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
Method
Single Pass
Explicit v0/g
Post-hoc s
Two-Pass (NFE = 2N)
CFG
✗
✓ †
✓
CFG + Interval
✗
✓ †
✓
CFG-Zero*
✗
✗ \lx@sectionsign
✓
APG
✗
∼‡
✓
Autoguidance
✗
✗ \lx@paragraphsign
✓
Table 1: Comparison of guidance methods. Explicit v0/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 v0 and g outputs. Top. A frozen CFG teacher provides the target vcfgteacher . The student matches it via a unified loss on v^0+s⋅g^ . Bottom. At inference, one backbone pass produces v0 and g through two LoRA adapters. The scale s multiplies g post-hoc. Three modulation gates ( w , α , φ ) 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
∼ 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 ( ↓ ) on ImageNet 256 × 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 ( ↓ ). 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 g^
3.07
22.31
Swap v^0
3.01
21.74
CtrlFlow
3.07
19.28
Table 4: Component swap, FID-50K ( ↓ ).
Figure 3: Decomposition maps for golden retriever . Guidance norm becomes spatially focused as structure emerges, while cos(v0,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
μ
Window center
0.5
Eq. 23
σmax
Window width at s=1
0.35
σ(s)
σmin
Window width at s=smax
0.10
σ(s)
Scale Dampening
sthresh
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 ( ↓ ) for all 25 configurations on ImageNet 256 × 256. The boxed cell marks the sole configuration exceeding the CFG-match threshold δ . 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 ( ↓ ) under the protocol of Table 2 . “Grid match” counts configurations with FID≤FIDCFG+δ 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 ( ↑ ) / Recall ( ↑ ) 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 ( ↑ ) 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 (%, ↑ ) 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
± 0.41
± 0.23
± 0.19
± 0.15
± 0.11
s=2
± 0.26
± 0.16
± 0.10
± 0.09
± 0.07
s=4
± 0.39
± 0.24
± 0.18
± 0.14
± 0.12
s=6
± 0.55
± 0.34
± 0.26
± 0.21
± 0.18
s=8
± 0.94
± 0.51
± 0.37
± 0.30
± 0.26
Appendix
Table 11: FID-50K standard deviation ( ± ) 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 ( ↓ ) 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 ( ↓ ) at four (N,s) configurations. Each row removes one mechanism while keeping the others active.
Active sub-components
(4,8)
(16,8)
w(t)⋅α(s)⋅φ(a) (full)
48.95
22.14
w(t)⋅α(s)
51.34
23.87
w(t)⋅φ(a)
52.18
24.15
α(s)⋅φ(a)
54.63
25.94
w(t) only
53.47
25.31
α(s) only
55.21
26.83
Appendix
Table 14: Per-modulation sub-ablation. FID-50K ( ↓ ). 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 ( ↓ ) 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 ( ↓ ) 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 + wα
2N
46.02
17.89
15.41
19.58
CFG + warp
2N
48.51
21.04
15.63
19.82
CFG + wα + warp
2N
42.31
17.42
14.88
19.04
CFG + wαφ + 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 ( ↓ ) on shared seed lists. Bold marks the best single-pass ( 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≥6 ). At s≤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 ( smax=8 ). FID-50K ( ↓ ) at N=25 .
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
DiT-XL/2 reference (from Table 2 )
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 ( ↓ ) on DiT-L/4 (458M) at four (N,s) configurations. CtrlFlow uses the same training protocol as on DiT-XL/2.
Component
t=0.8
t=0.5
t=0.2
t=0.05
cos(v^0,v0∗) (DiT-L/4)
0.931
0.958
0.946
0.898
cos(g^,g∗) (DiT-L/4)
0.812
0.907
0.889
0.843
DiT-XL/2 reference (from Table 23 )
cos(v^0,v0∗) (DiT-XL/2)
0.942
0.968
0.957
0.911
cos(g^,g∗) (DiT-XL/2)
0.836
0.924
0.908
0.867
Appendix
Table 20: Decomposition fidelity on DiT-L/4. Cosine similarity ( ↑ ) between learned and teacher-generated components, averaged over 1K samples at s=2 .
