TrajCFG: Token-Wise Classifier-Free Guidance via Denoising Trajectory Feedback

Jingyu Liu; Hao Jiang; Yikang Ding; Huafeng Shi; Xu Yang; Xiaoqiang Liu; Guoxin Zhang; Pengfei Wan

cs.CVOct 9, 2026

TrajCFG: Token-Wise Classifier-Free Guidance via Denoising Trajectory Feedback

Authors: Jingyu Liu, Hao Jiang, Yikang Ding, Huafeng Shi, Xu Yang, Xiaoqiang Liu, Guoxin Zhang, Pengfei Wan, +1 more

Organizations: Renmin University of China · Kling Team, Kuaishou Technology

Abstract

Classifier-free guidance (CFG) typically applies a shared scale to all latent tokens, although their velocity predictions evolve differently during sampling. We introduce Trajectory-Adaptive Classifier-Free Guidance (TrajCFG), a training-free method for flow-matching models that adapts token-wise guidance using historical changes in guided velocity direction. TrajCFG maps this feedback to bounded scales, assigning stronger guidance to larger directional changes and weaker guidance to more consistent directions. It combines adjacent-step and interval-normalized longer-span feedback, reusing standard CFG predictions without additional model evaluations. Controlled diagnostics and ablations examine the roles of directional feedback, token-wise allocation, and history-span fusion. Experiments on text-to-video, image-to-video, text-to-audio-video, text-to-image, and image-to-3D generation show improvements over base samplers across multiple backbones. For text-to-video generation, TrajCFG improves VBench Total scores by 0.30-0.75 points over the strongest evaluated baseline on each of three backbones. These results support historical directional feedback as a broadly applicable signal for fine-grained guidance allocation. The project page is available at https://jyliu-98.github.io/trajcfg-project-page/.

Figures & tables

Figure 1: Overview of TrajCFG. Standard CFG uses a fixed scale (purple), while TrajCFG (teal) adapts each token’s scale using guided velocities from the two most recent completed steps. Larger historical angles yield stronger guidance; smaller angles yield weaker guidance. Predictions are reused without extra model evaluations. For clarity, the illustration shows adjacent-step feedback. Trajectories are schematic.

Figure 2: Content-related structure in historical angles. Generated frames and historical velocity-angle maps at sampling step $k=5$ from Wan2.2-14B on Dynamic-Bench. Warmer colors indicate larger angular changes.

Figure 3: CFG response by historical angle group. Groups are fixed across candidate CFG scales. Clean-latent MSE differences are relative to CFG

=5

(negative is better); shading indicates standard deviations.

Method	Wan2.1-1.3B						Wan2.2-5B						Wan2.2-14B						Time
	VBench			Dynamic-Bench			VBench			Dynamic-Bench			VBench			Dynamic-Bench
	Qual.	Sem.	Total	Mot.	Over.	Avg.	Qual.	Sem.	Total	Mot.	Over.	Avg.	Qual.	Sem.	Total	Mot.	Over.	Avg.
Base	83.03	72.77	80.98	96.40	22.36	66.18	82.88	74.52	81.21	96.58	24.00	67.45	82.93	73.05	80.96	98.14	23.47	69.14	1.0 $\times$
CFG-Zero	82.69	73.95	80.94	97.13	23.81	66.64	83.79	74.84	82.00	96.68	24.19	67.94	83.32	75.27	81.71	98.02	25.18	69.41	1.0 $\times$
CFG-Ctrl	83.65	72.68	81.45	97.28	24.18	66.90	84.12	76.75	82.64	96.95	24.09	67.42	83.09	76.11	81.70	98.38	24.79	69.50	1.0 $\times$
PAG	83.24	73.92	81.38	97.15	23.89	67.25	83.37	76.14	81.93	96.91	24.11	67.63	83.07	73.31	81.12	98.11	24.90	69.46	1.5 $\times$

Table 1: Text-to-video generation results on VBench and Dynamic-Bench. VBench reports the official Quality (Qual.), Semantic (Sem.), and Total scores. For Dynamic-Bench, Mot., Over., and Avg. denote Motion Smoothness, Overall Consistency, and the average over evaluated VBench dimensions under the custom protocol. Higher scores are better. Runtime is measured relative to Base (

1.0\times

). Best and second-best scores in each metric column are bold and underlined, respectively.

