cs.CVOct 7, 2026

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

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

Organizations: Department of Computer Science Virginia Tech Alexandria, Virginia, USA · Verisign, Inc. Reston, Virginia, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Figure 1: From user brief to logo design. Each column is one brief. Top: the user prompt rendered directly. Bottom: the DOGS output, which fills in the unspecified design decisions to produce a coherent, professionally composed mark.
Figure 2: DOGS pipeline , organized into three stages. Brief Parsing maps the user brief into an initial state s0s_{0} , filling the design slots it specifies and leaving the rest undecided. Offline Grammar Induction re-annotates a logo corpus into typed nodes and builds a co-occurrence grammar G=(V,E)G=(\mathcal{V},\mathcal{E}) , pruned to define the per-step action space. GFlowNet Symbolic Completion draws a trajectory s0→⋯→sTs_{0}\to\cdots\to s_{T} with the sampler πF\pi_{F} over that action space; the terminal state sTs_{T} is rendered into a logo by a frozen T2I model, scored by the reward model, and the resulting trajectory probability trains πF\pi_{F} via the Trajectory Balance loss LTB\mathcal{L}_{\text{TB}} .
Figure 3: Qualitative comparison. Columns are methods. Comparison : DOGS composes a coherent design while baselines drift off-subject or render a stock image. Diversity : four samples per method as 2×22{\times}2 mosaics; DOGS yields four distinct yet usable logos, while baselines collapse or vary without comparable quality. Infringement : a brand prompt names a known trademark; baselines reproduce it, DOGS substitutes a different professional design (prompts truncated).
 SDXL-LightningFLUX.1-schnell
MethodQrec ⁣↑Q_{\text{rec}}\!\uparrowQaes ⁣↑Q_{\text{aes}}\!\uparrowDiversity ↑\uparrowQrec ⁣↑Q_{\text{rec}}\!\uparrowQaes ⁣↑Q_{\text{aes}}\!\uparrowDiversity ↑\uparrow
Original prompt9.26 ±\pm 0.154.03 ±\pm 0.170.00 ±\pm 0.009.70 ±\pm 0.104.96 ±\pm 0.170.00 ±\pm 0.00
Qwen rewrite7.41 ±\pm 0.213.19 ±\pm 0.152.25 ±\pm 0.167.75 ±\pm 0.203.64 ±\pm 0.182.49 ±\pm 0.16
Llama rewrite9.40 ±\pm 0.114.07 ±\pm 0.122.76 ±\pm 0.189.71 ±\pm 0.064.67 ±\pm 0.112.30 ±\pm 0.15
Promptist9.60 ±\pm 0.093.35 ±\pm 0.152.52 ±\pm 0.209.50 ±\pm 0.094.15 ±\pm 0.152.33 ±\pm 0.19
BeautifulPrompt9.61 ±\pm 0.073.18 ±\pm 0.113.34 ±\pm 0.169.45 ±\pm 0.083.65 ±\pm 0.113.54 ±\pm 0.16
Table 1: Quality battery. 150150 briefs, five samples per cell (metrics: Sec. 4.2 ). Bold: best among baselines and DOGS; abl. rows: ablations of Sec. 4.4 . Deterministic methods score zero diversity by construction.
 SDXL-LightningFLUX.1-schnell
MethodBrand ↓\downarrowCompany ↓\downarrowDesign ↓\downarrowBrand ↓\downarrowCompany ↓\downarrowDesign ↓\downarrow
Original prompt0.13 ±\pm 0.020.04 ±\pm 0.010.05 ±\pm 0.010.20 ±\pm 0.020.08 ±\pm 0.010.06 ±\pm 0.02
Qwen rewrite0.07 ±\pm 0.020.02 ±\pm 0.010.02 ±\pm 0.010.11 ±\pm 0.020.04 ±\pm 0.020.03 ±\pm 0.02
Llama rewrite0.12 ±\pm 0.010.04 ±\pm 0.010.05 ±\pm 0.010.18 ±\pm 0.020.07 ±\pm 0.010.07 ±\pm 0.02
Promptist0.10 ±\pm 0.020.04 ±\pm 0.010.04 ±\pm 0.010.19 ±\pm 0.020.07 ±\pm 0.020.07 ±\pm 0.02
BeautifulPrompt0.10 ±\pm 0.020.04 ±\pm 0.010.02 ±\pm 0.010.15 ±\pm 0.020.07 ±\pm 0.010.03 ±\pm 0.01
Table 2: Infringement battery. Leakage lift Δ=direct−baseline\Delta=\text{direct}-\text{baseline} , 100100 brands, one image per cell. Columns are the three prompt types of Sec. 4.1 .

Explore similar work

CardsList

  1. TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design

    May 20, 2026Haonan Zhu, Elad Hirsch, Alexandria Minetti +3VLM EvaluationT2I Generation Evaluation

    Lica World, San Francisco, USA. · Contra.Work Inc., Brooklyn, NY, USA.

  2. Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation

    Mar 12, 2026Joong Ho Kim, Nicholas Thai, Souhardya Saha Dip +2T2I GenerationT2I Generation Evaluation

    LSU ATHENA Lab, Baton Rouge, LA, U.S.A. · LSU Vision Lab, Baton Rouge, LA, U.S.A.

  3. Decomposing Subject-Driven Image Generation via Intermediate Structural Prediction

    May 20, 2026Hanzhong Guo, Yizhou YuSubject-Driven Image GenerationConditional Image Generation

    School of Computing and Data Science, The University of Hong Kong