Research map / Image generation / Fast diffusion sampling
Decompose the Distillation: Interpretable Single-Pass Guidance for Diffusion Models
Wuyang Zhang, Xiaolin Xu
Read the paperCheck your idea against 137,217 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 an interpretable single-pass guidance method for diffusion models that decomposes distillation of classifier-free guidance, avoiding the 2N forward passes of CFG.
Problem
Classifier-free guidance requires conditional and unconditional evaluations at every sampling step (2N forward passes for N steps); existing single-pass alternatives reduce this cost but the abstract (truncated) suggests they lack interpretability.
Approach
Decomposes the distillation of CFG into interpretable components to obtain guidance in a single forward pass while retaining an adjustable guidance scale; the full mechanism is not stated in the truncated abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add scale-adaptive modulation to the core student (ablation) | CtrlFlow-Core · ImageNet (N=4, s=8) | FID-50K: 73.41 -> 63.42 | Improved |
| Add time-grid warping to the core student (ablation) | CtrlFlow-Core · ImageNet (N=4, s=8) | FID-50K: 73.41 -> 59.83 | Improved |
| Add both modulation and warping (full model) (ablation) | CtrlFlow-Core · ImageNet (N=4, s=8) | FID-50K: 73.41 -> 48.95 | Improved |
| Use unified composed-output distillation instead of component-wise loss (ablation) | CtrlFlow · ImageNet (N=4, s=8) | FID-50K: FID degrades by 19.5 when removed | Improved |
| Dual-adapter decomposition alone vs opaque scale-conditioned student (ablation) | CtrlFlow-Core vs AGD-MultiScale · ImageNet (N=4, s=8) | FID-50K: 52.87 -> 73.41 | Hurt |
| Give windowing and dampening to the opaque student as time-varying scale input (ablation) | AGD-MultiScale · ImageNet DiT-XL/2 | FID-50K: 0.7 to 2.3 FID better | Improved |
| Add alignment gate that needs the v0/g split (ablation) | CtrlFlow-Full · ImageNet DiT-XL/2 | FID-50K: removing it costs 0.6 to 2.4 FID | Improved |
| Use three-phase curriculum rather than skipping Phase 1 or Phase 2 (ablation) | CtrlFlow · ImageNet DiT-XL/2 | training stability / high-scale FID | Improved |
| Use rank 64 adapters (ablation) | CtrlFlow · ImageNet DiT-XL/2 | FID / cost | No clear effect |
| Use conditional-field anchor v_cond + (s-1)g (ablation) | CtrlFlow-Core · ImageNet DiT-XL/2 | FID-50K: matches Core; trails Full by 1.1 to 2.6 FID | No clear effect |
| Apply windowing and dampening to exact two-pass CFG | CFG · ImageNet (N=4, s=8) | FID-50K: 53.84 -> 46.02 | Improved |
| Distill CFG into single pass with decomposed adapters (vs CFG) | CtrlFlow · ImageNet DiT-XL/2, N=25 | wall-clock latency (ms/img): 843 -> 437 | Improved |
Experimental setup
Baselines: classifier-free guidance
Builds on: classifier-free guidance
Closest papers by meaning
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.94)
- Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.93)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.92)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.92)
- Understanding and Improving Training-free Loss-based Diffusion Guidance (NeurIPS 2024 · similarity 0.91)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 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.
