acceptodds
Under review as a conference paper at ICLR 2027

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

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) increases reference influence according to 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.

Primary area:
generative models
Keywords:
video generation, autoregressive video model, subject customization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GZgZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents Custom Forcing, a training-free method for subject customization in autoregressive video generation. It stores reference images and a scene-specific image generated from them as persistent anchors in the KV cache of a frozen video model. To reduce identity drift, drift-adaptive value amplification adjusts the influence of reference features based on the similarity of generated frames to the reference images. Anchor contrast guidance further strengthens reference conditioning by contrasting attention outputs with and without the anchors. Experiments on two autoregressive backbones show improved identity preservation in videos up to two minutes, with comparable motion metrics and moderate inference overhead relative to fixed-anchor conditioning.

Critical strengths
  1. The paper addresses a practical problem: preserving a specific subject’s identity throughout long autoregressive videos without per-subject fine-tuning.
  2. The paper is clearly written and easy to follow.
Critical weaknesses
  1. The technical novelty is limited. Generating a video keyframe from reference images using an image generation model and inserting it into the autoregressive model’s KV cache is relatively common, as in FashionChameleon [1]. ACG mainly consists of attention extrapolation, L1-norm clipping, and blending, closely resembling NAG [2]. The main change is replacing positive/negative text conditioning with retaining/masking reference anchors. DVA is an adaptation of V-Warper [3].

  2. Appendix F.4 reports an average adaptive drift signal of approximately 0.022, while the constant baselines use 0.05, 0.1, and 0.2, corresponding to typical target gains of approximately 1.3 versus at least 3. Thus, the motion degradation of the constant baselines may primarily result from excessive injection strength. This strength mismatch leaves open whether a weaker constant gain or a predefined temporal schedule could achieve similar results, and does not sufficiently establish the necessity of online drift feedback.

  3. Comparisons with existing methods are missing. FashionChameleon also performs customized content generation within an autoregressive video framework. I understand that training-free methods cannot be directly compared with training-based methods, but I am interested in how large the performance gap is. Additionally, I would like to see results using human subjects to provide a more intuitive assessment of identity preservation.

[1] FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization [2] Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models [3] V-Warper: Appearance-Consistent Video Diffusion Personalization via Value Warping

Questions for authors

See Critical Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer GZgZ (1/2): novelty, constant and schedule

Thank you for the careful review and for pointing us to these works. All new results are in the revised PDF (changes in blue). Tables R1–R5 appear in these comments; all other table, figure, section, and appendix numbers refer to the revised PDF. We answer the first two weaknesses here and the third in the next comment.

W1. Novelty. Custom Forcing builds on existing operations, and we credit them. Our contribution is making a frozen autoregressive video model, with its weights unchanged, follow a user-given subject through inference-time control of its KV cache and attention, without customization training; none of the cited methods operates in this setting (Table R1), and each component is designed for it.

  • Custom anchor vs. FashionChameleon. FashionChameleon does not generate a keyframe with an image generator: it encodes the given reference and garment images with the VAE into a model trained for this conditioning, and refreshes the reference KV from the last frame when switching garments. Custom Forcing uses an image generator to place the subject in the prompt's scene before generation and stores this anchor with the real photos in the KV cache, so a model never trained on reference conditioning starts from the subject in the right scene.
  • ACG vs. NAG. ACG uses NAG's extrapolation, L1-norm clipping, and blending, which we credit (Sec. 3.4, App. A.3); we chose them because our backbone is a few-step distilled model. What differs is what is contrasted: two self-attention outputs with the cached reference anchors kept or masked, instead of positive and negative text prompts. This isolates the anchors' contribution without a negative prompt or a change to the text conditioning; as in PAG (Ahn et al., 2024) and SAG (Hong et al., 2023), which derive the guidance branch from self-attention rather than introducing a new guidance operator, the design lies in the branches contrasted.
  • DVA vs. V-Warper. V-Warper trains a per-subject LoRA and subject token and warps reference values to token correspondences within a short bidirectional clip. DVA uses neither: it measures identity drift during the causal rollout and amplifies the cached reference values only when drift appears.

Table R1. Relation to the cited works.

