Get Started
Home
Topics
Search
Library
Research questionHow can prompt embeddings be optimized during diffusion inference to balance aesthetics, prompt alignment, and resource use?Prompt-embedding search can change generated images without updating the diffusion model, but optimization choices affect both the resulting image and the resources required. The challenge is choosing a search strategy and objective weighting that meet the desired aesthetic and alignment goals efficiently.
Diffusion Models
Image Generation
Inference Optimization
Neural and Evolutionary Computing
Latest papersRecent research connected to this question, newest first.Inference-Time Optimization of Prompt Embeddings in Diffusion Models: A Comparison of sep-CMA-ES and AdamThe evidence concerns Stable Diffusion XL Turbo, 36 Parti Prompts, and comparisons between sep-CMA-ES and Adam. Candidate images are scored with weighted LAION Aesthetic Predictor V2 and CLIPScore objectives under aesthetics-only, balanced, and alignment-only settings; cosine similarity, SSIM, compute, and memory are also reported.research paper · Sep 2, 2026
Related questions
How can text-to-image diffusion models preserve prompt alignment without the extra sampling cost of classifier-free guidance?How can text-to-image models preserve variation in unspecified visual factors under long, semantically dense prompts?How can evolutionary prompt optimization improve accuracy without increasing prompt length?How can diffusion image and video generators be preference-aligned without inefficient training exploration or inference-time search?