A Creative Agent is Worth a 64-Token Template

Ruixiao Shi; Fu Feng; Yucheng Xie; Xu Yang; Jing Wang; Xin Geng

cs.CVMar 18, 2026

A Creative Agent is Worth a 64-Token Template

Authors: Ruixiao Shi, Fu Feng, Yucheng Xie, Xu Yang, Jing Wang, Xin Geng

Organizations: School of Computer Science and Engineering, Southeast University, Nanjing, China · Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China

Abstract

Text-to-image (T2I) models have substantially improved image fidelity and prompt adherence, yet their creativity remains constrained by reliance on discrete natural language prompts. When presented with fuzzy prompts such as a creative vinyl record-inspired skyscraper'', these models often fail to infer the underlying creative intent, leaving creative ideation and prompt design largely to human users. Recent reasoning- or agent-driven approaches iteratively augment prompts but incur high computational and monetary costs, as their instance-specific generation makes creativity'' costly and non-reusable, requiring repeated queries or reasoning for subsequent generations. To address this, we introduce CAT, a framework for Creative Agent Tokenization that encapsulates agents' intrinsic understanding of ``creativity'' through a Creative Tokenizer. Given the embeddings of fuzzy prompts, the tokenizer generates a reusable token template that can be directly concatenated with them to inject creative semantics into T2I models without repeated reasoning or prompt augmentation. To enable this, the tokenizer is trained via creative semantic disentanglement, leveraging relations among partially overlapping concept pairs to capture the agent's latent creative representations. Extensive experiments on Architecture Design, Furniture Design, and Nature Mixture tasks demonstrate that CAT provides a scalable and effective paradigm for enhancing creativity in T2I generation, achieving a $3.7\times$ speedup and a $4.8\times$ reduction in computational cost, while producing images with superior human preference and text-image alignment compared to state-of-the-art T2I models and creative generation methods.

Figures & tables

Figure 1: CAT: C reative A gent T okenization for Efficient Creative Generation. CAT generates the token template that directly injects creative intent into fuzzy prompts, enabling efficient, high-quality combinatorial creativity across Architecture Design , Furniture Design , and Nature Mixture tasks.

Figure 2: Breaking the Quality-Efficiency Bottleneck in Creative Generation. Compared with (a) direct generation, (b) think-then-generation and (c) agent-based approaches improve conceptual fusion but incur high time and API costs (see App. A ). In contrast, (d) CAT encapsulates the agent’s intrinsic understanding of “creativity” in a reusable token template that directly enhances semantic representations, delivering superior visual quality with minimal cost and inference time.

Figure 3: Overview of CAT. (a) We introduce a Creative Augmentor to augment fuzzy prompts and a Creative Evaluator to filter generated concepts, with valid ones stored in a Concept Pool . (b) The Creative Tokenizer maps each fuzzy prompt embedding to a corresponding token template, which is directly concatenated with the fuzzy embedding to form a creative embedding for creative generation. (c) The Creative Tokenizer is trained via semantic disentanglement, leveraging relations among partially overlapping concept pairs to capture the agent’s implicit understanding of creativity.

Figure 4: Performance of CAT on the Architecture Design and Furniture Design Tasks . We compare CAT with the representative open-source T2I model FLUX.1 Labs (2024) , proprietary models GPT-Image-1.5 OpenAI (2025) and Gemini 3.1 Flash Image Google (2026) (enhanced with Gemini-3.1-Pro for complex reasoning), as well as agent-driven creative generation methods T2I-Copilot Chen et al. (2025) and CREA Venkatesh et al. (2025) .

Figure 5: Performance of CAT on the Nature Mixture Task. We compare CAT with representative T2I models, including FLUX.1 Labs (2024) and Stable Diffusion 3 (SD3) Esser et al. (2024) , as well as state-of-the-art creative generation methods such as BASS Li et al. (2024) , AGSwap Zhang et al. (2025) , and CreTok Feng et al. (2025b) .

