Decompose the Distillation: Interpretable Single-Pass Guidance for Diffusion Models

Wuyang Zhang; Xiaolin Xu

cs.CVNov 3, 2026

Decompose the Distillation: Interpretable Single-Pass Guidance for Diffusion Models

Authors: Wuyang Zhang, Xiaolin Xu

Organizations: Department of Electrical and Computer Engineering Northeastern University, Boston, MA, USA

Abstract

Classifier-Free Guidance is a widely used inference-time guidance rule for conditional image generation with diffusion models, which exposes a guidance scale that adjusts conditioning strength after training. This flexibility requires conditional and unconditional model evaluations at every sampling step, so an N-step trajectory takes 2N forward passes. Single-pass alternatives reduce such cost, yet most collapse CFG's two components into one opaque guided prediction, losing the separable base flow and guidance direction that support per-step control and diagnostics, even when the scale remains adjustable. This work introduces CtrlFlow, a single-pass guidance method that preserves CFG's separable base-flow and guidance-direction interface. To preserve both components at one-pass cost, CtrlFlow attaches two lightweight LoRA adapters to a frozen backbone, each adding 2.5% parameters, and predicts the base flow and guidance direction from one pass of the shared frozen backbone. Because the guidance scale multiplies the guidance direction only after the forward pass, one evaluation yields the guided velocity at any scale, so a single trained checkpoint remains adjustable at inference without retraining. Preserving two outputs introduces a training mismatch: inference amplifies guidance-direction errors by the chosen scale, while naive component-wise losses weight both outputs equally. To align training with this composed inference-time error, CtrlFlow uses a unified multi-scale distillation loss. Residual guidance errors can still persist at extreme scales, where even small guidance-direction errors are amplified by the guidance scale. To reduce this inference-time error, CtrlFlow applies scale-adaptive modulation to damp velocity-field errors and time-grid warping to concentrate solver evaluations where the trajectory changes most rapidly.

Figures & tables

Two-Pass (NFE $=$ 2N)
Method	Single Pass	Explicit $\mathbf{v}_{0}/\mathbf{g}$	Post-hoc $s$
CFG	✗	✓ $\dagger$	✓
CFG + Interval	✗	✓ $\dagger$	✓
CFG-Zero*	✗	✗ $\lx@sectionsign$	✓
APG	✗	$\boldsymbol{\sim}$ $\ddagger$	✓
Autoguidance	✗	✗ $\lx@paragraphsign$	✓

Table 1: Comparison of guidance methods. Explicit

\mathbf{v}_{0}/\mathbf{g}

: base flow and guidance direction separately accessible. Post-hoc

s

: guidance scale adjustable without retraining.

Figure 1: CtrlFlow compresses CFG to one pass while preserving separate $\mathbf{v}_{0}$ and $\mathbf{g}$ outputs. Top. A frozen CFG teacher provides the target $\mathbf{v}_{\text{cfg}}^{\text{teacher}}$ . The student matches it via a unified loss on $\hat{\mathbf{v}}_{0}+s\cdot\hat{\mathbf{g}}$ . Bottom. At inference, one backbone pass produces $\mathbf{v}_{0}$ and $\mathbf{g}$ through two LoRA adapters. The scale $s$ multiplies $\mathbf{g}$ post-hoc. Three modulation gates ( $w$ , $\alpha$ , $\varphi$ ) stabilize guidance, and a warped time grid concentrates evaluations near high-curvature regions.

