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

1 INTRODUCTION

• Text-only autoregressive generation produces generic category instances, while existing customization approaches require per-subject optimization or joint bidirectional processing unsuitable for causal streaming.Causal image-to-video models lose the subject within 30 seconds, and a bidirectional model cannot generate long videos.Source: §1 · p. 2
• Fixed anchors establish subject identity but cannot prevent drift, because weak influence loses subject details, strong influence suppresses motion, and prompts favor generic subjects.Appearance errors in finalized chunks propagate to later chunks, causing progressive divergence from the reference.Source: §1 · p. 2
• Custom Forcing keeps DINO-I between 0.58 and 0.62 over two-minute rollouts, while fixed anchors fall from 0.58 to 0.42.The improvement reaches +0.163 DINO-I in the final window without reducing motion; it also exceeds bidirectional customization methods and matches a 14B reference-to-video model with a 1.3B backbone while generating each frame 9.5–28.5× faster.Source: Figure 2 · p. 2, §1 · pp. 2–3
• Custom Forcing modifies only the frozen backbone’s self-attention, storing reference images and a custom anchor in the KV cache without fine-tuning or an added conditioning network.DVA scales reference influence with measured drift, while ACG extrapolates away from attention output without anchors to counter the prompt-induced class prior.Source: §1 · pp. 2–3

2.1 STREAMING AUTOREGRESSIVE VIDEO DIFFUSION

• A causal video is divided into latent chunks, and its distribution is factorized over chunks conditioned on the text prompt and previous outputs.Within each transformer block, current-chunk tokens attend to cached tokens and tokens in the current chunk.Source: §2.1 · p. 3
• Each chunk is generated from a few-step diffusion model, and its finalized keys and values enter the KV cache; completed chunks are never revised.The cache is bounded as a video grows, using persistent attention sinks and a window of recent frames from which older frames are evicted.Source: §2.1 · p. 3

2.2 SUBJECT CUSTOMIZATION IN AUTOREGRESSIVE DIFFUSION

• The customization goal is to generate the prompted scene and action with a subject matching reference images throughout the rollout, while keeping the generator frozen.Identity drift is declining similarity to the references; corrections must be made online before the next chunk is generated.Source: §2.2 · p. 3
• Custom Forcing writes reference images and a custom anchor into the cache, uses DVA to set reference strength, and uses ACG to steer away from the generic subject.DVA rescales reference values; ACG modifies attention output.Source: Figure 3 · p. 4

3.1 OVERVIEW

• Custom Forcing uses persistent-cache anchors, DVA to adjust reference strength, and ACG to steer attention away from the generic subject prior.The method modifies the backbone’s self-attention rather than adding a conditioning network.Source: §3.1 · p. 4

3.2 CUSTOM K/V CONSTRUCTION

• Reference images and a custom opening-scene anchor occupy the cache; more custom-anchor copies favor scene stability, while more reference anchors restore motion.Reference anchors specify appearance, and the custom anchor specifies the opening scene.Source: §3.2 · pp. 4–5

3.3 DRIFT-ADAPTIVE VALUE AMPLIFICATION

• DVA measures drift from finalized chunks and amplifies reference-anchor values according to similarity loss, restricting amplification to subject tokens.It compares similarity with a level estimated from the same video and allocates more weight to tokens that are mostly subject.Source: §3.3 · p. 5

3.4 ANCHOR CONTRAST GUIDANCE

• ACG compares attention outputs with and without anchor tokens, then extrapolates their difference to strengthen anchor contributions against the class prior.The anchor-free output still reflects anchors indirectly through generated frames produced under their influence.Source: §3.4 · p. 6

4.1 EXPERIMENTAL SETUP

• The setup uses 10 DreamBooth subjects, 10 prompts each, and one 30 s and one 2 min video per prompt; the custom K/V contains five reference anchors and four copies of a generated custom anchor.The evaluation reports identity, prompt-alignment, consistency, and motion metrics, averaged over 100 prompts.Source: §4.1 · pp. 6–7