Method	Cosmos-Predict2.5-2B						Wan2.2-5B						Wan2.2-14B						Time
	VBench-I2V			PAI-Bench-G			VBench-I2V			PAI-Bench-G			VBench-I2V			PAI-Bench-G
	Img.	Mot.	Qual.	Img.	Mot.	Qual.	Img.	Mot.	Qual.	Img.	Mot.	Qual.	Img.	Mot.	Qual.	Img.	Mot.	Qual.
Base	98.46	99.09	66.87	96.76	99.09	61.32	98.74	99.02	67.49	97.14	98.72	59.56	98.51	98.77	66.87	96.69	98.50	58.86	1.0 $\times$
CFG-Zero	98.53	99.14	67.07	96.83	99.13	61.61	98.25	98.92	67.17	96.86	98.72	60.00	97.69	98.69	65.65	96.44	98.48	58.52	1.0 $\times$
CFG-Ctrl	98.13	98.97	66.88	95.85	98.37	60.93	98.37	99.10	66.37	97.23	98.97	59.81	98.48	98.79	66.63	96.42	98.52	58.18	1.0 $\times$
PAG	98.46	99.10	66.87	96.71	99.08	61.62	98.71	99.02	67.37	96.95	98.69	60.26	98.20	98.76	66.85	96.66	98.53	58.69	1.5 $\times$

Table 2: Image-to-video generation results on VBench-I2V and PAI-Bench-G. Img. (Image Alignment) denotes the mean of I2V Subject and I2V Background; Mot. denotes Motion Smoothness; and Qual. denotes the mean of Aesthetic Quality and Imaging Quality. Higher scores are better. Runtime is measured relative to Base (

1.0\times

). Best and second-best scores in each metric column are bold and underlined, respectively.

Panel of Figure 4 — Figure 4: Qualitative comparisons for video generation. (a) TrajCFG is compared with Base on Wan2.1-1.3B/VBench (left) and CFG-Zero on Wan2.2-14B/Dynamic-Bench (right). (b) TrajCFG is compared with CFG-Ctrl on Cosmos-Predict2.5-2B/VBench-I2V (left) and SRVS on Wan2.2-14B/PAI-Bench-G (right); input images appear beside the prompts. Guidance-scale rows visualize TrajCFG’s token-wise scales at sampling step $k=5$ , with warmer colors indicating stronger guidance. Red boxes highlight artifacts in the competing-method sequences.

	Variant	Wan2.1-1.3B			Wan2.2-5B
	Variant	Quality	Semantic	Total	Quality	Semantic	Total
	TrajCFG (Full)	84.06	75.37	82.32	84.82	77.45	83.35
Q1	Time-only	83.94	73.17	81.78	84.62	76.94	83.08
Q1	Sample-wise	83.83	74.02	81.87	84.48	76.34	82.85
Q2	Shuffled allocation	83.77	73.67	81.75	84.43	76.73	82.89
Q2	Reversed allocation	83.70	72.96	81.55	84.55	75.82	82.81

Table 3: Ablation study of TrajCFG for text-to-video generation on VBench. We report the official Quality, Semantic, and Total scores (higher is better). Fusion weights

(w_{1},w_{2})

correspond to adjacent-step (

h=1

) and stride-2 (

h=2

) guidance estimates, respectively; TrajCFG (Full) uses

(0.8,0.2)

. Best and second-best scores in each metric column are bold and underlined, respectively.

Backbone	Method	VABench					T2AV-Compass
Backbone	Method	Visual Realism $\uparrow$	Audio Aes. $\uparrow$	T-A $\uparrow$	A-V $\uparrow$	Desync $\downarrow$	VT $\uparrow$	AA $\uparrow$	T-A $\uparrow$	A-V $\uparrow$	DS $\downarrow$
LTX-2.3	Base	4.615	3.345	0.3118	0.2490	0.5911	7.295	6.709	0.1722	0.2469	0.7140
LTX-2.3	TrajCFG	4.638	3.351	0.3202	0.2567	0.5685	7.510	6.738	0.1735	0.2481	0.7016
Ovi	Base	4.522	3.282	0.3288	0.1564	0.7659	6.714	6.676	0.0897	0.1596	0.9624
Ovi	TrajCFG	4.547	3.364	0.3316	0.1654	0.7080	7.199	6.733	0.0922	0.1622	0.9356

Table 4: Text-to-audio-video results. Audio Aes., T-A, A-V, VT, AA, and DS denote Audio Aesthetics, text–audio alignment, audio–visual alignment, Video Technological, Audio Attribute, and DeSync. Lower is better for Desync and DS; best scores within each backbone are bold.