Control setting
GFT (sweep βGFT )
CtrlFlow (sweep s )
Weak guidance, s≈1
9.60
9.87
s≈2
3.18
3.07
s≈4
8.10
7.51
s≈8
21.40
19.28
Appendix
Table 21: CtrlFlow against the released GFT checkpoint, FID-50K ( ↓ ) at N=25 and matched NFE.
Table 23: Decomposition fidelity. Cosine similarity ( ↑ ) 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(v^0(sa),v^0(sb))
0.987
0.971
0.983
Appendix
Table 24: Cross-scale v0 consistency. Cosine similarity ( ↑ ) between 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, v0+sg (ours)
9.92
3.14
11.48
23.65
73.41
Conditional, vcond+(s−1)g
9.84
3.18
11.44
23.52
72.80
Appendix
Table 25: Anchor ablation without modulation or warping (Core), FID-50K ( ↓ ) 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 ( ↓ ).
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 .
Figure 7: Cross-class interpretability maps (extending Figure 3 , which shows golden retriever ). Each panel shows guidance norm ∥g∥ (top row) and alignment cos(v0,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 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 ( ↑ for all columns).
Figure 8: Guidance modulation profiles. Left. Temporal modulation w(t) narrows with increasing s . Center. Scale dampening α(s) activates above sthresh . Right. Alignment gating φ(a) attenuates guidance under misalignment.
Figure 9: Time-grid warping. Left. Warped time τ(t) deviates from the uniform diagonal at high s . Right. Dilation weight ω(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
✓
CFG-Zero*
Official repo
Default
N/A
[1,8]
1
✓
APG
Official repo
Default
N/A
[1,8]
1
✓
Autoguidance
Official repo
Default
N/A
[1,8]
1
✓
CFG + Interval
Official repo
Default
N/A
[1,8]
1
✓
AGD (fixed s )
Ours, AGD design
Ablation
2.0
[1,8]
1
×
Appendix
Table 28: Baseline protocol. All methods use DiT-XL/2 on ImageNet 256 × 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%
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.
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
Classifier-Free Guidance (CFG) is essential for high-fidelity conditional generation in flow matching, yet it imposes significant computational overhead by requiring dual forward passes at each sampling step. In this work, we address this bottleneck by introducing P-Guide, a framework that achieves high-quality guidance through a single inference pass by modulating only the initial latent state. We further show that, under a first-order approximation, P-Guide is equivalent to CFG in the sense that it steers generation from the prior space, without requiring explicit velocity field extrapolation during sampling. We consider both homoscedastic and heteroscedastic priors, and find that jointly modeling the mean and variance enables adaptive loss attenuation and improved robustness to data uncertainty. Extensive experiments demonstrate that P-Guide reduces inference latency by approximately 50% while maintaining fidelity and prompt alignment competitive with standard dual-pass CFG baselines.
Xin Peng, Ang Gao
School of Physical Science and Technology, Beijing University of Posts and Telecommunications, Beijing, China
Classifier-Free Guidance (CFG) is a cornerstone of modern conditional diffusion models, yet its reliance on the fixed or heuristic dynamic guidance weight is predominantly empirical and overlooks the inherent dynamics of the diffusion process. In this paper, we provide a rigorous theoretical analysis of the Classifier-Free Guidance. Specifically, we establish strict upper bounds on the score discrepancy between conditional and unconditional distributions at different timesteps based on the diffusion process. This finding explains the limitations of fixed-weight strategies and establishes a principled foundation for time-dependent guidance. Motivated by this insight, we introduce \textbf{Control Classifier-Free Guidance (C2FG)}, a novel, training-free, and plug-in method that aligns the guidance strength with the diffusion dynamics via an exponential decay control function. Extensive experiments demonstrate that C2FG is effective and broadly applicable across diverse generative tasks, while also exhibiting orthogonality to existing strategies.
Jiayang Gao, Tianyi Zheng, Jiayang Zou +9
Shanghai Jiao Tong University, Shanghai, China · vivo BlueImage Lab, vivo Mobile Communication Co., Ltd., China