		Architecture Design			Furniture Design			Nature Mixture
		T2I-Copilot	CREA	CAT	T2I-Copilot	CREA	CAT	BASS	AGSwap	CreTok	CAT
(a)	VQAScore $\uparrow$	0.80	0.80	0.85	0.87	0.88	0.91	0.76	0.76	0.85	0.89
	PickScore $\uparrow$	23.51	23.64	24.11	23.29	22.90	23.51	21.64	21.90	22.25	24.03
	ImageReward $\uparrow$	1.19	0.90	1.46	1.14	0.77	1.26	0.70	0.50	1.10	1.77
(b)	VLM-as-a-judge $\uparrow$	7.30	7.82	8.44	7.88	7.64	8.20	6.81	7.50	7.59	8.82
(c)	User Study (Aver. Rank) $\downarrow$	2.33	2.19	1.48	2.28	2.27	1.44	3.16	2.98	2.52	1.34

Table 1: Quantitative Comparison of Creative Generation. Performance is measured using automated image-text and human-aligned metrics, with creativity further evaluated by GPT-4o, and complemented by a user study reporting average human preference ranks.

Figure 6: Ablation Study on the Contribution of Each Loss Component. The baseline uses direct generation with FLUX; Variant #1 uses only $\mathcal{L}_{\text{align}}$ ; Variant #2 adds $\mathcal{L}_{\text{disen}}$ ; CAT corresponds to the full model with $\mathcal{L}_{\text{total}}$ .

Figure 7: Impact of Token Number on Token Template Performance. We evaluate the performance of the token template with 16, 32, 64 (Ours), and 128 tokens.

Figure 8: Token Template vs. Creative Augmentor. Continuous token template encodes richer creative intent, yielding more coherent and visually harmonious outputs than the Creative Augmentor using discrete language prompts.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

	Method	Integration	Alignment	Originality	Aesthetic	Comprehensive
Arch.	T2I-Copilot	6.80 $\pm$ 1.95	7.25 $\pm$ 1.94	6.97 $\pm$ 1.69	8.43 $\pm$ 0.67	7.30 $\pm$ 1.60
	CREA	7.60 $\pm$ 1.89	7.60 $\pm$ 1.70	8.03 $\pm$ 1.42	8.65 $\pm$ 0.75	7.82 $\pm$ 1.39
	CAT	8.45 $\pm$ 1.07	8.60 $\pm$ 0.82	7.83 $\pm$ 0.99	8.85 $\pm$ 0.49	8.44 $\pm$ 0.80
Furni.	T2I-Copilot	7.92 $\pm$ 1.64	8.10 $\pm$ 1.52	7.55 $\pm$ 1.36	8.50 $\pm$ 1.05	7.88 $\pm$ 1.23
	CREA	7.78 $\pm$ 1.44	7.20 $\pm$ 1.70	7.95 $\pm$ 1.65	8.15 $\pm$ 0.67	7.64 $\pm$ 1.17
	CAT	8.22 $\pm$ 1.71	8.18 $\pm$ 1.16	8.07 $\pm$ 1.55	8.50 $\pm$ 0.89	8.20 $\pm$ 1.21

Appendix

Table 2: Detailed Results of VLM-as-a-judge. We reported detailed sub-scores for Integration, Alignment, Originality, and Aesthetic.