Method	NFE	ms/img -2pt	$s{=}2$ (easy)		$s{=}4$	$s{=}8$ (hard)
			$N{=}25$	$N{=}4$	$N{=}8$	$N{=}25$	$N{=}4$
[0pt][0pt] CFG	$2N$	843	2.97	9.24	11.72	21.47	53.84
[0pt][0pt] CFG-Zero*	$2N$	854	2.81	8.93	11.14	20.67	52.67
[0pt][0pt] CFG + Interval	$2N$	$\sim$ 843	2.84	8.87	11.21	18.23	46.51
[0pt][0pt] APG	$2N$	852	3.21	10.12	12.84	13.56	39.21
[0pt][0pt] Autoguidance	$2N$	891	2.71	8.61	10.83	18.72	48.35

Table 2: FID-50K (

\downarrow

) on ImageNet 256

\times

256 with DiT-XL/2. Shaded rows use two passes, unshaded rows one. AGD-MultiScale appears in Table 3 and Appendix D.2 , and full statistics and grid in Appendices D.6 , D.3 , and D.1 .

Figure 2: FID-50K vs. scale at

N{=}25

. CtrlFlow tracks CFG while the fixed-scale AGD variant (“AGD”) drifts past its training scale

s=2

Variant	$(25,2)$	$(8,4)$	$(4,8)$
[0pt][0pt] AGD (fixed $s{=}2$ )	3.04	25.81	82.17
[0pt][0pt] AGD-MultiScale ( $r{=}128$ )	3.09	11.94	52.87
CtrlFlow-Core	3.14	11.48	73.41
CtrlFlow+Mod	3.12	11.21	63.42
CtrlFlow+Warp	3.15	11.35	59.83
[0pt][0pt] CtrlFlow-Full	3.07	10.98	48.95

Table 3: Contribution isolation, FID-50K (

\downarrow

). Core uses dual adapters and unified distillation, and +Mod/+Warp add stabilization. AGD-MultiScale is a matched-budget, scale-conditioned opaque distillation baseline.

Variant	$(25,2)$	$(25,8)$
CFG teacher	2.97	21.47
Swap $\hat{\mathbf{g}}$	3.07	22.31
Swap $\hat{\mathbf{v}}_{0}$	3.01	21.74
CtrlFlow	3.07	19.28

Table 4: Component swap, FID-50K (

\downarrow

Figure 3: Decomposition maps for golden retriever . Guidance norm becomes spatially focused as structure emerges, while $\cos(\mathbf{v}_{0},\mathbf{g})$ shifts from near-orthogonal to aligned. Extended classes appear in Appendix D.22 .

Appendix figures & tables33 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Description	Value	Used in
Temporal Modulation
$\mu$	Window center	0.5	Eq. 23
$\sigma_{\max}$	Window width at $s{=}1$	0.35	$\sigma(s)$
$\sigma_{\min}$	Window width at $s{=}s_{\max}$	0.10	$\sigma(s)$
Scale Dampening
$s_{\text{thresh}}$	Activation threshold	4.0	Sec. 3.4

Appendix

Table 5: Complete hyperparameter specification for CtrlFlow’s modulation and warping mechanisms.

Method		$N{=}2$	$N{=}4$	$N{=}8$	$N{=}16$	$N{=}25$
CFG	$s{=}1$	42.51	21.83	13.45	10.21	9.62
	$s{=}2$	21.73	9.24	4.82	3.45	2.97
	$s{=}4$	38.54	19.67	11.72	8.63	7.83
	$s{=}6$	64.28	36.18	20.47	16.02	14.21
	$s{=}8$	86.53	53.84	33.18	24.38	21.47
CtrlFlow	$s{=}1$	43.89	22.17	13.82	10.53	9.87

Appendix

Table 6: Full

(N,s)

grid. FID-50K (

\downarrow

) for all 25 configurations on ImageNet 256

\times

256. The boxed cell marks the sole configuration exceeding the CFG-match threshold

\delta

. Values in bold indicate CtrlFlow outperforms CFG at that configuration.

Method	Scale input	Params	$(25,2)$	$(8,4)$	$(25,8)$	$(4,8)$	Grid match
AGD (fixed $s{=}2$ )	none	14M	3.04	25.81	44.92	82.17	5/25
AGD-MultiScale ( $r{=}128$ )	before forward	34M	3.09	11.94	23.42	52.87	20/25
CtrlFlow-Full	after forward	34M	3.07	10.98	19.28	48.95	24/25

Appendix

Table 7: Matched comparison with the scale-conditioned opaque student, FID-50K (

\downarrow

) under the protocol of Table 2 . “Grid match” counts configurations with

\mathrm{FID}\leq\mathrm{FID}_{\mathrm{CFG}}+\delta

over all 25 cells.