FashionChameleonNAGV-WarperCustom Forcing
Frozen video model (no customization training)✗–✗ (per-subject LoRA)✓ (as released)
Anchor generated by an image model✗–✗✓
Guidance branches–Positive / negative text–Anchor kept / masked
Value manipulation––Correspondence warpingDrift-gated amplification
Online drift feedback✗✗✗✓
Causal long-video rollout✓–✗✓

W2. Weaker constant and predefined schedule. We ran these controls with all other settings unchanged (App. F.4, Table 13, Fig. 13): a constant b ≡ 0.022 matched to the mean b(t) of DVA, a weaker constant b ≡ 0.01, and a linear schedule with the same mean as DVA. b ≡ 0.022 and the schedule sit near DVA on automated metrics, so we acknowledge them as competitive baselines, while b ≡ 0.01 lowers DINO-I (Table R2). They differ in when they amplify. Please see Fig. 13 in App. F.4 of the revised PDF, which illustrates this distinction. Where little drift is measured, DVA keeps b(t) near zero, while b ≡ 0.022 and the schedule keep amplifying and can enlarge or deform the subject (red boxes), even when their DINO-I is higher (Fig. 13(b)). In a 30 s study, 30 participants ranked the videos of b ≡ 0.022, the schedule, and Custom Forcing for 20 prompts, with ties allowed; Custom Forcing was ranked better than each control more often than worse, for both identity and quality (Holm-adjusted p ≤ 0.009 in all four comparisons; App. D.7, Table 9(d); Table R3).

Table R2. Automated metrics, 30 s, n = 100 (subset of Table 13).

DINO-ICLIP-ICLIP-TSubj. Cons.Bg. Cons.Dynamic Degree
b ≡ 0.010.5760.7890.3580.9240.9440.44
b ≡ 0.0220.6430.8050.3530.9300.9470.37
Linear schedule0.6290.8020.3550.9320.9480.41
DVA (ours)0.6350.8030.3530.9280.9450.41

Table R3. Human ranking, 30 s, 30 participants, 20 prompts (Table 9(d)). Share of answers (%) in which Custom Forcing was ranked better than, the same as, or worse than each control; p: mixed-effects logistic regression on the untied answers, Holm-adjusted within each question.

Custom Forcing vs.Identity: better / same / worsepQuality: better / same / worsep
b ≡ 0.02259.7 / 26.8 / 13.50.000660.3 / 24.3 / 15.30.0001
Linear schedule50.0 / 34.3 / 15.70.008950.3 / 31.3 / 18.30.0057
AuthorsOpenReview ↗

Response to Reviewer GZgZ (2/2): comparison with existing methods, human subjects

W3. Comparison with existing methods and human subjects.

  • FashionChameleon. We agree it is closely related and would have liked to compare directly, but its model weights have not been released. Since you are interested in the gap to training-based methods, we compared Custom Forcing with four models that learn a reference-conditioning module, VACE, MAGREF, Phantom, and 1% Compute, at a common 5.1 s window (App. H.3, Table 17; Table R4). Custom Forcing has the highest DINO-I, equal or higher CLIP-I and prompt adherence, and perceptual quality within 0.041, but it moves less (Dynamic Degree 0.28 vs. 0.46–0.71). We now cite FashionChameleon in App. A.2.

Table R4. Common 5.1 s window, means over 100 videos (Table 17).

MethodDINO-ICLIP-ICLIP-TDynamicAesth.Imaging
VACE-1.3B0.3330.7030.2960.460.6310.673
MAGREF-14B0.4980.7860.3380.650.6080.693
Phantom-1.3B0.5180.7740.3500.710.6570.685
1% Compute0.5940.7690.3320.600.5470.599
Custom Forcing0.6140.8010.3550.280.6160.676
  • Human subjects (App. H.4, Table 19, Fig. 14; Table R5). We ran the 30 identities of the ConsisID evaluation set (five photos each) with four prompts of our own per identity, at 30 s and 2 min, with all hyperparameters unchanged; the photos are cropped around the face, and DVA uses head masks. Custom Forcing raises face similarity over Custom K/V only by 0.062 at 30 s and 0.074 at 2 min (FaceSim-Arc), with slightly lower FID and comparable CLIPScore, and stays nearly constant as videos get longer (0.204 → 0.192 vs. 0.142 → 0.118). ConsisID, a CogVideoX-5B model fine-tuned on human videos for face identity, reaches much higher face similarity on its 6.1 s clips (0.459 vs. 0.228), so we do not claim parity with face-specialized models (App. G). Custom Forcing is training-free and applies to any subject; its goal is to limit identity drift over long rollouts. Fig. 14 shows three identities over 30 s.

Table R5. Human subjects, official ConsisID metrics (Table 19). ConsisID, by design, takes a single photo (we use image 00); ours takes all five, and our 6.1 s rows are the first 6.1 s of the 30 s videos.

