Custom Forcing: Training-Free Subject Customization for Autoregressive Video Generation

Yunseung Ok; Hyunsoo Kim; Minseo Kim; Suhyun Kim

cs.CVOct 2, 2026

Custom Forcing: Training-Free Subject Customization for Autoregressive Video Generation

Authors: Yunseung Ok, Hyunsoo Kim, Minseo Kim, Suhyun Kim

Organizations: Kyung Hee University · The University of Texas at Austin

Abstract

Autoregressive video models can generate minute-long videos in real time, but they produce generic subjects from text rather than specific subjects from user-provided images. Existing customization methods either require costly per-subject optimization or use pretrained conditioning networks that jointly process all video frames with bidirectional attention. Neither approach is designed for causal streaming. We present Custom Forcing, a training-free method that stores reference-based anchor frames in the persistent KV cache of a frozen autoregressive video model. However, fixed anchors face two limitations: simple conditioning allows identity to drift, and the text prompt continues to favor a generic subject. To address these problems, drift-adaptive value amplification (DVA) scales reference influence with the degree of identity drift, while anchor contrast guidance (ACG) steers generation away from the generic class prior. Over two-minute rollouts, fixed anchors fall from 0.58 to 0.42 in DINO-I, while Custom Forcing keeps it between 0.58 and 0.62 without reducing motion. Custom Forcing also achieves higher subject similarity than bidirectional customization methods and better preserves identity over 30s than causal image-to-video and reference-to-video models, while generating each frame 9.5-28.5 times faster than these long-video baselines.

Figures & tables

\fnum@figure : DINO-I over two minutes ,

n=100

. Custom K/V only uses fixed anchor frames without DVA or ACG.

\fnum@figure : Overview of Custom Forcing . Top: the reference images and copies of a custom anchor generated by FLUX.2 [klein] ( Black Forest Labs, 2025 ) are written into the KV cache once as the custom K/V. Middle: in every self-attention layer of the frozen backbone, the current chunk reads the cache with and without the anchor frames, giving $z_{+}$ and $z_{-}$ , after DVA scales the reference values by $g(t)$ . Bottom left: DVA sets $g(t)$ from how far the similarity $s(t)$ of each finalized chunk falls below its initial level $\mu_{\mathrm{base}}$ . Bottom right: ACG extrapolates away from $z_{-}$ by $\kappa$ and blends the result with $z_{+}$ .

Bidirectional
				DINO-I over time $\uparrow$				Alignment $\uparrow$		Video quality
Method	Length	Long video	Training free	all	first	middle	last	CLIP-I	CLIP-T	Subj. Cons. $\uparrow$	Bg. Cons. $\uparrow$	Dyn. Degree
VACE-1.3B	5.1 s	✗	✓	0.333	–	–	–	0.703	0.296	0.921	0.951	0.46
SMRABooth	3.3 s	✗	✗	0.407	–	–	–	0.664	0.309	0.978	0.969	0.21
CustomCrafter	2.0 s	✗	✗	0.490	–	–	–	0.669	0.301	0.966	0.971	0.17
MotionBooth	3.0 s	✗	✗	0.435	–	–	–	0.642	0.293	0.907	0.928	0.76

Table 1: Main results. Long video marks whether a method can extend generation beyond a fixed length, and Training free whether it customizes the subject without training any model on it. Bold marks the best value in each column among the bidirectional methods and the

30

s rows, and the better of the two

2

min rows. Dynamic Degree is not bolded. Subject and Background Consistency favor the short bidirectional videos, which change little.

\fnum@figure : Two-minute generation. Custom K/V only drifts within the first minute, while Custom Forcing keeps both subjects for two minutes.