Method	$(25,2)$	$(25,8)$	$(8,2)$	$(8,8)$	$(4,2)$	$(4,8)$
CFG	.82 / .57	.87 / .32	.79 / .55	.84 / .28	.73 / .52	.78 / .22
AGD (fixed $s$ )	.81 / .55	.72 / .14	.78 / .53	.65 / .09	.71 / .50	.58 / .06
CtrlFlow	.82 / .56	.86 / .35	.79 / .54	.83 / .30	.73 / .51	.76 / .24

Appendix

Table 8: Precision (

\uparrow

) / Recall (

\uparrow

) at six representative

(N,s)

configurations. Best single-pass result per column in bold .

Method	$(25,2)$	$(25,8)$	$(8,2)$	$(8,8)$	$(4,2)$	$(4,8)$
CFG	271.3	198.4	248.6	167.2	201.5	108.3
AGD (fixed $s$ )	264.8	112.7	241.3	82.4	193.7	47.6
CtrlFlow	269.5	201.2	246.1	170.8	198.3	112.7

Appendix

Table 9: Inception Score (

\uparrow

) at six representative

(N,s)

configurations. Best single-pass result per column in bold .

Method	$(25,2)$	$(16,6)$	$(25,8)$	$(4,8)$
CFG	90.1 / 98.4	87.3 / 97.2	85.6 / 96.5	77.8 / 93.1
AGD-MultiScale	89.6 / 98.1	85.8 / 96.5	79.4 / 93.8	57.3 / 82.4
CtrlFlow-Core	89.4 / 98.0	85.1 / 96.2	83.7 / 95.6	69.2 / 89.1
CtrlFlow-Full	89.9 / 98.3	86.9 / 97.0	85.1 / 96.3	75.9 / 92.2

Appendix

Table 10: Class adherence. Top-1 / top-5 accuracy (%,

\uparrow

) of a frozen ResNet-50 on generated samples with respect to the conditioning class.

	$N{=}2$	$N{=}4$	$N{=}8$	$N{=}16$	$N{=}25$
$s{=}1$	$\pm$ 0.41	$\pm$ 0.23	$\pm$ 0.19	$\pm$ 0.15	$\pm$ 0.11
$s{=}2$	$\pm$ 0.26	$\pm$ 0.16	$\pm$ 0.10	$\pm$ 0.09	$\pm$ 0.07
$s{=}4$	$\pm$ 0.39	$\pm$ 0.24	$\pm$ 0.18	$\pm$ 0.14	$\pm$ 0.12
$s{=}6$	$\pm$ 0.55	$\pm$ 0.34	$\pm$ 0.26	$\pm$ 0.21	$\pm$ 0.18
$s{=}8$	$\pm$ 0.94	$\pm$ 0.51	$\pm$ 0.37	$\pm$ 0.30	$\pm$ 0.26

Appendix

Table 11: FID-50K standard deviation (

\pm

) across 3 generation seeds for CtrlFlow on DiT-XL/2.

Configuration	End of Phase 1	End of Phase 2	End of Phase 3
$(25,2)$	3.41	3.24	3.07
$(25,8)$	82.3	21.8	19.28

Appendix

Table 12: Training convergence. Validation FID-50K (

\downarrow

) at two configurations measured at the end of each curriculum phase.

Variant	$(4,2)$	$(4,8)$	$(16,2)$	$(16,8)$
Full CtrlFlow	9.47	48.95	3.52	22.14
w/o unified distillation	11.73	68.41	4.31	32.87
w/o guidance modulation	9.54	59.83	3.57	28.91
w/o time warping	9.51	63.42	3.54	24.67

Appendix

Table 13: Component ablation. FID-50K (

\downarrow

) at four

(N,s)

configurations. Each row removes one mechanism while keeping the others active.

