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

SPARKE: Scalable Prompt-Aware Diversity and Novelty Guidance in Diffusion Models via RKE Score

NeurIPS 2025 · Image generation › Text-to-image diffusion · approach: Inference-time T2I improvement · 2 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 SPARKE, a prompt-aware diversity guidance for latent diffusion models that uses a conditional order-2 Rényi kernel entropy (RKE) score, reducing guidance cost to O(n) to enable large-scale sampling.

Problem

Diversity of prompt-guided diffusion samples is hard to control across semantically varied prompts; existing entropy/Vendi-based guidance is prompt-agnostic or O(n^3) and does not scale to large batches.

Approach

Applies the conditional RKE score (Hadamard product of output and prompt kernel matrices) as a guidance potential in latent space of LDMs, using the order-2 Rényi entropy's Frobenius-norm form and a gradient simplification so that per-step guidance costs O(n) without prompt pre-clustering.

Key results

• On SD 2.1, SPARKE latent Cond-RKE raises Cond-Vendi score from 26.54 (no guidance) to 32.57 and Vendi score from 369.37 to 405.51 while CLIPScore goes 31.20 to 30.96.
• On SDXL, SPARKE improves Cond-Vendi from 27.54 to 31.17 and Vendi from 309.54 to 313.80 with CLIPScore 31.17 to 30.47.
• On SD-1.5, SPARKE adds about 0.13 s per sample runtime over the base model (1.752 vs 1.620 s) and about 0.05 GB peak memory, versus 3.079 s and 8.665 GB for c-VSG.

Measured effects reported in the paper

Surprising findings

• Unconditional RKE guidance becomes less effective over the course of generation, while the prompt-aware variant stays effective across 40k iterations.
• In the SD 2.1 table, c-VSG lowers CLIPScore (28.75) relative to no guidance (31.20).

Limitations

• Evaluation limited to image generation diffusion models; no video or text diffusion experiments.
• Combining with other diversity-enhancement techniques is left for future work.
• Reported fidelity (CLIPScore) drops slightly for SPARKE on SD 2.1, SDXL and PixArt-Sigma.

Experimental setup

Datasets: MS-COCO prompts GeoDE 2D Gaussian mixture benchmarks

Benchmarks: MS-COCO GeoDE 2D Gaussian mixture

Metrics: CLIPScore KID Vendi score Conditional-Vendi score in-batch similarity Density Coverage runtime GPU memory

Baselines: CADS Particle Guidance c-VSG VSG SD (No Guidance) latent RKE guidance

Builds on: Vendi score guidance contextualized Vendi score guidance (c-VSG) RKE score classifier-free guidance CADS

Closest papers by meaning

• Image Generation Diversity Issues and How to Tame Them (CVPR 2025 · similarity 0.88)
• Don't Play Favorites: Minority Guidance for Diffusion Models (ICLR 2024 · similarity 0.88)
• MGD$^3$ : Mode-Guided Dataset Distillation using Diffusion Models (ICML 2025 · similarity 0.88)
• Training Diffusion Models Towards Diverse Image Generation with Reinforcement Learning (CVPR 2024 · similarity 0.88)
• FlashEval: Towards Fast and Accurate Evaluation of Text-to-image Diffusion Generative Models (CVPR 2024 · similarity 0.87)
• Minority-Focused Text-to-Image Generation via Prompt Optimization (CVPR 2025 · similarity 0.87)