MethodSpanFaceSim-Arc ↑FaceSim-Cur ↑FID ↓CLIPScore ↑
ConsisID6.1 s0.4590.494173.8031.66
Custom K/V only6.1 s0.1910.203161.5131.69
Custom Forcing6.1 s0.2280.241160.6231.82
Custom K/V only30 s0.1420.152164.1731.58
Custom Forcing30 s0.2040.216163.9731.56
Custom K/V only2 min0.1180.127167.7730.85
Custom Forcing2 min0.1920.201167.7131.24

Thank you for the detailed response. The additional experiments address most of my concerns in W2 and W3. However, the contribution remains incremental, and, as the authors acknowledge, performance on human subjects is still limited, which restricts its practical usefulness. On balance, I am raising my recommendation from weak reject to weak accept.

AuthorsOpenReview ↗

Thank you for your follow-up and updated assessment

Thank you for reviewing our additional results and for raising your score. We are glad that the new experiments addressed your concerns in W2 and W3. Your feedback helped us strengthen our evaluation and clarify our contribution.

We also appreciate your remaining points. We agree that performance on human subjects is still a limitation, and we will discuss it more clearly in the limitations section of the final version as an important direction for future work. We will also make the contribution of Custom Forcing more explicit, as a training-free approach to subject customization for autoregressive video generation.

Thank you again for your time and thoughtful review. Please let us know if you have any further questions before the discussion period ends.

Reviewer_qgFVOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Custom Forcing preserves a reference subject in causal video diffusion. Five reference images and four copies of a scene-specific image generated by FLUX.2 [klein] occupy persistent KV-cache slots. Drift-adaptive value amplification (DVA) adjusts masked reference values using online similarity; anchor contrast guidance (ACG) extrapolates between attention outputs with and without anchors. The video backbone remains frozen, with auxiliary image generation, segmentation, and visual encoding models.

Experiments cover ten DreamBooth subjects (two cats, four dogs, a backpack, a monster toy, an RC car, and a robot toy), 100 subject-prompt combinations, and approximately 30-second and two-minute videos. Comparisons include fixed anchors, component ablations, customization methods, and image/reference-to-video systems. Rolling Forcing and Deep Forcing show improved identity retention; additional evidence includes sensor substitution and a 30-participant study.

Critical strengths

(1) A substantial, controlled improvement on the central task. Against the same backbone and custom anchors, Table 1 (p. 7) improves last-window DINO-I from 0.475 to 0.616 at 30 seconds and from 0.415 to 0.578 at two minutes. This directly supports identity retention beyond initial conditioning. Table 6 (p. 21) reports improvement for every subject and 88/100 prompts, with a subject-bootstrap interval of [0.093, 0.189] for the mean final-window gain. The principal result therefore has support beyond selected examples and cross-model comparisons.

(2) Useful corroboration and practical integration. Table 5 (p. 20) repeats the effect on Deep Forcing without changing the method's constants. Table 7 (p. 21) replaces the drift sensor with ConvNeXt-B and retains essentially the same DINO-I, addressing the shared-sensor/evaluator concern. The human study reports 60% identity preference over fixed anchors (Table 9, p. 23). These complementary checks, together with the reported 11-13% overhead on Rolling Forcing (p. 8), support the value of a method requiring no subject-specific optimization.

Critical weaknesses

(1) The constant-strength comparison does not isolate the value of adaptation. Appendix F.4 reports mean adaptive drift of 0.022, corresponding to mean target amplification about 1.3, but Table 13 tests only fixed drift values 0.05, 0.1, and 0.2, beginning at amplification 3 (pp. 25-26; Eq. (7)). Those controls apply substantially stronger conditioning. Their poor identity/motion tradeoff therefore does not establish that a suitably chosen constant cannot match DVA. Table 3 and Fig. 9 convincingly show that DVA helps relative to unmodified anchors, but a fixed-strength comparison near the observed operating range is needed to substantiate the specific contribution of online adaptation.

(2) The quality tradeoff against contemporary S2V methods and its generality remain unclear. The abstract and introduction claim superiority over bidirectional customization methods, but Table 1 covers VACE-1.3B, SMRABooth, and two older U-Net methods, MotionBooth and CustomCrafter. Phantom is cited but not evaluated; MAGREF and Learning Zero-Shot Subject-Driven Video Generation Using 1% Compute are neither cited nor evaluated. The latter also reports a Wan2.1-1.3B variant. Table 2's stronger long-video systems are useful, but do not resolve how this AR approach compares with these reference-conditioned S2V alternatives. Moreover, Table 1 compares native 2-5.1 s bidirectional outputs against 30 s causal outputs without a common-duration quality evaluation. A comparison should establish how closely AR generation approaches contemporary S2V identity, prompt adherence, motion, and perceptual quality, alongside its streaming and long-rollout benefits. Uniformly outperforming larger bidirectional models is not necessary for the contribution to be valuable; measuring the quality tradeoff is.