Active sub-components	$(4,8)$	$(16,8)$
$w(t)\cdot\alpha(s)\cdot\varphi(a)$ (full)	48.95	22.14
$w(t)\cdot\alpha(s)$	51.34	23.87
$w(t)\cdot\varphi(a)$	52.18	24.15
$\alpha(s)\cdot\varphi(a)$	54.63	25.94
$w(t)$ only	53.47	25.31
$\alpha(s)$ only	55.21	26.83

Appendix

Table 14: Per-modulation sub-ablation. FID-50K (

\downarrow

). Each row activates only the listed sub-components.

Variant	FID $(16,2)$	FID $(16,8)$	Converged?
Full 3-phase	3.52	22.14	Yes
Skip Phase 1	5.21	35.28	Unstable
Skip Phase 2	3.74	37.93	Overfits
All $(N,s)$ from start	4.91	28.64	Slow

Appendix

Table 15: Curriculum ablation. FID-50K (

\downarrow

) at two configurations.

Rank $r$	Extra params	FID $(25,2)$	FID $(25,8)$	FID $(4,8)$
16	1.3%	3.54	24.18	61.34
32	2.5%	3.21	20.87	53.46
64 (default)	5.0%	3.07	19.28	48.95
128	10.0%	3.03	19.02	48.17

Appendix

Table 16: LoRA rank sensitivity. FID-50K (

\downarrow

) and parameter overhead at three configurations. Rank 64 sits at the knee of the cost-quality curve.

Variant	NFE	$(4,8)$	$(25,8)$	$(16,6)$	$(8,6)$
CFG	$2N$	53.84	21.47	16.02	20.47
CFG + $w\alpha$	$2N$	46.02	17.89	15.41	19.58
CFG + warp	$2N$	48.51	21.04	15.63	19.82
CFG + $w\alpha$ + warp	$2N$	42.31	17.42	14.88	19.04
CFG + $w\alpha\varphi$ + warp	$2N$	41.03	16.31	14.22	18.51
AGD-MultiScale	$N$	52.87	23.42	16.90	21.40

Appendix

Table 17: Mechanism transfer, FID-50K (

\downarrow

) on shared seed lists. Bold marks the best single-pass (

\text{NFE}=N

) entry per column.

(16,6)

and

(8,6)

were never used for hyperparameter selection.

Figure 4: FID ratio (CtrlFlow / CFG) across the $(N,s)$ grid. Left. Without time warping, dead zones appear at low $N$ and high $s$ (ratio $>1.10$ , four configurations at $s\geq 6$ ). At $s\leq 4$ , the warping magnitude is zero and both panels are identical. Right. With time warping, all configurations except $(2,8)$ stay at or below 1.04. High- $s$ configurations improve over CFG (green cells, ratio $<1$ ).

Method	$s=8$	$s=10$	$s=12$
CFG	21.47	31.24	42.81
AGD (fixed $s$ )	44.92	62.38	78.14
CtrlFlow	19.28	28.41	45.92

Appendix

Table 18: Scale extrapolation beyond training range (

s_{\max}{=}8

). FID-50K (

\downarrow

) at

N{=}25

DiT-XL/2 reference (from Table 2 )
Method	$(25,2)$	$(25,8)$	$(8,4)$	$(4,8)$
CFG (DiT-L/4)	5.52	34.82	20.14	82.41
AGD, fixed $s$ (DiT-L/4)	5.78	72.34	42.18	118.52
CtrlFlow (DiT-L/4)	5.71	32.14	19.23	76.38
CFG (DiT-XL/2)	2.97	21.47	11.72	53.84
CtrlFlow (DiT-XL/2)	3.07	19.28	10.98	48.95

Appendix

Table 19: Cross-backbone generalization. FID-50K (

\downarrow

) on DiT-L/4 (458M) at four

(N,s)

configurations. CtrlFlow uses the same training protocol as on DiT-XL/2.