				DINO-I over time $\uparrow$				Alignment $\uparrow$
Method	Backbone	Time	Causal	all	first	middle	last	CLIP-I	CLIP-T
FramePack	HunyuanVideo 13B	9.4×	✗	0.625	0.616	0.620	0.623	0.795	0.357
FramePack-F1	HunyuanVideo 13B	9.5×	✓	0.477	0.534	0.418	0.392	0.773	0.355
SkyReels-V2 DF	Wan2.1 1.3B	11.5×	✓	0.449	0.552	0.393	0.307	0.767	0.361
SkyReels-V3	Wan2.1 14B	28.5×	✓	0.641	0.693	0.620	0.564	0.794	0.347
Custom Forcing (Ours)	Wan2.1 1.3B	1.0×	✓	0.635	0.619	0.626	0.616	0.803	0.353

Table 2: Comparison with image- and reference-to-video models ,

30

n=100

, identity and prompt alignment only. Backbone gives the architecture and size of each video diffusion transformer, Time is the generation time per frame relative to Custom Forcing on one H100 GPU, and Causal marks whether each part of the video is final before the next is generated. Bold marks the best score and underline the second best.

\fnum@figure : Qualitative comparison with image- and reference-to-video models. The image-to-video models start from the custom anchor, and SkyReels-V3 starts from the reference images, so its scenes differ. Custom Forcing keeps both subjects to the end.

			DINO-I $\uparrow$
	DVA	ACG	first	mid.	last
Custom K/V only	✗	✗	0.579	0.506	0.475
+ DVA	✓	✗	0.600	0.598	0.598
+ DVA + ACG (Ours)	✓	✓	0.619	0.626	0.616

Table 3: Component ablation ,

30

\fnum@figure : Adding the components one at a time. Red boxes mark where the subject departs from the reference images.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

	Symbol	Value
Anchor slots	$N$	$9$
Reference anchors	$\|\mathcal{R}\|$	$5$
Retained history frames	$H$	$3$
Chunk size / tokens per frame	$F$ / $L$	$3$ / $1560$
Injection budget	$\alpha_{\text{inj}}$	$300$
Drift cap	$b_{\max}$	$0.2$

Appendix

Table 4: Hyperparameters of Custom Forcing . No value is tuned per subject or per prompt, and only

b(t)

adapts to each video.

				DINO-I over time $\uparrow$				Alignment $\uparrow$		Video quality
Method	Length	DVA	ACG	all	first	middle	last	CLIP-I	CLIP-T	Subj. Cons. $\uparrow$	Bg. Cons. $\uparrow$	Dyn. Degree
Vanilla AR model	30 s	–	–	0.167	0.169	0.158	0.149	0.692	0.364	0.873	0.916	0.79
Custom K/V only	30 s	✗	✗	0.556	0.592	0.516	0.505	0.789	0.355	0.890	0.923	0.81
+ DVA	30 s	✓	✗	0.614	0.614	0.584	0.562	0.787	0.347	0.878	0.921	0.62
+ DVA + ACG (Ours)	30 s	✓	✓	0.641	0.634	0.623	0.615	0.798	0.348	0.889	0.923	0.87
Custom K/V only	2 min	✗	✗	0.518	0.586	0.466	0.447	0.780	0.354	0.859	0.911	0.84

Appendix

Table 5: Results on Deep Forcing ,

n=100

. Columns and bold follow Tab. 1 , and rows follow Tab. 3 .

Panel of Figure 11 — \fnum@figure : Custom Forcing on Deep Forcing. The custom anchor is on the left of each row. (a) $30$ s. Top: “A toy marching along a paved park path lined with autumn trees, …”. Bottom: “A backpack standing upright on a quiet city sidewalk lined with shopfronts, …”. (b) Two minutes. Top: “A toy rolling across a wide stone plaza in a historic square, …”. Bottom: “A dog walking along a quiet city sidewalk lined with shopfronts, …”. Custom Forcing keeps the red body, yellow belly, and large eyes of the toy and the red color and shape of the backpack to the end of the $30$ s videos, and the white body, black visor, and blue circles of the robot and the red-brown coat and drooping ears of the dog for two minutes.