Backbone	Method	GenEval							DPG-Bench
Backbone	Method	Single Obj. $\uparrow$	Two Obj. $\uparrow$	Counting $\uparrow$	Colors $\uparrow$	Position $\uparrow$	Attr. Bind. $\uparrow$	Overall $\uparrow$	Global $\uparrow$	Entity $\uparrow$	Attribute $\uparrow$	Relation $\uparrow$	Other $\uparrow$	Overall $\uparrow$
Qwen-Image	Base	0.9875	0.9318	0.6094	0.8324	0.4950	0.5825	0.7412	89.39	90.84	89.48	91.76	90.63	87.11
Qwen-Image	TrajCFG	0.9906	0.9394	0.6188	0.8617	0.4950	0.6050	0.7517	91.03	93.20	90.40	93.21	94.35	87.19
SD3-Medium	Base	0.9969	0.8611	0.6125	0.8351	0.3125	0.5650	0.7021	86.98	84.24	86.26	87.30	85.19	83.56
SD3-Medium	TrajCFG	1.0000	0.8813	0.6344	0.8484	0.3225	0.5975	0.7140	88.35	86.37	86.55	89.38	86.15	83.99

Table 5: Text-to-image results. Higher is better; best scores within each backbone are bold, including ties.

Backbone	Method	Shape				Texture
Backbone	Method	ULIP-T $\uparrow$	ULIP-I $\uparrow$	Uni3D-T $\uparrow$	Uni3D-I $\uparrow$	CMMD $\downarrow$	$\mathbf{FID}_{\mathbf{CLIP}}\downarrow$	CLIP-score $\uparrow$	LPIPS $\downarrow$
Hunyuan3D 2.0	Base	0.1753	0.2277	0.2190	0.2714	0.0571	0.0606	0.2503	0.1015
Hunyuan3D 2.0	TrajCFG	0.1762	0.2290	0.2198	0.2719	0.0532	0.0581	0.2505	0.1008

Table 6: Image-to-3D results on GSO. Metrics follow Hunyuan3D 2.0; -T and -I denote mesh–text and mesh–image similarity. Arrows indicate direction; best scores are bold.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Backbone	$\bm{s_{\mathrm{base}}}$	$\bm{s_{\min}}$	$\bm{s_{\max}}$	$\bm{\rho_{\mathrm{lo}}}$	Steps
T2V	Wan2.1-1.3B	6.0	5.0	10.0	0.95	40
	Wan2.2-5B	5.0	3.0	10.0	0.95	40
	Wan2.2-14B	3.5	1.0	10.0	0.90	40
I2V	Cosmos-Predict2.5-2B	5.0	4.0	9.0	0.90	35
	Wan2.2-5B	5.0	2.0	10.0	0.95	40
	Wan2.2-14B	2.5	1.0	7.0	0.95	40

Appendix

Table 7: Hyperparameter settings for each task and backbone.

Figure 5: Additional qualitative comparisons for text-to-video generation. Results are shown for Wan2.1-1.3B, Wan2.2-5B, and Wan2.2-14B on VBench and Dynamic-Bench. Each example presents selected frames from Base and TrajCFG together with the token-wise guidance map produced by TrajCFG; red boxes mark representative Base failures.

Figure 6: Additional qualitative comparisons for image-to-video generation. The three rows show Cosmos-Predict2.5-2B, Wan2.2-5B, and Wan2.2-14B, while the left and right columns use VBench-I2V and PAI-Bench-G, respectively. Each example presents the input image, selected frames from Base and TrajCFG, and the token-wise guidance map produced by TrajCFG; red boxes mark representative Base failures.

Figure 7: Qualitative comparisons for text-to-audio-video generation. The four examples show Ovi on VABench (upper left), Ovi on T2AV-Compass (upper right), LTX-2.3 on VABench (lower left), and LTX-2.3 on T2AV-Compass (lower right). Each example pairs selected video frames with the corresponding audio waveform. Arrows mark sound-event intervals, while red boxes highlight spatial or temporal discrepancies in Base.

Figure 8: Qualitative comparisons for text-to-image generation. The four groups show SD3-Medium on GenEval (upper left), SD3-Medium on DPG-Bench (upper right), Qwen-Image on GenEval (lower left), and Qwen-Image on DPG-Bench (lower right). Within each group, Base and TrajCFG are shown in the top and bottom rows, respectively. Red text highlights selected prompt requirements.

Figure 9: Qualitative comparisons for image-to-3D generation. Input images and two rendered views from Base and TrajCFG. Red boxes highlight appearance discrepancies in Base.