(3) Coverage is also limited to ten DreamBooth identities, 70 distinct prompt sentences, and one seed per subject-prompt pair (pp. 17, 21, 26). Cats and dogs provide articulated subjects, but no human-face results or targeted evaluation under substantial deformation, expression changes, or occlusion is reported. Appendix G acknowledges that whole-subject drift can miss local changes. The applicable single-subject categories of OpenS2V-Eval, including separate identity and naturalness measures, would provide useful external validation.

(4) Motion preservation is also insufficiently resolved: equal Dynamic Degree at 30 seconds (0.41 versus 0.41) measures thresholded motion detections, not equal subject-action quality (official implementation). The paper itself shows that scene drift can increase this metric (F.2). The 45.8% human quality preference over fixed anchors is described as nonsignificant without a reported test or interval, and covers only 30-second videos (D.7). These limitations weaken the generality of the claimed identity/motion tradeoff, while leaving the controlled improvement over fixed anchors intact.

Questions for authors

(1) Decision-relevant evidence request: With the same masks, ACG, initialization, and EMA, how does DVA compare with constant drift values around the observed mean of 0.022, or a constant matched to its mean amplification? Would the identity/motion advantage remain?

(2) Decision-relevant clarification: What statistical analysis supports the preference and nonsignificance statements in D.7, accounting for repeated participants and prompts? Please provide the uncertainty for the quality comparison with fixed anchors and clarify whether the motion-preservation conclusion should be limited to the reported metrics and duration.

(3) Decision-relevant comparison and positioning request: How close is Custom Forcing to contemporary reference-conditioned S2V methods at a common short duration? Please extend the existing VACE comparison to Phantom, MAGREF, and Learning Zero-Shot Subject-Driven Video Generation Using 1% Compute, or explain concrete incompatibilities and limit the superiority claim accordingly. Prioritize comparable-scale variants where available, report model size and reference-image access (including the generated custom anchor), and assess identity, prompt adherence, motion, and perceptual quality. Streaming efficiency and long-rollout drift should remain complementary evaluation axes. Please also cite and discuss the currently omitted MAGREF and 1% Compute works; in particular, distinguish shared learned reference conditioning with no per-subject test-time tuning from your frozen-backbone inference-time control without a separate customization-training stage. This comparison would materially clarify the quality cost or benefit of obtaining causal customization.

(4) Decision-relevant generalization request: What evidence supports generalization beyond the ten DreamBooth identities? Please provide results on applicable OpenS2V-Eval single-subject categories, or explain the benchmark mismatch and bound the claim accordingly. In particular, is human facial identity supported, and what evidence separates true identity loss from legitimate large pose/deformation changes when driving DVA? The concern is broader coverage and subject-level fidelity, not the absence of all articulated examples.

6 comments
AuthorsOpenReview ↗

Response to Reviewer qgFV (1/4): Q1 constants and schedule

Thank you for the detailed and constructive review. All new results are in the revised PDF (changes in blue). Tables R1–R11 appear in these comments; all other table, figure, section, and appendix numbers refer to the revised PDF. We answer Q1 here and Q2–Q4 in the next three comments.

Q1/W1. Constants around 0.022 and a schedule. With the same masks, ACG, initialization, and EMA, we replaced b(t) with a constant matched to the observed mean (b ≡ 0.022), a weaker constant (b ≡ 0.01), and a linear schedule with the same mean (App. F.4, Table 13, Fig. 13; Table R1). b ≡ 0.022 and the schedule sit near DVA on automated metrics, so we acknowledge them as competitive baselines; b ≡ 0.01 lowers DINO-I. They differ from DVA in when they amplify. Please see Fig. 13 in App. F.4 of the revised PDF, which illustrates this distinction. Where little drift is measured, DVA keeps b(t) near zero, while b ≡ 0.022 and the schedule keep amplifying and can enlarge or deform the subject (red boxes), even when their DINO-I is higher (Fig. 13(b)). Where drift is measured, DVA raises b(t) and all three stay plausible (Fig. 13(a)). In a 30 s study, 30 participants ranked the videos of b ≡ 0.022, the schedule, and Custom Forcing for 20 prompts, with ties allowed. Custom Forcing was ranked better than each control more often than worse, for both identity and quality (Holm-adjusted p ≤ 0.009 in all four comparisons; App. D.7, Table 9(d); Table R2). The advantage of adaptive feedback therefore appears in human judgments, not as a higher average identity score.