Task	Concept	Style
Architecture Design	(10 items): Clock Tower, Museum, Observation Tower, Observatory, Pavilion, Shopping Mall, Skyscraper, Stadium, Suspension Bridge, Windmill	(55 items): Accordion, Armadillo Shell, Bamboo, Bat Wing, Bird Nest, Bookshelf, Bubble, Butterfly, Cactus, Camera Lens, Candle, Caterpillar, Circuit Board, Coral Reef, Corn, Crystal, Dandelion, Deer Antler, DNA Double Helix, Dolphin, Fish, Folding Fan, Gyroscope, Honeycomb, Hourglass, Jellyfish, Lantern, Leaves, Lotus, Mountain, Mushroom, Onion, Origami Crane, Peacock Feather, Peanut, Piano, Pinecone, Pipe Organ, Pocket Watch, Prism, Ribcage, Rose, Sailing Ship, Sea Urchin, Seashell, Shark, Snail Shell, Spider Web, Stained Glass, Sunflower, Tortoise Shell, Tree Branches, Vinyl Record, Walnut, Woven Basket
Furniture Design	(20 items): Bar Stool, Bed, Bookshelf, Cabinet, Chair, Chandelier, Coat Rack, Coffee Table, Cradle, Dining Table, Dresser, Fireplace, Floor Lamp, Mirror, Sofa, Swing, Table Lamp, Television, Tub, Vase	(38 items): Accordion, Armadillo Shell, Bamboo, Bat Wing, Bird Nest, Bubble, Butterfly, Cactus, Circuit Board, Coral Reef, Crystal, Dandelion, Deer Antler, Folding Fan, Honeycomb, Hourglass, Jellyfish, Lantern, Leaves, Lotus, Mushroom, Origami Crane, Peacock Feather, Pinecone, Prism, Ribcage, Rose, Sea Urchin, Seashell, Snail Shell, Spider Web, Stained Glass, Sunflower, Tortoise Shell, Tree Branches, Vinyl Record, Walnut, Woven Basket
Task	Concept
Nature Mixture	(60 items): Ant, Antelope, Bat, Bear, Bee, Broccoli, Butterfly, Cabbage, Cactus, Camel, Carrot, Cat, Cauliflower, Chameleon, Cherry, Chicken, Coconut, Crab, Crocodile, Deer, Dog, Dolphin, Eagle, Elephant, Fox, Frog, Giraffe, Gorilla, Grape, Hippo, Horse, Jellyfish, Kangaroo, Lemon, Lettuce, Lion, Mango, Mushroom, Octopus, Owl, Peacock, Penguin, Persimmon, Pineapple, Potato, Pumpkin, Rabbit, Rhino, Snail, Snake, Spider, Starfruit, Strawberry, Sunflower, Tiger, Tomato, Turtle, Watermelon, Whale, Zebra

Appendix

Table 3: Overview of the Expanded CangJie Dataset. For the Architecture Design and Furniture Design tasks, each pair consists of a primary concept and a style-guiding concept. For the Nature Mixture task, the two concepts in each pair are combined in equal proportion. All items are listed in alphabetical order.

Figure 9: Interface of the User Study. Each volunteer is assigned a subset of questions from three tasks: Architecture Design (20 questions), Furniture Design (20 questions), and Nature Mixture (30 questions). For each participant, 10 questions are randomly selected from each task. For each question, participants rank the presented images based on perceived creativity.

Figure 10: Qualitative results of CAT on the Architecture Design task.

Figure 11: Qualitative results of CAT on the Furniture Design task.

Figure 12: Qualitative results of CAT on the Nature Mixture task.

T2I Generation

Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.

See all 299 papers

May 19, 2026

Self-Creative Text-to-Object Generation using Semantic-Aware Spatial Weighting

Instilling creativity in text-to-image (T2I) generation presents a significant challenge, as it requires synthesized images to exhibit not only visual novelty and surprise, but also artistic value. Current T2I models, however, are largely optimized for literal text-image alignment with their data distribution, and their noise prediction networks constrain the generation to high-probability regions, consequently generating outputs that lack authentic creativity. To address this, we propose a Self-Creative Diffusion (SCDiff) model for meaningful T2I generations featuring two core modules: a learnable spatial weighting (LSW) module and a visual-semantic mixing loss (VSML). The LSW module designs a parametric Kaiser-Bessel window to reinforce central image features, fostering novel and surprising generation. The VSML module introduces a dual loss function: a similarity loss constrains that the new images align with its textual description, while a diversity loss maximizes its distinction from the original image, enhancing both semantic value and visual novelty. Extensive experiments demonstrate that our model substantially improves creativity, semantic alignment, and visual coherence, offering a simple yet powerful framework for generating creative objects.