4.2 MAIN RESULTS

• Custom Forcing stays nearly flat in DINO-I over 30 s (0.619 to 0.616) and declines only to 0.578 over 2 min, while Dynamic Degree matches Custom K/V only.Custom K/V only falls from 0.579 to 0.475 within 30 s and from 0.583 to 0.415 over 2 min.Source: §4.2 · p. 7, Table 1 · p. 7
• With a 1.3B backbone, Custom Forcing matches 14B SkyReels-V3 in whole-video DINO-I (0.635 against 0.641) and exceeds it in the last window (0.616 against 0.564).It generates each frame 9.4–28.5× faster than the four compared models; 30 participants preferred it over every included method except SkyReels-V3, where the difference was not significant.Source: §4.2 · pp. 7–8, Table 2 · p. 8
• On one H100 GPU, Custom Forcing takes 1.13× Custom K/V only’s generation time at 30 s and 1.11× at 2 min; ACG alone adds 4%.Most overhead comes from measuring drift.Source: §4.2 · p. 8

4.3 ABLATION STUDY

• Adding DVA keeps DINO-I near 0.60 from the first window to the last, and adding ACG raises each window by a further 0.02–0.03.Qualitatively, DVA delays identity changes, while adding ACG keeps both subjects as in the reference images.Source: §4.3 · p. 8, Table 3 · p. 8, Figure 6 · p. 9

5 CONCLUSION

• Custom Forcing customizes autoregressive video generation through the KV cache without per-subject optimization or an additional conditioning network, keeping identity close to references for two minutes without reducing motion.It holds DINO-I between 0.58 and 0.62 while fixed anchors fall to 0.42, and adds 11–13% to fixed-anchor generation time.Source: §5 · p. 9

A.1 AUTOREGRESSIVE VIDEO GENERATION

• Autoregressive models generate frames or chunks in temporal order and can stream outputs, while bidirectional models denoise all frames jointly; prior long-rollout methods maintain consistency with generated context, whereas Custom Forcing targets external references.The related-work discussion also describes methods using sinks, persistent memory, longer contexts, and spatial memory.Source: §A.1 · p. 16

A.2 SUBJECT-DRIVEN VIDEO CUSTOMIZATION

• Video customization includes subject-specific optimization and reusable conditioning systems, but the cited methods primarily target bounded, jointly denoised videos rather than open-ended causal rollouts.V-Warper injects reference appearance through value representations but uses image-domain adaptation and bidirectional video denoising; Custom Forcing stores references in a streaming model’s persistent cache.Source: §A.2 · pp. 16–17

A.3 INFERENCE-TIME ATTENTION CONTROL

• ACG masks reference frames to form its negative attention branch, scales the attention share queries already assign to references, and requires a second attention read rather than an additional forward pass.DVA adjusts reference values during sampling using an online subject-similarity measurement.Source: §A.3 · p. 17

B IMPLEMENTATION DETAILS

• Rolling Forcing generates 832 × 480 video at 16 fps in chunks of three latent frames, with queries attending to nine anchor, three recent, and 15 denoised frames.The custom K/V stores five DreamBooth reference images followed by four copies of the custom anchor; αinj = 300 and κ = 4 are used.Source: B IMPLEMENTATION DETAILS · p. 17
• No hyperparameter is tuned per subject or prompt; only b(t) adapts to each video, and the ACG blend weight warms up over five chunks after its base value is set.The implementation keeps κ fixed during the warm-up.Source: B IMPLEMENTATION DETAILS · p. 17, Table 4 · p. 18
• The custom anchor is one 832 × 480 image generated by FLUX.2 [klein] 4B from five reference images and an image prompt adapted from the video prompt.The generated image is used without editing and with seed 0.Source: B IMPLEMENTATION DETAILS · pp. 17–18
• The bidirectional baselines use distinct training recipes, including 200-epoch CustomCrafter training, 2,000-step SMRABooth training, and 300-step MotionBooth training using UCF-101 clips in place of unavailable Panda-70M clips.The autoregressive LoRA baseline fine-tunes Rolling Forcing with a new subject token and adapters on attention projections.Source: B IMPLEMENTATION DETAILS · pp. 18–19
• The image- and reference-to-video baselines use their released checkpoints and specified native sampling settings, while SkyReels-V3 uses four reference images and extends an initial 5 s generation to 32.7 s.Custom Forcing receives the same custom anchor and all five reference images; generation time is averaged per frame over 100 videos on one H100 GPU.Source: B IMPLEMENTATION DETAILS · p. 19

