PaperlayerView exact revision on arXiv

Source-linked AI summary

DOGS: Design-Space Sampling for Prompt-Driven Logo Generation

Ganyu Zou, Chen Dai, Nathan Self, Kevin Piper, Ramachandra Rao Seethiraju, Karthik Shyamsunder, Chang-Tien Lu, Naren Ramakrishnan

arXiv:2610.10760v1cs.CV

TL;DR

Prompt optimization for T2I generation largely rewrites text, but brief logo prompts leave design choices unspecified and linear sequences cannot capture their relational structure. DOGS samples designs from a corpus-induced graph grammar using a reward that combines recognizability, aesthetics, and originality. Across two open-source renderers and nine baselines, it produces logos that are more recognizable and aesthetic, more diverse, and less prone to trademark infringement.

  • Problem

    Text-rewriting prompt optimization targets general image quality, while logo briefs leave essential design choices unspecified and linear prompts do not represent the relational priors needed for logo design.

    Source: §1 · p. 2

  • Method

    DOGS uses a corpus-induced typed graph grammar and a GFlowNet sampler to complete symbolic logo designs in proportion to a reward combining recognizability, aesthetics, and corpus-relative originality.

    Source: §1 · p. 3

  • Results

    Across two open-source renderers and nine baselines, DOGS produces logos that are more recognizable and aesthetic, substantially more diverse, and far less prone to trademark infringement.

    Source: §1 · pp. 3–4

  • Takeaways & Limitations

    The results suggest that the design-space formulation extends beyond logos to structured visual design.

    Source: §5 · p. 14

Abstract

from arXiv · show

Prompt optimization for text-to-image (T2I) generation has been pursued almost entirely as text rewriting, in which a short user brief is expanded into a longer, model-preferred token sequence. We argue that such a language-space formulation is ill-suited to structured visual design tasks such as logo creation, where a one-line brief leaves most design decisions unspecified. These decisions depend on relational priors that a linear sequence cannot encode, and they leave an uncontrolled channel through which protected marks may be reproduced. We therefore recast logo prompting as sampling within a structured design space, and instantiate this idea as DOGS (Design-space prompting with an Originality-aware GFlowNet Sampler). From a large corpus of real-world logos, we mine a typed, graph-structured design grammar whose edges record empirical co-occurrence. A GFlowNet sampler then generates design graphs with probability proportional to a terminal reward that combines recognizability, aesthetics, and corpus-relative originality. Every slot draws only from a closed design-level vocabulary, and any infringement-inducing or harmful token is removed during parsing. The originality reward further penalizes proximity to existing logos, thereby incorporating infringement avoidance into the method by construction. On two open-source renderers and against nine baselines, DOGS produces logos that are more recognizable and aesthetic, substantially more diverse, and far less prone to trademark infringement.

1 Introduction

Logo briefs leave most design decisions unspecified, while text-rewriting methods do not represent the relationships and combinations needed to search structured designs. DOGS instead samples graph-structured logo designs and reports stronger results across quality, diversity, and infringement measures.

1 Introduction

  • A one-line brief leaves most design decisions unspecified, although usable logos require recognizability, originality, and conformance to design conventions beyond generic image quality.Logo users are often non-experts, including brand owners, startup teams, and independent content creators.Source: §1 · p. 2
  • Text-rewriting methods tune phrasing but do not choose the logo's entity, layout, or typography, and linear sequences cannot encode relational priors or deliberately search combinations.The paper identifies infringement risk as another gap: free-text pipelines lack an explicit objective that measures and penalizes proximity to existing logos.Source: §1 · pp. 2–3
  • DOGS mines a typed graph grammar from real-world logos, with design elements as nodes and empirical co-occurrence and design relations represented in the grammar.The grammar captures which elements co-occur, their roles, and spatial relations.Source: §1 · p. 3
  • A design-level whitelist excludes trademark names during parsing, while a corpus-relative originality reward penalizes proximity to existing logos during sampling.The terminal reward also combines recognizability and aesthetics.Source: §1 · p. 3
  • DOGS makes symbolic design composition the search variable and uses GFlowNet sampling to distribute probability across high-reward combinations and multiple valid designs.The grammar supplies relational structure absent from linear prompt rewriting.Source: §1 · p. 3
  • Across two open-source renderers and nine baselines, DOGS produces logos that are more recognizable and aesthetic, substantially more diverse, and far less prone to trademark infringement.The evaluation includes brand, company, and design prompts at decreasing levels of explicitness.Source: §1 · pp. 3–4

2 Related Work

Related work spans language-space prompt optimization, generation with explicit structure, and reward-driven training. DOGS combines structured design-space search with reward-proportional sampling, while keeping the renderer fixed.