Yue Yu, Haibo Chen, Shuo Chen +2

Nanjing University of Science and Technology · Nanjing University

T2I Generation Computational Creativity T2I Diffusion Models

Ask PDF

Aug 5, 2026

ToolArtist: Tool-Using Unified Multimodal Models for Agentic Image Generation

Text-to-image (T2I) models can produce visually compelling images, yet they remain limited on open-world tasks that require complex semantic understanding, multi-step reasoning, and the integration of external world knowledge. Existing efforts introduce agent capabilities into image generation, but they either prescribe a fixed workflow or place only a subset of the open-world image generation process under agent control. Consequently, reasoning, tool invocation, and image generation are not coordinated by a single policy. We propose ToolArtist, a fully agentic image generation model obtained by post-training a Unified Multimodal Model (UMM). ToolArtist dynamically orchestrates reasoning, external tool use, and native image generation within one unified policy. During Supervised Fine-Tuning (SFT), we equip a teacher agent with search tools alongside an image-generation tool. We then convert the collected trajectories into a UMM compatible format, where the image-generation tool is concealed while the resulting generated images are retained. During Reinforcement Learning (RL), we develop an agentic RL infrastructure for UMMs and introduce Reason-Act-Draw GRPO (RAD-GRPO), which uses complementary intent and quality rewards to jointly optimize the model. Experiments show that placing the entire open-world image-generation process under an agent policy consistently outperforms approaches with fixed pipelines or only partially agent-controlled components. We release the training data and the complete post-training infrastructure.

Jiahao Zhao, Xiaomin Yu, Zhongxiang Sun +5

1RUC · 2HKUST(GZ) · 3NUS +1

Unified Multimodal Models Agentic Image Generation Tool-Using Vision-Language Agents RL for Image Generation

Ask PDF

May 21, 2026

AtelierEval: Agentic Evaluation of Humans & LLMs as Text-to-Image Prompters

Text-to-image (T2I) systems increasingly rely on upstream prompters, either humans or multimodal large language models (MLLMs), to translate user intent into detailed prompts. Yet current benchmarks fix the prompt and only evaluate T2I models, leaving the prompting proficiency of this upstream component entirely unmeasured. We introduce AtelierEval, the first unified benchmark that quantifies prompting proficiency across 360 expert-crafted tasks. Grounded in a cognitive view, it spans three task categories and instantiates tasks using a taxonomy of real-world challenges, with a dual interface for both humans and MLLMs. To enable scalable and reliable evaluation, we propose AtelierJudge, a skill-based, memory-augmented agentic evaluator. It produces subjective and objective scores for prompt-image pairs, achieving a Spearman correlation of 0.79 with human experts, approaching human performance. Extensive experiments benchmark 8 MLLMs against 48 human users across 4 T2I backends, validate AtelierEval as a robust diagnostic tool, and reveal the superiority of mimicry over planning, advocating for an image-augmented direction for future prompters. Our work is released to support future research.

Hanjun Luo, Zhimu Huang, Sylvia Chung +6

New York University Abu Dhabi · Zhejiang University · University of Electronic Science and Technology of China +2

T2I Generation Agent Evaluation Automated Evaluation T2I Generation Evaluation

Ask PDF

Abstract

T2I Generation

Explore similar work

Self-Creative Text-to-Object Generation using Semantic-Aware Spatial Weighting

ToolArtist: Tool-Using Unified Multimodal Models for Agentic Image Generation

AtelierEval: Agentic Evaluation of Humans & LLMs as Text-to-Image Prompters