Revisiting Classifier-Free Guidance Methods in Latent Diffusion Models

Artem Sergievskii; Artyom Turevich; Sergey Kastryulin

cs.CVAug 17, 2026

Revisiting Classifier-Free Guidance Methods in Latent Diffusion Models

Authors: Artem Sergievskii, Artyom Turevich, Sergey Kastryulin

Organizations: HSE University · Yandex, HSE University

Abstract

Inference-time quality-enhancement methods are an effective and widely adopted means of improving diffusion models without expensive retraining. We study a family of training-free techniques conceptually rooted in Classifier-Free Guidance (CFG), most of which were originally proposed on older U-Net diffusion models and validated using metrics that assess image quality in isolation, without accounting for compositional alignment or semantic correspondence between the generated image and its associated text prompt. We re-evaluate eight such methods on two open-weight rectified-flow transformers under a fixed per-model protocol and three compositional-alignment benchmarks. No method consistently improves on CFG across the measured criteria. APG obtains several nominal best scores, but the corresponding gains often remain within the estimated evaluation uncertainty. Attention-perturbation methods provide isolated gains on SD3.5 Medium and more frequent degradations on FLUX.2 [klein] 4B Base, while CFG remains a competitive lower-cost baseline.

Figures & tables

	SD3.5 Medium					FLUX.2 [klein] 4B Base
Method	GE	DPG	Obj	Txt	Rsn	GE	DPG	Obj	Txt	Rsn
95% $\Delta$ margin	$\pm$ 0.027	$\pm$ 0.76	$\pm$ 0.019	$\pm$ 0.044	$\pm$ 0.005	$\pm$ 0.028	$\pm$ 0.85	$\pm$ 0.018	$\pm$ 0.042	$\pm$ 0.004
no-CFG	0.367	73.51	0.526	0.037	0.160	0.346	73.22	0.666	0.177	0.188
CFG	0.655	84.35	0.713	0.443	0.257	0.786	83.30	0.782	0.736	0.281
CFG++	0.617	84.29	0.749	0.459	0.232	0.704	83.35	0.784	0.753	0.284
CFG-Zero*	0.638	84.19	0.720	0.409	0.231	0.726	82.91	0.792	0.723	0.279

Table 1: Headline scores for both models. GE is GenEval; DPG is mean

\times

100; Obj, Txt, and Rsn are the selected OneIG families. Gray denotes differences within one shared reference margin of CFG; green/red denote larger gains/losses. A stronger shade beyond two margins is used only for visual emphasis. Bold marks the column maximum.

Panel of Figure 1 — Figure 1: Selected-setting outputs for SD3.5 Medium (left) and FLUX.2 [klein] 4B Base (right). Rows are methods. Columns use four shared prompts: a portrait of a woman wearing round glasses in a library; a dog to the right of a tie; a misty mountain valley with a river at sunrise; and a woman riding a white horse through red poppies.

Panel of Figure 2 — Figure 2: Selected FLUX.2 [klein] 4B Base cases where alternatives introduce errors relative to CFG (top) or improve its output (bottom). Columns are methods and rows are prompts.

Method	Original eval vs CFG	Comp.?	GenEval ( $\Delta$ )
CFG++	FID, CLIP, inversion (SD/SDXL)	no	$0.617$ ( $-0.038$ )
CFG-Zero*	T2I-CompBench, CLIP, Aesth. (SD3/3.5)	yes	$0.638$ ( $-0.017$ )
APG	FID, recall, saturation	no	$0.671$ ( $+0.016$ )
TCFG	FID, CLIP (MS-COCO; SD3)	no	$0.644$ ( $-0.011$ )
SAG	FID, IS (ImageNet, SD1.5)	no	$\mathbf{0.715}$ ( $+0.060$ )
PAG	FID, IS (SD1.5, ImageNet)	no	$0.595$ ( $-0.060$ )

Table 2: What each method’s original paper measured against CFG, versus our result. “Comp.?” marks whether the paper used a compositional-alignment benchmark comparable to ours. The last column is our SD3.5 GenEval Overall (CFG

=0.655

) with the signed difference.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Settings
no-CFG	$w=1$
CFG	$w=4.0$
CFG++	$\lambda=0.75$
CFG-Zero*	$w=3.5$ , zero-init steps $=0$
APG	$w=7$ , step radius $=20$ , momentum $=-0.4$
TCFG	$w=6$ , rank $=1$

Appendix

Table 3: Final hyperparameters per method, SD3.5 Medium (25 steps).

