Research map / Image generation / Fast diffusion sampling
TFG: Unified Training-Free Guidance for Diffusion Models
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 TFG, a unified algorithmic framework and hyper-parameter search strategy for training-free guidance of diffusion models, benchmarked across 7 diffusion models, 16 tasks and 40 targets.
Problem
Training-free guidance steers an unconditional diffusion model with an off-the-shelf predictor trained only on clean data; existing methods lack theoretical grounding, broad benchmarks, and a principled way to choose hyper-parameters, and can fail even on simple CIFAR10 label guidance.
Approach
A single algorithm-agnostic design space (recurrence, mean guidance, variance guidance, Gaussian-smoothed implicit dynamics, and scalar strengths) that contains DPS, LGD, FreeDoM, MPGD and UGD as special cases; hyper-parameters are chosen by analyzing structure invariance and a beam-search over scalar strengths using small-sample trials.
Key results
- TFG improves guidance validity over the best existing method by 7.4% on average across 14 settings and wins on 13 of 14 (paper abstract reports 8.5% average improvement).
- On CIFAR10 label guidance, increasing recurrence from 1 to 4 raises accuracy from 52.0% to 77.1% (FID 91.7 to 73.9), narrowing the gap to training-based guidance (85%).
- On fine-grained bird guidance (525 species), TFG reaches 2.24% accuracy versus 0% for unconditional generation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TFG instead of the strongest existing method | training-free guidance · Across benchmark tasks | guidance validity: 7.4% average improvement | Helped |
| Increase recurrence from 1 to 2 (ablation) | TFG · CIFAR10 label guidance | accuracy: 52.0 -> 66.8 | Helped |
| Increase recurrence from 1 to 4 (ablation) | TFG · CIFAR10 label guidance | accuracy: 52.0 -> 77.1 | Helped |
| Increase recurrence from 1 to 2 (ablation) | TFG · ImageNet label guidance | accuracy: 40.9 -> 52.3 | Helped |
| Increase recurrence from 1 to 4 (ablation) | TFG · ImageNet label guidance | accuracy: 40.9 -> 59.8 | Helped |
| Vary Monte-Carlo sample count from 1 to 4 (ablation) | TFG variance-only guidance · CIFAR10 label guidance | FID: 90.6; 91.0; 90.7 | No clear effect |
Surprising findings
- Generation quality improves and then deteriorates as recurrence count increases, and FID worsens when N_recur or N_iter is too large.
- Number of Monte-Carlo samples in the Gaussian-smoothed expectation has little effect; adding noise matters more than estimation precision.
Limitations
- Gap to training-based guidance remains large on some tasks such as molecule property guidance.
- Molecules with better validity tend to have lower generation fidelity; the framework does not cover all existing approaches.
- Only the 'increase', 'decrease', and 'constant' schedule structures are studied.
Experimental setup
Datasets: CIFAR10 ImageNet CelebA-HQ Cat QM9
Benchmarks: CIFAR10 label guidance ImageNet label guidance CelebA combined guidance Cat deblur and super-resolution QM9 molecule property guidance audio declipping and inpainting Stable Diffusion style transfer
Metrics: accuracy FID LPIPS KID MAE valid ratio DTW FAD CLIP score
Baselines: DPS LGD FreeDoM MPGD UGD training-based classifier guidance
Builds on: DPS LGD FreeDoM MPGD Universal Guidance
Closest papers by meaning
- No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.94)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.92)
- Token Perturbation Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.91)
- Gradient Guidance for Diffusion Models: An Optimization Perspective (NeurIPS 2024 · similarity 0.91)
- Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.91)
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.