Backbone	Method	Aesthetic	Appearance	Background	Color	Dynamic	Human	Imaging	Motion	Multiple	Object
Backbone	Method	Quality	Style	Consistency	Color	Degree	Action	Quality	Smoothness	Objects	Class
Wan2.1-1.3B	Base	64.11	22.04	96.70	84.71	65.28	91.00	62.92	98.00	71.19	87.42
	CFG-Zero	65.26	22.16	95.61	87.67	68.90	93.00	64.91	97.41	69.74	86.31
	CFG-Ctrl	64.49	22.12	96.77	86.69	68.06	93.00	64.80	98.09	68.90	84.97
	PAG	64.76	22.24	96.58	85.62	66.67	93.00	64.71	97.89	70.50	85.76
	SRVS	64.87	22.37	96.64	82.77	66.67	92.00	64.94	98.19	71.49	88.92

Appendix

Table 8: Complete per-dimension results for text-to-video (T2V) generation on VBench.

Backbone	Variant	Aesthetic	Appearance	Background	Color	Dynamic	Human	Imaging	Motion	Multiple	Object
Backbone	Variant	Quality	Style	Consistency	Color	Degree	Action	Quality	Smoothness	Objects	Class
Wan2.1-1.3B	TrajCFG (Full)	65.48	22.45	96.90	86.82	70.83	93.75	64.71	98.12	75.68	89.14
	Time-only	64.49	22.08	97.14	85.06	69.44	95.00	65.02	98.16	72.26	86.55
	Sample-wise	64.17	22.13	97.02	86.77	69.44	95.00	64.90	98.15	71.04	87.10
	Shuffled allocation	64.12	22.17	96.99	84.68	69.44	94.00	64.63	98.15	72.26	87.74
	Reversed allocation	64.01	22.28	96.98	86.67	69.44	94.00	64.32	98.16	69.59	85.21

Appendix

Table 9: Complete per-dimension ablation results for text-to-video (T2V) generation on VBench.

Backbone	Method	Subject	Background	Motion	Aesthetic	Imaging	Overall
Backbone	Method	Consistency	Consistency	Smoothness	Quality	Quality	Consistency
Wan2.1-1.3B	Base	85.54	91.42	96.40	45.17	56.21	22.36
	CFG-Zero	85.75	91.38	97.13	46.39	55.40	23.81
	CFG-Ctrl	86.80	92.47	97.28	44.40	56.27	24.18
	PAG	87.27	92.55	97.15	46.23	56.40	23.89
	SRVS	87.27	92.65	97.38	47.07	56.39	24.15

Appendix

Table 10: Complete per-dimension results for text-to-video (T2V) generation on Dynamic-Bench.

Backbone	Method	I2V	I2V	Subject	Background	Motion	Aesthetic	Imaging	Overall
Backbone	Method	Subject	Background	Consistency	Consistency	Smoothness	Quality	Quality	Consistency
Cosmos-Predict2.5-2B	Base	98.18	98.74	96.00	96.18	99.09	62.23	71.51	23.48
	CFG-Zero	98.26	98.79	96.39	96.42	99.14	62.46	71.67	23.50
	CFG-Ctrl	97.83	98.43	95.37	95.53	98.97	62.21	71.55	23.36
	PAG	98.18	98.74	95.95	96.16	99.10	62.25	71.49	23.44
	SRVS	97.77	98.47	94.68	96.07	99.11	62.27	71.48	23.42

Appendix

Table 11: Complete per-dimension results for image-to-video (I2V) generation on VBench-I2V.

Backbone	Method	I2V	I2V	Subject	Background	Motion	Aesthetic	Imaging	Overall
Backbone	Method	Subject	Background	Consistency	Consistency	Smoothness	Quality	Quality	Consistency
Cosmos-Predict2.5-2B	Base	96.27	97.25	92.43	94.17	99.09	52.46	70.18	21.18
	CFG-Zero	96.33	97.32	92.70	94.38	99.13	52.83	70.38	21.19
	CFG-Ctrl	95.33	96.36	91.23	93.54	98.37	51.59	70.27	20.94
	PAG	96.21	97.21	92.14	94.08	99.08	52.87	70.36	21.17
	SRVS	95.40	96.65	91.67	93.71	99.10	52.86	70.43	21.15

Appendix

Table 12: Complete per-dimension results for image-to-video (I2V) generation on PAI-Bench-G.