\fnum@figure : Effect of the ACG warm-up on Deep Forcing , Without warm-up, the generation remains close to the custom-anchor pose and exhibits limited motion. With warm-up, the subject shows greater motion while preserving its identity.

Subject	Custom anchor	Custom K/V only	Custom Forcing	Gain
		DINO-I, last window
backpack	0.669	0.647	0.662	$+0.015$
cat	0.695	0.425	0.636	$+0.211$
cat2	0.788	0.481	0.702	$+0.221$
dog	0.656	0.509	0.617	$+0.108$
dog3	0.460	0.055	0.316	$+0.261$

Appendix

Table 6: Results per subject ,

30

10

prompts each.

	Scored with
Drift sensor	DINO-I	ConvNeXt-I
DINOv2-B (Ours)	0.635	0.626
ConvNeXt-B	0.634	0.623

Appendix

Table 7: Drift sensor and evaluator ,

30

n=100

\fnum@figure : DINO-I over time ,

n=100

. Left:

30

s with each component. Right:

2

min.

	DINO-I $\uparrow$			$s_{\text{prior}}$ $\downarrow$
Method	first	middle	last	first	middle	last
Custom K/V only	0.579	0.506	0.475	0.395	0.414	0.418
+ ACG	0.595	0.551	0.550	0.380	0.385	0.378

Appendix

Table 8: Similarity to the reference images and to the class prior ,

30

n=100

\fnum@figure : User study interface. Participants see the prompt, the reference images and choose A or B for each question before moving to the next pair.

	Human		Qwen3-VL-235B		Claude Sonnet 5		Claude Opus 5.5
Custom Forcing vs.	Subject	Quality	Subject	Quality	Subject	Quality	Subject	Quality
Custom K/V only	60.0	45.8	60.0	50.0	40.0	35.0	80.0	60.0
FramePack	60.0	70.0	50.0	45.0	45.0	50.0	55.0	65.0
FramePack-F1	62.5	57.5	65.0	70.0	80.0	75.0	95.0	85.0
SkyReels-V2 DF	64.2	59.2	75.0	80.0	80.0	75.0	90.0	85.0
SkyReels-V3	45.8	51.7	80.0	90.0	70.0	80.0	60.0	80.0

Appendix

Table 9: User study and VLM judges. Percentage of judgments that prefer Custom Forcing over each method;

50\%

means no preference. Human:

120

votes per cell from

30

participants. VLM judges: Qwen3-VL-235B, Claude Sonnet 5, and Claude Opus 5.5, each with one judgment on each of

20

pairs per cell and

100

in All .

Table 19

\fnum@figure : Number of reference anchors at

N=9

30

n=100

. Reference images are repeated in the shaded region. Left: Subject Consistency and Dynamic Degree, with the two axes aligned at the default

|\mathcal{R}|=5

. Right: DINO-I.

		all	first	middle	last	Dynamic
		DINO-I
R5K4	Custom K/V only	0.541	0.579	0.506	0.475	0.41
R5K4	Custom Forcing (Ours)	0.635	0.619	0.626	0.616	0.41
G5K4	Custom K/V only	0.571	0.601	0.550	0.523	0.34
G5K4	Custom Forcing (Ours)	0.630	0.628	0.628	0.606	0.30
G8K1	Custom K/V only	0.556	0.577	0.538	0.512	0.40

Appendix

Table 12: What the custom K/V holds ,

30

n=100

. R, G, and K count reference images, generated images, and copies of the custom anchor. Bold marks the best value in each column, and Dynamic Degree is not bolded.

	DINO-I	CLIP-I	CLIP-T	Subj. Cons.	Bg. Cons.	Dynamic
Custom Forcing (Ours)	0.635	0.803	0.353	0.928	0.945	0.41
$b\equiv 0.05$	0.651	0.804	0.335	0.887	0.926	0.28
$b\equiv 0.1$	0.403	0.729	0.329	0.793	0.875	0.08
$b\equiv 0.2$	0.158	0.643	0.309	0.691	0.823	0.02
no subject mask	0.622	0.803	0.352	0.913	0.936	0.30

Appendix

Table 13: Adaptive amplification and subject mask ,

30

n=100

. The largest constant,

0.2

, equals the cap

b_{\max}

. Bold marks the best value in each column, and Dynamic Degree is not bolded.

