Research map / Image generation / Fast diffusion sampling
Fast Direct: Query-Efficient Online Black-box Guidance for Diffusion-model Target 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 Fast Direct, a query-efficient online black-box guidance method that steers a pre-trained diffusion model's noise sequence toward a pseudo-target built from a GP surrogate or historical best samples.
Problem
Guiding pre-trained diffusion models toward targets defined by expensive, non-differentiable black-box objectives (e.g., LLM-rated prompt alignment, molecular binding affinity) with few online queries and no offline dataset.
Approach
Guided noise sequence optimization moves each noise in the universal direction x* - x_K, and Fast Direct replaces the unknown target with a pseudo-target set nonparametrically (GP gradient step or historical best), enabling black-box guidance without training or differentiating the objective.
Key results
- On 12 image tasks with a 50-batch-query budget, Fast Direct reaches near-full Gemini score and is 6x to 10x more batch-query-efficient than DDPO, DPOK, and D3PO given 200 queries.
- On 6 molecule tasks, Fast Direct is 11x up to 44x more query-efficient than baselines (average gains of 26.67x vs DDPO and 44.44x vs DPOK, per the table).
- Fast Direct is insensitive to step size and batch size in the ablation on the deer-elephant task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the universal update direction x* - xK instead of x* - xk (ablation) | Guided noise sequence optimization · Target image generation | image quality | Helped |
| Use the predicted-clean-image approximation x* - x̂K instead of x* - xK (ablation) | Guided noise sequence optimization · Target image generation | target accuracy | Hurt |
| Use x* - xK′ with a smaller K′ (K/2, K/4, or K/8) (ablation) | Guided noise sequence optimization · Target image generation | image quality | Hurt |
| Vary the step size α across the tested settings (ablation) | Fast Direct · Image target generation Task-1 | Gemini rating | No clear effect |
| Vary batch size B across the tested settings (ablation) | Fast Direct · Image target generation Task-1 | Gemini rating | No clear effect |
| Use EDM instead of DDIM (ablation) | Fast Direct · Aesthetic quality | aesthetic score | Helped |
Surprising findings
- Guidance with the noisy direction x* - x_K still yields clean on-manifold images even when the target is heavily noised (variance 9 x I).
- The direction x* - x_k (intermediate noisy state) degrades image quality, while x* - x_K works well.
Limitations
- Relies on an LLM judge (Gemini 1.5 Flash) and a docking simulator as objectives, not real human or wet-lab evaluations.
- No theoretical explanation for why the universal direction works on the data manifold; left as future work.
- Baseline comparisons use default hyperparameters and extra budgets only for selected tasks.
Experimental setup
Datasets: CrossDocked2020
Benchmarks: 12 image prompt-alignment tasks (SDXL-Lightning, Gemini rating) 6 TargetDiff molecule binding tasks (Vina score) DDPO compressibility, incompressibility, aesthetic quality
Metrics: Gemini rating (1-5) Vina score (kcal/mol) batch-query-efficiency gain JPEG size aesthetic score
Baselines: DDPO DPOK D3PO DNO
Builds on: DNO DDPO DPOK D3PO DDIM
Closest papers by meaning
- TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Simple Guidance Mechanisms for Discrete Diffusion Models (ICLR 2025 · similarity 0.88)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.88)
- Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.88)
- Greed is Good: A Unifying Perspective on Guided Generation (NeurIPS 2025 · similarity 0.88)
- Gradient Guidance for Diffusion Models: An Optimization Perspective (NeurIPS 2024 · similarity 0.88)
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.