2 Related Work

  • Prompt-optimization methods either rewrite short inputs with language models trained on aesthetic or preference signals, or directly optimize discrete prompts through search methods.Examples include supervised fine-tuning, reinforcement learning, preference optimization [2] [9] [17] [26], hard-prompt search [34], genetic search [22], and shortcut text gradients [33].Source: §2 · p. 4
  • Explicit-structure generation either generates structured objects such as graphs and trees or constrains a pretrained decoder with a formal grammar.Graphics work also defines and learns visual design spaces from grammars and layout or scene corpora, including GRAINS [14], LayoutTransformer [6], and LayoutDM [10].Source: §2 · p. 4
  • Reward-driven methods include expected-reward reinforcement learning, preference-pair optimization, and approaches that spread probability across multiple high-reward modes.GFlowNets sample trajectories in proportion to terminal reward [1], with trajectory balance as a canonical objective [16]; PAG applies this to T2I prompt tokens [36].Source: §2 · pp. 4–5
  • DOGS induces a corpus grammar, parses a brief into a partial state, completes it with a GFlowNet, and renders and scores the terminal design using a frozen T2I model.The resulting trajectory probability trains the sampler with the Trajectory Balance loss.Source: Figure 2 · p. 5

3 Methodology

DOGS completes a brief by sampling a structured symbolic logo design rather than rewriting the brief as text. A corpus-derived grammar guides completion, and a reward balances recognizability, aesthetics, and originality.

3.1 Overview

  • A brief becomes a partial design state, which a GFlowNet completes by sampling terminal states in proportion to their reward.The state is converted deterministically into a renderer prompt and passed to a frozen T2I model.Source: §3.1 · p. 5

3.2 Symbolic Design Space

  • The symbolic design space represents logos as typed compositions with entity nodes, object slots, visual attributes, and relations that may remain undecided.Entity and object nodes can also carry role and repetition modifiers.Source: §3.2 · p. 6

3.3 Corpus-Induced Design Grammar

  • A corpus-mined graph records empirical co-occurrence among design elements and constrains legal next actions during completion.Its edges carry co-occurrence counts after a support threshold filters low-support pairs.Source: §3.3 · pp. 6–7

3.4 GFlowNet for Design Completion

  • The sampler is pretrained on masked corpus designs, then finetuned with a GFlowNet objective to draw states in proportion to terminal reward.Pretraining labels actions that restore an annotated terminal state as positive and other actions as negative; a KL anchor and entropy bonus stabilize finetuning.Source: §3.4 · pp. 8–9

3.5 Reward Computation

  • The terminal-state reward combines CLIP-based recognizability, judged aesthetics, and corpus-relative originality measured visually and symbolically.The originality views compare rendered images with corpus logos and assess how commonplace design-element pairings are.Source: §3.5 · pp. 9–10

4 Experiments

The experiments compare DOGS with nine prompt-engineering baselines on two open-source renderers, using separate logo-quality and trademark-infringement evaluations. DOGS leads jointly on aesthetics and diversity, and shows low trademark leakage across prompt types.

4.1 Experimental Setup

  • Evaluation compares prompt-side procedures on two open-source renderers, using 150 logo briefs and a disjoint 100-brand set across three leakage prompt types.Training uses 58,243 of 61,263 re-annotated logos; the co-occurrence graph has 3,312 nodes and 126,686 edges.Source: §4 · p. 10, §4.1 · pp. 10–12

4.2 Metrics

  • Quality scores recognizability, aesthetics, and within-cell diversity; infringement is measured as leakage lift and reported separately by prompt type.Leakage lift subtracts similarity to a randomly chosen other brand from similarity to the target brand.Source: §4.2 · p. 12

4.3 Results

  • DOGS is the only method leading jointly on aesthetics and diversity, and its recognizability reaches 9.79 on FLUX, the top FLUX score.It matches or exceeds the LLM rewriters on aesthetics on both renderers.Source: §4.3 · pp. 12–13, Table 1 · p. 13
  • Under the brand prompt, DOGS leakage lift is 0.02 on both renderers versus baseline values of 0.13 / 0.20 for Original; across three regimes it stays ≈0.02, except SDXL/design (0.03 vs 0.01).Leakage lift is defined as direct minus baseline.Source: §4.3 · p. 13, Table 2 · p. 14

4.4 Ablation study

  • Reward-conditioned finetuning yields the highest diversity (3.71/3.41) and best recognizability on both renderers, while largely preserving aesthetics.Graph retrieval is deterministic and scores zero diversity; supervised completion recovers diversity before reward-conditioned finetuning.Source: Table 1 · p. 13, §4.4 · p. 14

5 Conclusion

DOGS replaces language-space rewriting with sampling over a corpus-induced graph grammar of typed design elements. Across two renderers and nine baselines, it produces more recognizable, aesthetic, and diverse logos with less trademark infringement.

5 Conclusion

  • DOGS recasts logo prompt optimization as sampling design graphs with a GFlowNet reward over recognizability, aesthetics, and corpus-relative originality.The graph grammar is induced from real-world logos and its edges encode empirical co-occurrence.Source: §5 · p. 14
  • Across two renderers and nine baselines, DOGS produces logos that are more recognizable, aesthetic, and diverse, and far less prone to trademark infringement.Source: §5 · p. 14
  • The conclusion suggests that design-space prompting may extend beyond logos to structured visual design.Source: §5 · p. 14

Related papers

Loading 2610.10760v1…