Method	Settings
no-CFG	guidance off
CFG	$w=7.0$
CFG++	$w=1.2$
CFG-Zero*	$w=7.0$ , zero-init steps $=2$
APG	$w=10$ , momentum $=-0.5$ , $\eta=0$ , step radius $=0$
TCFG	$w=12$ , rank $=1$

Appendix

Table 4: Final hyperparameters per method, FLUX.2 [klein] 4B Base (30 steps).

Method	single	two	count	color	pos	attr	Overall	single	two	count	color	pos	attr	Overall
	SD3.5 Medium							FLUX.2 [klein] 4B Base
no-CFG	0.767	0.500	0.194	0.419	0.097	0.226	0.367	0.800	0.242	0.225	0.596	0.120	0.090	0.346
CFG	1.000	0.833	0.484	0.806	0.226	0.581	0.655	0.988	0.889	0.788	0.862	0.560	0.630	0.786
CFG++	1.000	0.833	0.484	0.742	0.161	0.484	0.617	0.988	0.848	0.550	0.840	0.450	0.550	0.704
CFG-Zero*	1.000	0.733	0.484	0.839	0.258	0.516	0.638	1.000	0.818	0.663	0.883	0.410	0.580	0.726
APG	1.000	0.833	0.581	0.774	0.258	0.581	0.671	1.000	0.889	0.738	0.883	0.620	0.620	0.792

Appendix

Table 5: GenEval category scores for both models. Overall is the macro average.

Method	DPG	Attr	Entity	Other	Global	Relation	DPG	Attr	Entity	Other	Global	Relation
	SD3.5 Medium						FLUX.2 [klein] 4B Base
no-CFG	73.51	79.87	82.48	68.89	77.57	91.65	73.22	83.70	82.13	61.60	81.76	91.80
CFG	84.35	88.47	90.58	84.44	83.18	94.03	83.30	89.82	90.15	81.60	81.46	93.19
CFG++	84.29	87.07	90.08	81.11	85.05	94.03	83.35	89.62	90.03	81.20	82.07	93.73
CFG-Zero*	84.19	88.47	90.58	80.00	79.44	93.68	82.91	89.64	90.03	83.20	81.16	93.54
APG	84.54	88.82	90.99	80.00	83.18	94.75	83.50	89.48	90.40	78.00	81.46	93.93

Appendix

Table 6: DPG-Bench L1 scores for both models. DPG is mean

\times

100.

Method	color	other	shape	size	texture	color	other	shape	size	texture
	SD3.5 Medium					FLUX.2 [klein] 4B Base
no-CFG	83.46	78.78	78.46	66.67	78.60	87.05	82.71	75.54	69.01	83.49
CFG	92.93	87.50	80.00	70.97	87.64	93.18	87.47	82.97	76.86	89.90
CFG++	91.73	85.47	78.46	77.42	85.06	93.68	86.92	83.41	78.10	88.75
CFG-Zero*	93.53	86.92	81.54	74.19	86.53	93.33	86.47	83.84	76.86	89.59
APG	93.83	87.50	84.62	72.04	86.90	92.98	87.25	83.84	75.62	89.29

Appendix

Table 7: DPG-Bench L2 attribute scores for both models.

Method	part	state	whole	global	count	text	non-sp.	spatial
no-CFG	81.05	77.54	83.49	77.57	67.61	73.68	85.71	92.07
CFG	86.27	85.14	91.97	83.18	83.10	89.47	94.64	93.99
CFG++	79.74	82.97	92.36	85.05	81.69	78.95	96.43	93.86
CFG-Zero*	84.31	84.42	92.29	79.44	77.46	89.47	92.86	93.73
APG	84.31	85.51	92.61	83.18	77.46	89.47	91.07	95.01
TCFG	83.01	82.61	91.39	83.18	83.10	89.47	92.86	94.37

Appendix

Table 8: DPG-Bench L2 non-attribute scores for SD3.5 Medium.