DiT-XL/2 reference (from Table 23 )
Component	$t{=}0.8$	$t{=}0.5$	$t{=}0.2$	$t{=}0.05$
$\cos(\hat{\mathbf{v}}_{0},\,\mathbf{v}_{0}^{*})$ (DiT-L/4)	0.931	0.958	0.946	0.898
$\cos(\hat{\mathbf{g}},\,\mathbf{g}^{*})$ (DiT-L/4)	0.812	0.907	0.889	0.843
$\cos(\hat{\mathbf{v}}_{0},\,\mathbf{v}_{0}^{*})$ (DiT-XL/2)	0.942	0.968	0.957	0.911
$\cos(\hat{\mathbf{g}},\,\mathbf{g}^{*})$ (DiT-XL/2)	0.836	0.924	0.908	0.867

Appendix

Table 20: Decomposition fidelity on DiT-L/4. Cosine similarity (

\uparrow

) between learned and teacher-generated components, averaged over 1K samples at

s{=}2

Control setting	GFT (sweep $\beta_{\text{GFT}}$ )	CtrlFlow (sweep $s$ )
Weak guidance, $s\approx 1$	9.60	9.87
$s\approx 2$	3.18	3.07
$s\approx 4$	8.10	7.51
$s\approx 8$	21.40	19.28

Appendix

Table 21: CtrlFlow against the released GFT checkpoint, FID-50K (

\downarrow

) at

N{=}25

and matched NFE.

Method	NFE	FID-10K	CLIPScore
CFG, two-pass	$2N$	9.82	0.309
Scale-conditioned opaque student	$N$	11.37	0.300
CtrlFlow	$N$	10.41	0.304

Appendix

Table 22: Stable Diffusion 2.1 pilot. Zero-shot COCO FID-10K (

\downarrow

) and CLIPScore (

\uparrow

) at

N{=}25

s{=}7.5

Component	$t{=}0.8$	$t{=}0.5$	$t{=}0.2$	$t{=}0.05$
$\cos(\hat{\mathbf{v}}_{0},\,\mathbf{v}_{0}^{*})$	0.942	0.968	0.957	0.911
$\cos(\hat{\mathbf{g}},\,\mathbf{g}^{*})$	0.836	0.924	0.908	0.867

Appendix

Table 23: Decomposition fidelity. Cosine similarity (

\uparrow

) between learned and teacher-generated components, averaged over 1K samples at

s{=}2

	$s{=}2$ vs. $s{=}4$	$s{=}2$ vs. $s{=}8$	$s{=}4$ vs. $s{=}8$
$\cos(\hat{\mathbf{v}}_{0}^{(s_{a})},\,\hat{\mathbf{v}}_{0}^{(s_{b})})$	0.987	0.971	0.983

Appendix

Table 24: Cross-scale

\mathbf{v}_{0}

consistency. Cosine similarity (

\uparrow

) between

\hat{\mathbf{v}}_{0}

extracted at different guidance scales, averaged over 1K samples at

t{=}0.5

Anchor	$(25,1)$	$(25,2)$	$(8,4)$	$(25,8)$	$(4,8)$
Unconditional, $\mathbf{v}_{0}+s\,\mathbf{g}$ (ours)	9.92	3.14	11.48	23.65	73.41
Conditional, $\mathbf{v}_{\text{cond}}+(s-1)\,\mathbf{g}$	9.84	3.18	11.44	23.52	72.80

Appendix

Table 25: Anchor ablation without modulation or warping (Core), FID-50K (

\downarrow

) on shared class-balanced seed lists.