\fnum@figure : A constant amplification raises DINO-I but breaks the scene. Frames at $17.5$ s of two prompts, whose DINO-I rises from $0.612$ to $0.681$ and from $0.482$ to $0.530$ .

Slug	Prompt
backpack (“backpack”)
00_park_bench	A cinematic video of a backpack resting on a weathered wooden park bench in an autumn park, orange leaves turning overhead and drifting past behind it.
01_forest_rock	A cinematic video of a backpack sitting on a flat mossy rock beside a winding forest trail, dappled sunlight flickering through the canopy and ferns swaying at the edges.
02_city_sidewalk	A cinematic video of a backpack standing upright on a quiet city sidewalk lined with shopfronts, traffic drifting past far behind it and shop awnings rippling in the breeze.
03_beach	A cinematic video of a backpack resting on flat open sand at a wide beach, fine sand streaming past it and waves breaking far behind.
04_mountain_ridge	A cinematic video of a backpack sitting on a rocky mountain ridge above a deep valley, wind tugging at its straps and clouds drifting across the valley below.

Appendix

Table 14: Evaluation prompts , grouped by subject. The noun used in the prompts is in parentheses.

Explore similar work

From Scores to Samples: Elastic Forcing for Autoregressive Video Generation
Sep 28, 2026Chi Zhang, Yueyi Liu, Shi Haoyang +5Autoregressive Video Generation Distribution Matching Distillation
College of AI, Tsinghua University · IAIR, Xi’an Jiaotong University · Xianghui Academy, Fudan University +1
Video-Mirai: Autoregressive Video Diffusion Models Need Foresight
Jun 2, 2026Yonghao Yu, Lang Huang, Runyi Li +2Autoregressive Video Diffusion Models Diffusion Models
The University of Tokyo · National Institute of Informatics · Peking University
Head Forcing: Long Autoregressive Video Generation via Head Heterogeneity
May 14, 2026Jiahao Tian, Yiwei Wang, Gang Yu +1Autoregressive Video Diffusion Models Autoregressive Video Generation
AGI Lab, Westlake University · University of California at Merced · StepFun

Sep 28, 2026, cs.CVJ/K move · Enter open · S save

From Scores to Samples: Elastic Forcing for Autoregressive Video Generation

Chi Zhang, Yueyi Liu, Shi Haoyang, Ruichuan An+4

College of AI, Tsinghua University · IAIR, Xi’an Jiaotong University · Xianghui Academy, Fudan University+1

Autoregressive Video Generation Distribution Matching Distillation Scores

Open paper AskPDF arXiv

Few-step autoregressive video generation commonly relies on Distribution Matching Distillation (DMD), requiring a bidirectional diffusion teacher and an online fake-score model. We instead learn the rollout distribution directly from reference videos, eliminating both score models during post-training. Our framework minimizes maximum mean discrepancy (MMD) in frozen self-supervised video representation spaces, using a hybrid Nyström--Monte Carlo estimator to balance approximation bias and sampling variance. Memory-efficient replay and gradient subsampling make this objective practical. Using the same architecture and initialization as Self-Forcing, our 1.3B model improves the VBench Total score from 83.80 to 84.64 while retaining 17 FPS. Removing auxiliary score models also enables 14B post-training on eight H200 GPUs. Beyond distillation, learning from reference videos enables the acquisition of new visual styles, semantic concepts, and spatial priors without a target-specific diffusion teacher.

Abstract

Explore similar work

From Scores to Samples: Elastic Forcing for Autoregressive Video Generation

Video-Mirai: Autoregressive Video Diffusion Models Need Foresight

Head Forcing: Long Autoregressive Video Generation via Head Heterogeneity