Table R1. Automated metrics, 30 s, n = 100 (subset of Table 13).

DINO-ICLIP-ICLIP-TSubj. Cons.Bg. Cons.Dynamic Degree
b ≡ 0.010.5760.7890.3580.9240.9440.44
b ≡ 0.0220.6430.8050.3530.9300.9470.37
Linear schedule0.6290.8020.3550.9320.9480.41
DVA (ours)0.6350.8030.3530.9280.9450.41

Table R2. Human ranking, 30 s, 30 participants, 20 prompts (Table 9(d)). Share of answers (%) in which Custom Forcing was ranked better than, the same as, or worse than each control; p: mixed-effects logistic regression on the untied answers, Holm-adjusted within each question.

Custom Forcing vs.Identity: better / same / worsepQuality: better / same / worsep
b ≡ 0.02259.7 / 26.8 / 13.50.000660.3 / 24.3 / 15.30.0001
Linear schedule50.0 / 34.3 / 15.70.008950.3 / 31.3 / 18.30.0057
AuthorsOpenReview ↗

Response to Reviewer qgFV (2/4): Q2 statistics and motion

Q2/W4. Statistical analysis and the scope of the motion claim.

  • Analysis. Each participant made 20 judgments and each prompt appears 30 times, so we analyze the 600 judgments of each question with a mixed-effects logistic regression (Laplace) with crossed random intercepts for participants and prompts and the screen side as a covariate, and apply Holm correction across the five comparisons (App. D.7, Table 9(a); Table R3).
  • Identity. Pooled over the five comparisons, participants prefer Custom Forcing (59.9% [51.1, 68.1], p = 0.027). Four of the five point the same way, including fixed anchors, but none is significant alone after Holm, so D.7 and Sec. 4.2 now report the pooled preference.
  • Quality. Against fixed anchors, the uncertainty you requested is 45.8% [33.5, 58.6], with no significant difference; the only significant quality difference favors Custom Forcing, over FramePack (72.9% [61.0, 82.2], p = 0.002).
  • Two minutes. In a new two-minute study (30 participants, 20 prompts, A or B with positions balanced, 600 answers per question, the same model with an A/B position term), participants chose Custom Forcing over Custom K/V only for identity in 73.7% of answers and for quality in 63.3% (App. D.7, Table 9(c); Table R4).
  • Motion. We agree the conclusion should be limited. Dynamic Degree counts whole-frame optical flow, so camera and scene motion also raise it, and it does not measure how well the subject performs the action (now stated in App. C). We removed "without reducing motion" from the abstract, Sec. 1, and Sec. 5, and report motion only as measured (Dynamic Degree 0.41 vs. 0.41 at 30 s, 0.45 vs. 0.47 at 2 min) together with the quality judgments above, whose question also asks raters to consider how natural the motion looks.

Table R3. 30 s pairwise study (Table 9(a)). Votes: share of 120 votes (600 in All) preferring Custom Forcing; Model: mixed-effects estimate [95% CI]; p: Holm-adjusted (All: one test).

Custom Forcing vs.Identity votesIdentity modelpQuality votesQuality modelp
Custom K/V only60.061.8 [49.4, 72.9]0.1545.845.8 [33.5, 58.6]1.00
FramePack60.062.6 [50.1, 73.6]0.1570.072.9 [61.0, 82.2]0.002
FramePack-F162.564.5 [52.1, 75.2]0.09057.559.0 [46.2, 70.8]0.50
SkyReels-V2 DF64.265.7 [53.4, 76.3]0.06759.260.5 [47.6, 72.0]0.44
SkyReels-V345.845.3 [33.3, 57.8]0.4651.751.2 [38.6, 63.7]1.00
All58.559.9 [51.1, 68.1]0.02756.857.9 [48.8, 66.5]0.090

Table R4. Two-minute study (Table 9(c)). Chosen: share of the 600 answers per question that chose Custom Forcing; Model: mixed-effects estimate [95% CI].