Anchor	$(25,8)$ Full	$(4,8)$ Full
Unconditional (ours)	19.28	48.95
Conditional	20.41	51.58

Appendix

Table 26: Anchor ablation under the deployed modulation and warping path, FID-50K (

\downarrow

Figure 5: Qualitative comparison at three $(N,s)$ operating points (same ImageNet class, golden retriever , per row). Rows show $(N{=}25,s{=}2)$ , $(N{=}25,s{=}8)$ , and $(N{=}4,s{=}8)$ , and columns show CFG ( $2N$ passes), the fixed-scale AGD variant ( $s{=}2$ , $N$ passes, labeled “AGD”), and CtrlFlow ( $N$ passes). At $(25,2)$ all methods produce comparable quality. At $(25,8)$ AGD exhibits color saturation and loss of detail. At $(4,8)$ AGD collapses to noisy outputs while CtrlFlow preserves recognizable structure. Aggregate FID values across all 1000 ImageNet classes are reported in Table 2 , and additional classes are in Figure 6 .

Figure 6: Extended qualitative comparison on two additional ImageNet classes ( giant panda , lion ) across the same three $(N,s)$ operating points and three methods as Figure 5 . Each block of three rows corresponds to one class, and columns show CFG ( $2N$ passes), the fixed-scale AGD variant ( $N$ passes, labeled “AGD”), and CtrlFlow ( $N$ passes) at a shared seed. The visual story is consistent across classes. AGD oversaturates and collapses at large $s$ , CtrlFlow tracks CFG closely, and the severity tracks the FID gaps reported in Table 2 .

Panel of Figure 7 — Figure 7: Cross-class interpretability maps (extending Figure 3 , which shows golden retriever ). Each panel shows guidance norm $\|\mathbf{g}\|$ (top row) and alignment $\cos(\mathbf{v}_{0},\mathbf{g})$ (bottom row) at four timesteps. The temporal pattern matches the main-text figure. Guidance is diffuse at $t{=}0.8$ and concentrates on class-discriminative regions by $t{=}0.05$ . The spatial focus varies by class, with wing patterns for butterfly and crater and smoke for volcano , confirming that $\mathbf{g}$ encodes class-specific semantic structure.

Map	Pointing acc.	Energy-in-box	Deletion AUC	Insertion AUC
Guidance norm, CtrlFlow	0.81	0.613	0.624	0.681
Guidance norm, two-pass teacher	0.80	0.601	0.612	0.669
Grad-CAM	0.74	0.548	0.587	0.642
Random	0.14	0.219	0.501	0.503

Appendix

Table 27: Localization of guidance-norm maps against ground-truth boxes (

\uparrow

for all columns).

Figure 8: Guidance modulation profiles. Left. Temporal modulation

w(t)

narrows with increasing

s

. Center. Scale dampening

\alpha(s)

activates above

s_{\text{thresh}}

. Right. Alignment gating

\varphi(a)

attenuates guidance under misalignment.

Figure 9: Time-grid warping. Left. Warped time

\tau(t)

deviates from the uniform diagonal at high

s

. Right. Dilation weight

\omega(t)

increases at trajectory endpoints for large guidance scales.

Figure 10: FID-50K vs. NFE at

s{=}4

. CtrlFlow (solid orange) at NFE

{}=N

achieves comparable FID to CFG (dashed gray) at NFE

{}=2N

, halving compute for equivalent quality.

Method	Source	Hyperparams	Train $s$	Eval $s$	Models	$s$ -free
CFG	Official DiT	Default	N/A	$[1,12]$	1	$\checkmark$
CFG-Zero*	Official repo	Default	N/A	$[1,8]$	1	$\checkmark$
APG	Official repo	Default	N/A	$[1,8]$	1	$\checkmark$
Autoguidance	Official repo	Default	N/A	$[1,8]$	1	$\checkmark$
CFG + Interval	Official repo	Default	N/A	$[1,8]$	1	$\checkmark$
AGD (fixed $s$ )	Ours, AGD design	Ablation	2.0	$[1,8]$	1	$\times$

Appendix

Table 28: Baseline protocol. All methods use DiT-XL/2 on ImageNet 256

\times

256. “Train

s

” is the guidance scale used during training. “Eval

s

” is the range tested. “

s

-free” indicates whether one checkpoint supports variable

s

at inference without retraining.

