Research map / Image generation / Text-to-image diffusion
Redefining <Creative> in Dictionary: Towards an Enhanced Semantic Understanding of Creative Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CreTok, which learns a single new token <CreTok> for "creative" in Stable Diffusion 3 by aligning text embeddings of restrictive and adaptive prompts over a new CangJie text-pair dataset, enabling training-free combinatorial creative generation.
Problem
Diffusion models generate out-of-distribution concepts easily but struggle with abstract "creative" combinatorial objects (e.g., a creative mixture of a lettuce and a mantis); existing creative methods require per-output retraining or reference images.
Approach
Redefine the abstract word "creative" as a single learnable token whose text embedding is optimized (image-free) so that adaptive prompts ("a photo of a <CreTok> mixture") match restrictive prompts ("a t1 t2") across many sampled text pairs, giving a universal adjective usable without further training.
Key results
- CreTok achieves VQAScore 0.835 vs 0.805 for SD 3.5 and 0.793 for SD 3, with ImageReward 1.065 vs 0.896 for SD 3.
- User study average rank 1.9 for CreTok vs 3.1 for BASS and 3.4 for SD 3.
- Generation takes about 4s per image versus 40s for BASS and 120s for ConceptLab, with about 30 minutes of training on a single 4090 GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace “creative” with the trained universal <CreTok> token | Stable Diffusion 3 · CangJie / BASS text pairs | VQAScore: ↑0.03 | Helped |
| Replace “creative” with the trained universal <CreTok> token | Stable Diffusion 3 · CangJie / BASS text pairs | PickScore: ↑0.009 | Helped |
| Replace “creative” with the trained universal <CreTok> token | Stable Diffusion 3 · CangJie / BASS text pairs | ImageReward: ↑0.169 | Helped |
| Replace “creative” with the trained universal <CreTok> token | Stable Diffusion 3 · CangJie / BASS text pairs | Average rank: 1.9 vs. 3.1 | Helped |
| Use the trained universal <CreTok> token instead of the baseline generation methods | Creative image generation · CangJie / BASS text pairs | generation time per image: 4s vs. 120s per image; 30→ speedup | Helped |
| Use the trained universal <CreTok> token instead of BASS | Creative image generation · CangJie / BASS text pairs | generation time per image: 4s vs. 40s per image; 10→ speedup | Helped |
| Set the similarity-loss threshold to ε = 0.5 (ablation) | CreTok · Combinatorial object generation | concept integration | Helped |
| Use a low similarity-loss threshold (ε = 0.1–0.4) (ablation) | CreTok · Combinatorial object generation | concept integration | Hurt |
| Use a high similarity-loss threshold (ε = 0.6–0.9) (ablation) | CreTok · Combinatorial object generation | concept integration | Hurt |
Surprising findings
- Newer SOTA models such as SD 3.5 and Kandinsky 3 are outperformed by CreTok built on the older SD 3 backbone on alignment and preference metrics.
Limitations
- Improper loss threshold either yields independent concepts or overfitting to one concept.
- Evaluation of creativity relies on proxy metrics and GPT-4o judgments since existing metrics are insufficient.
Experimental setup
Datasets: CangJie BASS text pairs
Benchmarks: CangJie TP2O text pairs
Metrics: VQAScore PickScore ImageReward GPT-4o creativity scores (integration, alignment, originalit user study rank generation time
Baselines: Stable Diffusion 3 Stable Diffusion 3.5 Kandinsky 3 DALL-E 3 Midjourney v6.1 BASS MagicMix Black Scholes ConceptLab
Builds on: ConceptLab BASS Textual Inversion Stable Diffusion 3
Closest papers by meaning
- Enhancing Creative Generation on Stable Diffusion-based Models (CVPR 2025 · similarity 0.88)
- An analytic theory of creativity in convolutional diffusion models (ICML 2025 · similarity 0.87)
- Seeking the Unfamiliar but Memorable: Conceptual Creativity as Meta-Learning (NeurIPS 2026 · similarity 0.86)
- CreBench: Human-Aligned Creativity Evaluation from Idea to Process to Product (AAAI 2026 · similarity 0.86)
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.86)
- DreamDistribution: Learning Prompt Distribution for Diverse In-distribution Generation (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.