Backbone	Method	Speech	Audio	T–V	T–A	A–V	Lip–Sync	Desync $\downarrow$	Alignment	Expressiveness	Visual	Audio	Audio	Visual
Backbone	Method	Q&N	Aes.	Align.	Align.	Align.		Desync $\downarrow$	Alignment	Expressiveness	Realism	Realism	QA	QA
LTX-2.3	Base	2.698	3.345	0.1928	0.3118	0.2490	2.752	0.5911	4.470	4.299	4.615	4.032	0.7049	0.6616
LTX-2.3	TrajCFG	2.763	3.351	0.1963	0.3202	0.2567	2.555	0.5685	4.479	4.321	4.638	4.066	0.7178	0.6730
Ovi	Base	3.181	3.282	0.1968	0.3288	0.1564	4.420	0.7659	4.189	4.138	4.522	4.063	0.5251	0.5948
Ovi	TrajCFG	3.463	3.364	0.1977	0.3316	0.1654	4.780	0.7080	4.179	4.141	4.547	4.074	0.5257	0.5953

Appendix

Table 13: Complete per-dimension results for text-to-audio-video (T2AV) generation on VABench. Lower is better for Desync.

Backbone	Method	Video Quality		Audio Quality		Cross-modal Alignment
Backbone	Method	VT $\uparrow$	VA $\uparrow$	AA $\uparrow$	SQ $\uparrow$	A–V $\uparrow$	T–A $\uparrow$	T–V $\uparrow$	DS $\downarrow$	LS $\uparrow$
LTX-2.3	Base	7.295	4.921	6.709	1.562	0.2469	0.1722	0.0432	0.7140	2.484
LTX-2.3	TrajCFG	7.510	4.924	6.738	1.578	0.2481	0.1735	0.0423	0.7016	2.820
Ovi	Base	6.714	4.516	6.676	1.566	0.1596	0.0897	0.0297	0.9624	1.939
Ovi	TrajCFG	7.199	4.555	6.733	1.665	0.1622	0.0922	0.0316	0.9356	2.449

Appendix

Table 14: Complete per-dimension results for text-to-audio-video (T2AV) generation on T2AV-Compass. Lower is better for DS.

Explore similar work

Classifier-Free Guidance in Flow Matching: Non-Autonomous Potentials, Overshoot, and Posterior-Mean Control
Sep 21, 2026Jishen Peng, Zheng MaFlow Matching Classifier-Free Guidance
School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai, China · Institute of Natural Sciences, MOE-LSC, Shanghai Jiao Tong University, Shanghai, China · Qing Yuan Research Institute, Shanghai Jiao Tong University, Shanghai, China +1
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
VAGS: Velocity Adaptive Guidance Scale for Image Editing and Generation
May 15, 2026Yan Luo, Ahmadou Aidara, Jingyi Lu +3Classifier-Free Guidance Diffusion-Based Image Generation
Harvard AI and Robotics Lab, Harvard University · School of Computing and Data Science, The University of Hong Kong · Kempner Institute for the Study of Natural and Artificial Intelligence, Harvard University

Sep 21, 2026J/K move · Enter open · S save

Classifier-Free Guidance in Flow Matching: Non-Autonomous Potentials, Overshoot, and Posterior-Mean Control

Jishen Peng, Zheng Ma

School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai, China · Institute of Natural Sciences, MOE-LSC, Shanghai Jiao Tong University, Shanghai, China · Qing Yuan Research Institute, Shanghai Jiao Tong University, Shanghai, China+1

Flow Matching Classifier-Free Guidance Conditional Image Generation Flow-Based Generative Modeling

Open paper AskPDF arXiv

Classifier-free guidance (CFG) improves conditional generation in Flow Matching, but strong guidance can distort the generated distribution and reduce diversity. We provide a geometric account of this behavior by viewing Flow Matching as a time-varying gradient flow and characterizing how CFG reshapes its underlying potential. This view explains how stronger alignment can be accompanied by mean displacement and trajectory concentration, and motivates controlling guidance through the model-implied terminal posterior mean. We therefore propose Posterior-Mean-Capped CFG (PMC-CFG), a training-free, per-sample method that adaptively retains the strongest feasible guidance without additional network evaluations. Experiments on synthetic and large-scale image-generation benchmarks show that PMC-CFG limits guidance-induced distortion and concentration while improving the alignment--diversity trade-off, with particularly strong benefits when nominal guidance is large.

Abstract

Explore similar work

Classifier-Free Guidance in Flow Matching: Non-Autonomous Potentials, Overshoot, and Posterior-Mean Control

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

VAGS: Velocity Adaptive Guidance Scale for Image Editing and Generation