply achieve higher creativity by edit more?
  • More importantly, I think it is important to measure the preservation of the reference design while generating the design variation. Otherwise, it means that we can generate a total independent design from the reference design with an arbitrary "creativity degree". For example, for results shown in Fig. 1 and Fig. 3, what are the common properties or features preserved between design? And if yes, how exactly they are preserved?
    • For me, I guess the only preservation is done implicitly comes from the input prompt. But for professional graphic design, designers usually require more accurate and detailed controls instead of relying coarse input such as text prompt. Therefore, I encourage authors to think about how to achieve better content preserving in the method, and where to apply "creativity for variation".
Questions for authors

Please check "Critical Weakness".

Reviewer_d9gkOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces creativity-controllable graphic design generation. Given a raster reference design, a text intention, and a scalar β ∈ [0, 1], CreatCtrl generates a layered design that is meant to stay related to the reference and be more creative, with β setting how much more. Training data is synthesized as follows. A Qwen3-VL-based creativity scorer is fine-tuned on 6k Awwwards designs rated by humans. It scores a 17k-design corpus, and the ranked corpus is split into a low-creativity reference pool and a high-creativity target pool. Each reference is paired with its nearest neighbour in each of five creativity strata. β is the per-reference normalized score gap, and pairs are re-weighted by a VLM consistency check. The generator is an autoregressive Transformer over element-attribute tokens, conditioned on CLIP features via cross-attention, with β injected through AdaLN. Rendering uses SDXL/retrieval for images and SVGs and an LLM for text content. The paper reports gains over three baselines on FID, CLIP-score and C-FID, plus a 10-designer user study.

Critical strengths

1.New, relevant problem. Controlling creativity relative to a reference matches a real designer workflow, and as far as I know it has not been formulated this way before for layered design generation. 2.Layered, editable output. This is practically valuable, and it avoids the distorted-text failure that the designers noted for the editing baselines. 3.Professional designers in the study. The user study uses professional designers, and the "creativity inspiration" result (59.6%) directly supports the motivation.

Critical weaknesses

W1. The whole method rests on human creativity annotations whose reliability is never established. The labels feed the scorer, which in turn determines the pool partition, the stratified pairing, and β. Yet the paper describes the annotation in a single sentence ("creativity ratings from human designers on a scale of 0 to 10").

Protocol not reported. The paper does not say who rated, how many raters scored each design, which rating guidelines were used, or whether the scores are Awwwards' existing jury scores or were collected by the authors. No inter-rater agreement. There is no ICC, Krippendorff's α, or per-design rating variance. For a construct as subjective as "creativity", this is the first thing that needs to be shown. The indirect evidence does not substitute for agreement. A test MAE of 0.50 shows that the scorer fits the labels, not that the labels are consistent. Without the human–human disagreement as a ceiling, 0.50 cannot be interpreted. The 0.71 model–human consistency is measured only on easy triplets (top/middle/bottom 20%). The number of participants, the exact metric (full ranking or pairwise), and the human–human agreement are not reported. Most importantly, it does not test the fine-grained differences within the high-creativity pool that β depends on. The 82% VLM–human agreement validates a different model (the one used for reweighting), and the 0.82 user-study consistency concerns rankings of generated outputs. Neither speaks to the training labels. Domain shift. Awwwards is a web-design award site, whereas the target domain is posters and social-media designs. The scorer is never validated on human ratings in the target domain.

If the labels are noisy, β has no reliable monotonic meaning, which may also explain the weak controllability in W3.

W2. Limited technical novelty: the method is essentially "construct paired data, then train a standard conditional generator".

The model is assembled from existing components, and the paper says so. The tokenized design representation and autoregressive generator follow GOL (Yang & Cao 2025). Reference and intention conditioning follows IP-Adapter. β injection via AdaLN follows DiT. CFG, two-stage training, and a weighted NLL loss are standard. Rendering relies on SDXL+IP-Adapter, generate-or-retrieve (Koh et al. 2023), SVG retrieval and an LLM. Nothing in the architecture or objective is specific to creativity; β is an ordinary scalar condition. No mechanism actually targets creativity. The model imitates a nearby, more highly scored existing design, so at best it reproduces the D_high distribution. It is never encouraged to go beyond the training targets. The scorer is never used as a reward or guidance signal (e.g., RL, DPO or classifier guidance), so the behaviour is closer to learned conditional retrieval than to creative generation. The contribution therefore rests on the data pipeline, which is heuristic and unvalidated. The 5:3:1 split, five strata, per-reference min–max β, ω = 5 vs. 1, and the "all tuples positive" rule are not justified or ablated. A task-plus-data paper can be a valid ICLR contribution, but then the data needs rigorous validation (W1) and ablations (W6), and both are missing.

W3. Controllability is weak. In the paper's own user study, CreatCtrl's Kendall τ is 0.307, versus 0.867 for the prompt-based LMM baseline. Because β is min–max normalized over five pairs per reference, it acts as about five discrete levels, and "continuous" control is never evaluated, for example with a dense β sweep or s(output) − s(reference) as a function of β.

W4. C-FIDβ does not measure creativity. It measures distance from the references, so any drift scores higher. NNS reaches 99.14 at β = 0.9 while its reference relevance is only 8%.

Questions for authors

1.What is the annotation protocol (who rated, how many raters per design, guidelines)? Are these Awwwards' own jury scores? Please report inter-rater agreement (ICC or Krippendorff's α) and the human–human MAE as a ceiling for the scorer. 2.Can you validate the scorer on fine-grained pairs (e.g., adjacent strata Q_k, Q_{k+1}) and on human ratings collected in the target domain (posters and social posts)? 3.Beyond the data pipeline, what is technically new in the model? How does CreatCtrl compare with (a) the same model fine-tuned without β and (b) a globally normalized β?

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

Related papers

Discussion (0)

Sign in to comment.