Method	part	state	whole	global	count	text	non-sp.	spatial
no-CFG	80.47	81.52	82.43	81.76	57.50	78.00	84.91	92.25
CFG	84.77	84.69	91.82	81.46	78.50	94.00	86.79	93.61
CFG++	85.16	84.79	91.61	82.07	77.50	96.00	87.42	94.15
CFG-Zero*	85.74	85.11	91.48	81.16	80.00	96.00	86.79	93.98
APG	85.94	84.16	92.14	81.46	74.00	94.00	85.53	94.48
TCFG	83.98	84.16	91.95	81.16	79.00	94.00	86.16	93.73

Appendix

Table 9: DPG-Bench L2 non-attribute scores for FLUX.2 [klein] 4B Base.

Panel of Figure 3 — Figure 3: Guidance-scale sweep for the white-horse prompt.

Panel of Figure 4 — Figure 4: Guidance-scale sweep for the portrait prompt.

Panel of Figure 5 — Figure 5: Guidance-scale sweep for the mountain-valley prompt.

Explore similar work

Guidance Matters: Rethinking the Evaluation Pitfall for Text-to-Image Generation
Feb 26, 2026Dian Xie, Shitong Shao, Lichen Bai +5Diffusion Model Guidance T2I Generation Evaluation
The Hong Kong University of Science and Technology (Guangzhou) · Harbin Institute of Technology (Shenzhen) · Cogniser Information Technology
Guiding Diffusion Models with Semantically Degraded Conditions
Mar 11, 2026Shilong Han, Yuming Zhang, Hongxia WangDiffusion Model Guidance T2I Diffusion Models
College of Science, National University of Defense Technology
Guiding a Diffusion Model by Swapping Its Tokens
Apr 9, 2026Weijia Zhang, Yuehao Liu, Shanyan Guan +4Diffusion Model Guidance Diffusion Guidance
MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University · vivo Mobile Communication Co., Ltd.

Feb 26, 2026J/K move · Enter open · S save

Guidance Matters: Rethinking the Evaluation Pitfall for Text-to-Image Generation

Dian Xie, Shitong Shao, Lichen Bai, Zikai Zhou+4

The Hong Kong University of Science and Technology (Guangzhou) · Harbin Institute of Technology (Shenzhen) · Cogniser Information Technology

Diffusion Model Guidance T2I Generation Evaluation Classifier-Free Guidance

Open paper AskPDF arXiv

Classifier-free guidance (CFG) has helped diffusion models achieve great conditional generation in various fields. Recently, more diffusion guidance methods have emerged with improved generation quality and human preference. However, can these emerging diffusion guidance methods really achieve solid and significant improvements? In this paper, we rethink recent progress on diffusion guidance. Our work mainly consists of four contributions. First, we reveal a critical evaluation pitfall that common human preference models exhibit a strong bias towards large guidance scales. Simply increasing the CFG scale can easily improve quantitative evaluation scores due to strong semantic alignment, even if image quality is severely damaged (e.g., oversaturation and artifacts). Second, we introduce a novel guidance-aware evaluation (GA-Eval) framework that employs effective guidance scale calibration to enable fair comparison between current guidance methods and CFG by identifying the effects orthogonal and parallel to CFG effects. Third, motivated by the evaluation pitfall, we design Transcendent Diffusion Guidance (TDG) method that can significantly improve human preference scores in the conventional evaluation framework but actually does not work in practice. Fourth, in extensive experiments, we empirically evaluate recent eight diffusion guidance methods within the conventional evaluation framework and the proposed GA-Eval framework. Notably, simply increasing the CFG scales can compete with most studied diffusion guidance methods, while all methods suffer severely from winning rate degradation over standard CFG. Our work would strongly motivate the community to rethink the evaluation paradigm and future directions of this field.

Abstract

Explore similar work

Guidance Matters: Rethinking the Evaluation Pitfall for Text-to-Image Generation

Guiding Diffusion Models with Semantically Degraded Conditions

Guiding a Diffusion Model by Swapping Its Tokens