C EVALUATION DETAILS

• DINO-I and CLIP-I measure similarity to reference images, while CLIP-T measures agreement with the prompt using normalized features from every eighth video frame.DINO-I over time is computed across three 10.5 s windows for short videos.Source: C EVALUATION DETAILS · p. 19
• VBench Subject and Background Consistency measure temporal stability, while Dynamic Degree measures motion, allowing the evaluation to examine stability alongside motion.These scores are computed on each video at its native frame rate.Source: C EVALUATION DETAILS · p. 19

D.1 MAIN RESULTS IN DETAIL

• CUSTOM FORCING keeps both subjects for two minutes while Custom K/V only shows drift within the first minute; its CLIP-T stays within 0.007 of the 30-second maximum.The bidirectional methods remain below Custom K/V only in CLIP-I and CLIP-T.Source: §D.1 · p. 19

D.2 DEEP FORCING

• On Deep Forcing, DINO-I rises from 0.556 to 0.641 at 30 s and from 0.518 to 0.620 at 2 min; ACG on top of DVA gives the highest value in every window.DVA alone lowers Dynamic Degree from 0.81 to 0.62, while adding ACG restores it; CUSTOM FORCING takes 1.30× the time of Custom K/V only at 30 s and 1.26× at 2 min.Source: §D.2 · p. 20, Table 5 · p. 20
• With ACG warm-up on Deep Forcing, the subject moves more while preserving identity; without warm-up, generation remains near the anchor pose with limited motion.Figure 7 shows Custom Forcing preserving distinctive subject appearance through 30-second and two-minute Deep Forcing videos.Source: Figure 7 · p. 20, Figure 8 · p. 21

D.3 RESULTS PER SUBJECT

• CUSTOM FORCING raises last-window DINO-I over Custom K/V only for every subject, by 0.142 on average and in 88 of 100 prompts.The gain is largest for animals, which show the most identity drift under Custom K/V only, and smallest for the backpack and RC car.Source: §D.3 · p. 21, Table 6 · p. 21

D.4 DRIFT SENSOR

• Replacing the DINOv2-B drift sensor with ConvNeXt-B changes identity scores by at most 0.003 across the two evaluators.Source: §D.4 · p. 21, Table 7 · p. 21

D.5 IDENTITY OVER TIME

• Over time, Custom K/V only declines from 0.61 to 0.44 over 30 s and from 0.59 to 0.41 over 2 min, while CUSTOM FORCING stays near 0.60 at 30 s and ends at 0.58 at 2 min.DVA stops the decline after about 10 s, whereas ACG alone starts highest but still declines.Source: §D.5 · pp. 21–22, Figure 9 · p. 22

D.6 CLASS PRIOR

• Adding ACG lowers similarity to the vanilla AR class prior by 0.015 to 0.040 in every third and keeps it near 0.38, while DINO-I rises by 0.016 to 0.076.Source: §D.6 · p. 22, Table 8 · p. 22

D.7 USER STUDY AND VLM JUDGES

• Participants prefer CUSTOM FORCING for subject similarity over four baselines in 60.0–64.2% of judgments, but show no significant preference against SkyReels-V3.They prefer it for quality over FramePack (70.0%) and overall (56.8%); three VLM judges favor it overall for subject and quality, but differ from participants on SkyReels-V3 and Custom K/V only.Source: §D.7 · p. 23, Table 9 · p. 23