Custom Forcing vs.Identity chosenIdentity modelpQuality chosenQuality modelp
Custom K/V only73.780.1 [68.4, 88.2]< 0.000163.366.7 [50.4, 79.8]0.045
AuthorsOpenReview ↗

Response to Reviewer qgFV (3/4): Q3 contemporary S2V methods

Q3/W2. Contemporary S2V methods at a common short duration. We extended the VACE-1.3B comparison to Phantom-1.3B, MAGREF-14B (its only released size), and 1% Compute (its only released checkpoint, the first-version LoRA on CogVideoX-5B; the later CogVideoX and Wan models in its paper are not public). All use their official settings on the same 100 prompts with seed 0 and are evaluated on a common 5.1 s window; for our methods and the text-only backbone we use the first 5.1 s (81 frames) of the 30 s rollouts (App. H.3, Tables 16–18; Tables R5–R7). Table 1 now also includes these models, with a short-duration block for our methods, and Sec. 4.1–4.2 discuss them.

Table R5. Model size, reference access, and training (Table 16). Training-free: no customization-training stage; tuning-free: no per-subject optimization. Photos are DreamBooth images 00 (MAGREF, 1% Compute), 00–03 (Phantom), and 00–04 (VACE and ours); the anchor is generated per prompt by FLUX.2 [klein] 4B.

MethodBackbone (size)Reference inputTraining-freeTuning-free
VACEWan2.1 (1.3B)5 photos✗✓
MAGREFWan2.1 (14B)1 photo✗✓
PhantomWan2.1 (1.3B)4 photos✗✓
1% ComputeCogVideoX (5B)1 photo✗✓
Custom ForcingRolling Forcing (Wan2.1, 1.3B)5 photos + 1 generated anchor✓✓

Table R6. Common 5.1 s window, means over 100 videos (Table 17; Subject/Background Consistency, Motion Smoothness, and Flicker are also there).

MethodDINO-ICLIP-ICLIP-TDynamicAesth.Imaging
VACE-1.3B0.3330.7030.2960.460.6310.673
MAGREF-14B0.4980.7860.3380.650.6080.693
Phantom-1.3B0.5180.7740.3500.710.6570.685
1% Compute0.5940.7690.3320.600.5470.599
Text-only backbone0.1620.6780.3580.440.6120.691
Custom K/V only0.5940.8020.3540.310.6150.677
Custom Forcing0.6140.8010.3550.280.6160.676

Results. Custom Forcing has the highest DINO-I; against the S2V models, its CLIP-I and prompt adherence are equal or higher, and its perceptual quality is within 0.041. It moves less (Dynamic Degree 0.28 vs. 0.46–0.71); consistency, smoothness, and flicker favor such low-motion videos, so we do not count them as advantages. Part of the low motion comes from the causal backbone itself (0.44 on the same window) and part from the reference K/V (0.31); over the full 30 s, Custom Forcing is not less dynamic than the text-only backbone (0.41 vs. 0.36). The identity, in contrast, does not come from the backbone, which reaches only 0.162 DINO-I without references. 1% Compute comes closest in DINO-I (0.594), but its prompt adherence and perceptual quality are lower. We narrowed the abstract and Sec. 1 to identity higher than or comparable to same- or larger-size S2V models at a common short duration, and Sec. 1 notes that Custom Forcing moves less.

Efficiency. We keep efficiency and long-rollout drift as separate axes. The baselines generate a fixed 5–6 s clip, whereas Custom Forcing generates 30 s causally; because it runs on an autoregressive backbone, it also generates video much faster (Table R7). Its times include custom-anchor generation and the VAE encoding of the references and anchor, and Table R7 also times a 5.06 s clip generated directly with the same settings.

Table R7. Generation time on one H100, excluding model loading (Table 18).

MethodLengthTime per videoTime per video second
VACE-1.3B5.06 s220.0 s43.5 s
MAGREF-14B5.13 s469.0 s91.4 s
Phantom-1.3B5.06 s193.3 s38.2 s
1% Compute6.13 s138.0 s22.5 s
Custom Forcing (5.06 s clip)5.06 s16.0 s3.2 s
Custom Forcing (30 s rollout)31.3 s67.2 s2.1 s

Positioning. App. A.2 now cites and discusses MAGREF and 1% Compute. Like VACE and Phantom, they train a shared reference-conditioning module once (tuning-free), whereas Custom Forcing trains nothing and controls a frozen causal backbone at inference time (training-free). Table 1 now splits these two columns and no longer labels VACE training-free.

Reviewer_z98hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Custom Forcing, a training-free method for subject customization in streaming autoregressive video generation. It stores reference images and a scene-specific generated anchor in the persistent KV cache of a frozen video model. Drift-adaptive value amplification (DVA) adjusts reference conditioning based on measured identity drift, while anchor contrast guidance (ACG) strengthens reference contributions by contrasting attention outputs with and without the anchors. Experiments on ten DreamBooth subjects show improved identity preservation over videos lasting up to two minutes. The evaluation includes comparisons with customization and long-video baselines, component ablations, transfer to Deep Forcing, an alternative drift encoder, and a 30-participant user study.

