Research map / Image generation / Text-to-image diffusion

Seeking the Unfamiliar but Memorable: Conceptual Creativity as Meta-Learning

NeurIPS 2026 · Image generation › Text-to-image diffusion · approach: Inference-time T2I improvement · 0 citations (OpenAlex)

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 a Creator-Appraiser bilevel meta-learning framework that rewards generators for producing stimuli that are initially unfamiliar but quickly learnable by an adaptive observer, instantiated with a frozen diffusion model on MNIST (autoencoder appraiser) and Stable Diffusion v1.5 with a CLIP-LoRA appraiser on natural images.

Problem

Defining and computationally producing creative concepts: how to generate new concepts that are unfamiliar to an observer yet rapidly learnable, rather than retrieving familiar ones or producing incoherent noise.

Approach

A Creator generates a candidate; a fast-adapting Appraiser fits it in a few inner-loop steps from a reset initialization; the Appraiser's learning progress is the reward whose meta-gradient updates the Creator's latent or text embedding at each denoising step, with gating on initial and residual loss for the autoencoder variant.

Key results

• On natural images, the CLIP-LoRA appraiser's inner-loop reward rises from 0.25 at rank 4 to 0.70 at rank 32, while the VLM relevance acceptance rate falls from about 65% to 12%.
• Median nearest-neighbor DINOv2 distance of r=8 Creator-Appraiser outputs from 1000 vanilla SD samples is about 5x the intra-vanilla distance (0.58 vs 0.11).
• On MNIST, only the proposed method reaches the unfamiliar-and-learnable region of the (L0, LT) plane; repulsive guidance stays on the no-learning diagonal and noise perturbation stays near the origin.

Surprising findings

• Relative improvement (L0 - LT)/L0 inverts the desired ranking, rewarding familiar inputs over structured-novel ones.
• Fine-tuning only the first (high-level) decoder layer recovers most of the full-decoder reward, while the last layer yields near-zero improvement.

Limitations

• Creativity is measured only as perceptual value, not functional value, and the criterion is plausibly necessary but not sufficient.
• The Appraiser is memoryless across creations.
• Higher LoRA rank increases off-class drift and lowers VLM relevance acceptance.

Experimental setup

Datasets: MNIST Omniglot ImageNet

Benchmarks: MNIST Omniglot

Metrics: initial loss L0 residual loss LT shaped reward CLIP similarity sim0 CLIP similarity improvement delta sim VLM relevance pass rate DINOv2 nearest-neighbor distance

Baselines: Vanilla DDPM noise perturbation repulsive guidance Vanilla Stable Diffusion

Builds on: MAML CLIP DDPM Stable Diffusion LoRA compression progress

Closest papers by meaning

• An analytic theory of creativity in convolutional diffusion models (ICML 2025 · similarity 0.89)
• Enhancing Creative Generation on Stable Diffusion-based Models (CVPR 2025 · similarity 0.88)
• The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.87)
• Redefining <Creative> in Dictionary: Towards an Enhanced Semantic Understanding of Creative Generation (CVPR 2025 · similarity 0.86)
• Concept Reachability in Diffusion Models: Beyond Dataset Constraints (ICML 2025 · similarity 0.86)
• Emergence of Hidden Capabilities: Exploring Learning Dynamics in Concept Space (NeurIPS 2024 · similarity 0.85)