Method	Params	Extra	NFE	ms/img	Train (GPU-hr)
CFG	675M	0%	$2N$	843	0
CFG-Zero*	675M	0%	$2N$	854	0
APG	675M	0%	$2N$	852	0
Autoguidance	1015M	50%	$2N$	891	120
AGD (fixed $s$ )	689M	2%	$N$	432	32
CtrlFlow	709M	5%	$\boldsymbol{N}$	437	48

Appendix

Table 29: Computational cost comparison on DiT-XL/2 at

N{=}25

Component	ms/image
Frozen backbone, 25 steps at 16.9 ms	421.5
Adapter matrix multiplications, both branches	12.6
Modulation and composition	2.9
Warp precomputation (cached)	$<1$
Measured total	437

Appendix

Table 30: Latency breakdown of CtrlFlow at

N{=}25

, batch size 1, A100, 50 warmup and 200 timed iterations under CUDA synchronization.

Diffusion Guidance

Diffusion guidance steers generative sampling toward prompts, rewards, measurements, safety constraints, or other desired outcomes.

See all 116 papers

CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance
Mar 3, 2026Hanyang Wang, Yiyang Liu, Jiawei Chi +3Diffusion Model Guidance Classifier-Free Guidance
Tsinghua University
P-Guide: Parameter-Efficient Prior Steering for Single-Pass CFG Inference
May 7, 2026Xin Peng, Ang GaoFlow Matching Classifier-Free Guidance
School of Physical Science and Technology, Beijing University of Posts and Telecommunications, Beijing, China
C $^2$ FG: Control Classifier-Free Guidance via Score Discrepancy Analysis
Mar 9, 2026Jiayang Gao, Tianyi Zheng, Jiayang Zou +9Adaptive CFG Classifier-Free Guidance
Shanghai Jiao Tong University, Shanghai, China · vivo BlueImage Lab, vivo Mobile Communication Co., Ltd., China

Mar 3, 2026J/K move · Enter open · S save

CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance

Hanyang Wang, Yiyang Liu, Jiawei Chi, Fangfu Liu+2

Tsinghua University

Diffusion Model Guidance Classifier-Free Guidance Diffusion Model Alignment T2I Diffusion Models Sliding Mode Control

Open paper AskPDF arXiv

Classifier-Free Guidance (CFG) has emerged as a central approach for enhancing semantic alignment in flow-based diffusion models. In this paper, we explore a unified framework called CFG-Ctrl, which reinterprets CFG as a control applied to the first-order continuous-time generative flow, using the conditional-unconditional discrepancy as an error signal to adjust the velocity field. From this perspective, we summarize vanilla CFG as a proportional controller (P-control) with fixed gain, and typical follow-up variants develop extended control-law designs derived from it. However, existing methods mainly rely on linear control, inherently leading to instability, overshooting, and degraded semantic fidelity especially on large guidance scales. To address this, we introduce Sliding Mode Control CFG (SMC-CFG), which enforces the generative flow toward a rapidly convergent sliding manifold. Specifically, we define an exponential sliding mode surface over the semantic prediction error and introduce a switching control term to establish nonlinear feedback-guided correction. Moreover, we provide a Lyapunov stability analysis to theoretically support finite-time convergence. Experiments across text-to-image generation models including Stable Diffusion 3.5, Flux, and Qwen-Image demonstrate that SMC-CFG outperforms standard CFG in semantic alignment and enhances robustness across a wide range of guidance scales. Project Page: https://hanyang-21.github.io/CFG-Ctrl

Abstract

Diffusion Guidance

Explore similar work

CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance

P-Guide: Parameter-Efficient Prior Steering for Single-Pass CFG Inference

C2^22FG: Control Classifier-Free Guidance via Score Discrepancy Analysis

C $^2$ FG: Control Classifier-Free Guidance via Score Discrepancy Analysis