Critical strengths
  1. The method addresses a practical problem with a lightweight intervention. Preserving a user-specified subject throughout a causal rollout is distinct from maintaining consistency with previously generated frames. Custom Forcing uses the existing KV-cache interface without subject-specific training or changes to model weights. Its modest additional generation cost makes the approach practically appealing.

  2. The main identity-preservation gains have several complementary sources of support. On two-minute videos, the final-window DINO-I increases from 0.415 with fixed anchors to 0.578 with the full method (Table 1). Improvements also transfer to Deep Forcing, and replacing the drift encoder with ConvNeXt produces similar results. The subject-level analysis reports gains for all ten subjects, while the user study favors the full method over fixed anchors for identity preservation. Together, these results make the central empirical finding convincing within the evaluated setting.

Critical weaknesses
  1. The constant-amplification comparison does not adequately isolate the benefit of adaptivity. Appendix F.4 compares DVA with fixed drift values of 0.05, 0.1, and 0.2, while reporting that the adaptive drift averages only 0.022. The fixed alternatives therefore use substantially stronger conditioning, making their reduced motion and scene distortion difficult to attribute specifically to the absence of feedback. A smaller fixed value matched to the adaptive method’s effective strength, or a simple time-dependent schedule, would provide a more informative control. This matters because online drift adaptation is a central methodological contribution.

  2. The evaluation provides limited evidence of generalization beyond the selected subjects and generation seed. All videos use seed 0, and the benchmark contains ten subjects, with related animal categories sharing prompt templates. Although the subject-level bootstrap and second autoregressive model are useful, they do not establish robustness to generation randomness or a broader range of identities and reference conditions. The human study covers 30-second videos, leaving the two-minute identity-preservation claim supported primarily by automated metrics and qualitative examples.

Questions for authors
  1. How does DVA compare with fixed amplification matched to its average effective strength, such as a fixed drift near 0.022, and with a simple time-dependent schedule? Does adaptive feedback still improve the identity–motion trade-off?

  2. Were the injection and guidance hyperparameters selected using the same subjects and prompts used for evaluation? Can you provide results on held-out subjects or additional random seeds?

  3. Do human judgments confirm the identity-preservation improvement over the full two-minute videos? The current user study evaluates 30-second outputs, whereas longer rollouts are a central motivation.

  4. Does the timing comparison include custom-anchor generation, reference-mask preparation, and KV-cache construction? Since the baselines use different resolutions and frame rates, reporting absolute latency and generated video seconds per wall-clock second would help interpret the practical speed advantage.

4 comments
AuthorsOpenReview ↗

Response to Reviewer z98h (1/2): Q1 matched constant and schedule, Q2 hyperparameters, held-out subjects, and seeds

Thank you for the thoughtful and constructive review. All new results are in the revised PDF (changes in blue). Tables R1–R5 appear in these comments; all other table, figure, section, and appendix numbers refer to the revised PDF. We answer Q1–Q2 here and Q3–Q4 in the next comment.

Q1/W1. Matched constant and time-dependent schedule. We added a constant b ≡ 0.022, matched to the mean b(t) of DVA, and a linear schedule with the same mean, keeping all other settings (App. F.4, Table 13, Fig. 13). On automated metrics the three sit at nearly the same identity–motion point, so we acknowledge both as competitive baselines (Table R1). They differ in when they amplify. Please see Fig. 13 in App. F.4 of the revised PDF, which illustrates this distinction. Where little drift is measured (Fig. 13(b)), DVA keeps b(t) near zero, while the controls keep amplifying and can enlarge or deform the subject (red boxes), even when their DINO-I is higher. Where drift is measured (Fig. 13(a)), DVA raises b(t) and all three stay plausible. Human judgments reflect this difference: in a 30 s study, 30 participants ranked the videos of the three methods for 20 prompts, with ties allowed, and Custom Forcing was ranked better than each control more often than worse, for both identity and quality (Holm-adjusted p ≤ 0.009 in all four comparisons; App. D.7, Table 9(d); Table R2). The advantage of adaptive feedback therefore appears in human judgments, not as a higher average identity score.

Table R1. Automated metrics, 30 s, n = 100 (subset of Table 13).