E QUALITATIVE OBSERVATIONS

• Across two-minute examples, Custom K/V-only changes the toy’s eyes and body and turns a gray merle dog black, while CUSTOM FORCING retains both subjects’ appearance.Source: E QUALITATIVE OBSERVATIONS · p. 24
• FramePack and SkyReels-V3 preserve animal fur through 30 s with ten times Custom Forcing’s parameters, while CUSTOM FORCING’s DINO-I stays between 0.616 and 0.626 across windows.The image-to-video models receive only the custom anchor; SkyReels-V3 sees reference images only in its first 5 s, unlike Custom Forcing’s persistent KV cache.Source: E QUALITATIVE OBSERVATIONS · p. 24
• DVA delays visible identity changes relative to Custom K/V only, and adding ACG keeps both the dog and toy as in the reference images.Source: E QUALITATIVE OBSERVATIONS · p. 24

F.1 INJECTION AND GUIDANCE STRENGTH

• The operating point balances identity and motion: increasing guidance beyond κ = 4 saturates DINO-I and lowers Dynamic Degree, while larger injection budgets reduce consistency and motion.At αinj = 420, last-third DINO-I is 0.041 below the first third, versus 0.003 at 300; κ = 6 lowers CLIP-I, Subject Consistency, and Dynamic Degree.Source: §F.1 · p. 24, Table 10 · p. 24, Table 11 · p. 24

F.2 COMPOSITION OF THE CUSTOM K/V

• At N = 9, |R| = 5 is the smallest reference-image count that reaches Custom K/V-only Dynamic Degree (0.41), while both consistency scores remain higher.At |R| = 1, Dynamic Degree is 0.14 and motion is limited; at |R| = 8, consistency falls below Custom K/V only as the scene drifts.Source: §F.2 · p. 24, Figure 11 · p. 25

F.3 GENERATED IMAGES IN PLACE OF THE REFERENCE IMAGES

• Replacing reference images with generated images lowers Dynamic Degree from 0.41 to 0.30 in G5K4, despite DINO-I only 0.005 below the default; G8K1 is 0.020 lower in DINO-I.DVA and ACG raise DINO-I most in the last window in every setting.Source: §F.3 · p. 25, Table 12 · p. 25

F.4 CONSTANT AMPLIFICATION AND SUBJECT MASK

• A constant b ≡ 0.05 raises DINO-I by 0.016 but makes the subject fill the frame; removing the subject mask also lowers DINO-I, Subject Consistency, and Dynamic Degree.Measured b(t) averages 0.022 and exceeds 0.05 in only 14% of chunks; amplifying background tokens ties the video to reference-image backgrounds.Source: §F.4 · p. 25, Table 13 · p. 26, Figure 12 · p. 26

• Identity control operates on the subject as a whole, so changes limited to a small part may barely affect the drift signal and cannot be targeted independently.The passage gives a wheel or backpack strap as examples.Source: G LIMITATIONS · p. 26
• Appearance the frozen backbone cannot render remains out of reach because Custom Forcing changes only how it attends to anchors.Source: G LIMITATIONS · p. 26
• Evaluation covers one subject per video for up to two minutes; multiple subjects and longer horizons remain future work.Source: G LIMITATIONS · p. 26

• 100 evaluation prompts contain 70 distinct sentences, and every prompt names the subject only by its category.The table groups prompts by subject, with the subject noun shown in parentheses.Source: H EVALUATION PROMPTS · p. 26, Table 14 · p. 27
• Cat and cat2, dog and dog3, and dog7 and dog8 each share the same ten prompts.Source: H EVALUATION PROMPTS · p. 26
• The ten robot-toy prompts describe five scenes twice, once with the toy moving and once holding still.Source: H EVALUATION PROMPTS · p. 26