DINO-ICLIP-ICLIP-TSubj. Cons.Bg. Cons.Dynamic Degree
b ≡ 0.0220.6430.8050.3530.9300.9470.37
Linear schedule0.6290.8020.3550.9320.9480.41
DVA (ours)0.6350.8030.3530.9280.9450.41

Table R2. Human ranking, 30 s, 30 participants, 20 prompts (Table 9(d)). Share of answers (%) in which Custom Forcing was ranked better than, the same as, or worse than each control; p: mixed-effects logistic regression on the untied answers, Holm-adjusted within each question.

Custom Forcing vs.Identity: better / same / worsepQuality: better / same / worsep
b ≡ 0.02259.7 / 26.8 / 13.50.000660.3 / 24.3 / 15.30.0001
Linear schedule50.0 / 34.3 / 15.70.008950.3 / 31.3 / 18.30.0057

Q2/W2. Hyperparameter selection, held-out subjects, and seeds. Yes: α_inj = 300 and κ = 4 were chosen from the sweeps in Tables 10–11 on the same 100 evaluation videos; there was no separate development set, and App. H.1 now states this. Keeping every hyperparameter fixed, we ran two more video seeds and 20 held-out DreamBooth subjects with 10 new prompts each, which were not used for any selection (App. H.1, Table 14; Table R3). Custom Forcing improves DINO-I over Custom K/V only in every setting. Beyond DreamBooth, App. H.4 evaluates human faces (30 ConsisID identities) and App. H.5 evaluates OpenS2V-Eval, where each prompt has a single reference photo.

Table R3. DINO-I at 30 s, Custom K/V only → Custom Forcing (Table 14).

SubjectsVideo seedRolling ForcingDeep Forcing
Original 1000.541 → 0.6350.556 → 0.641
Original 1010.544 → 0.6350.539 → 0.641
Original 1020.543 → 0.6320.545 → 0.638
Held-out 2000.533 → 0.6100.535 → 0.638
AuthorsOpenReview ↗

Response to Reviewer z98h (2/2): Q3 two-minute human evaluation, Q4 timing

Q3/W2. Two-minute human evaluation. We ran a two-minute study: 30 participants compared the 2 min videos of Custom Forcing and Custom K/V only for 20 prompts, with the same two questions as the 30 s study, choosing A or B with the positions of the two methods balanced (600 answers per question). The identity question asks which video keeps the subject more similar to the reference photos from start to end, so each answer covers the full two minutes. Participants chose Custom Forcing for identity in 73.7% of answers and for overall quality in 63.3% (Table R4). Both new studies are in App. D.7 (Table 9(c), 9(d)), and Sec. 4.2 reports the two-minute result.

Table R4. Two-minute study, 30 participants, 600 answers per question (Table 9(c)). Model: estimate of the mixed-effects logistic regression with crossed random intercepts for participants and prompts and an A/B position term, with its 95% CI.

Custom Forcing vs.QuestionChosenModel [95% CI]p
Custom K/V onlyIdentity73.7%80.1% [68.4, 88.2]< 0.0001
Custom K/V onlyOverall quality63.3%66.7% [50.4, 79.8]0.045

Q4. What the timing includes. The submitted Table 2 included the KV-cache prefill (0.3 s per video) but excluded custom-anchor generation (FLUX.2 [klein] 4B, 3.0 s per prompt) and the VAE encoding of the references and anchor (0.6 s). The revised Table 2 and the new Table 15 (App. H.2) include anchor generation for every method that takes the anchor (all but SkyReels-V3) and VAE encoding for our two rows. Reference masks (SAM ViT-H, 0.20 s per image) are made once per subject and are not included. Because the methods differ in resolution and frame rate, Table 15 reports absolute latency and generated video seconds per wall-clock second (Table R5).

Table R5. One H100 80GB, native settings, means over 100 videos (Table 15).

MethodResolution, fpsVideoTime per video ↓Video s / wall-clock s ↑
FramePack832×480, 3032.4 s1,165 s0.028
FramePack-F1832×480, 3032.4 s1,173 s0.028
SkyReels-V2 DF960×544, 2431.4 s1,103 s0.028
SkyReels-V31280×720, 2432.7 s2,836 s0.012
Custom K/V only832×480, 1631.3 s59.6 s0.525
Custom Forcing832×480, 1631.3 s67.2 s0.466
Reviewer_z98hOpenReview ↗

Response to the authors' rebuttal

Thanks for your detailed response. The additional experimental results have addressed my main concerns. And I will maintain the score at 3.

AuthorsOpenReview ↗

Thank you for your follow-up and continued support

Thank you for reviewing our additional results and for your positive assessment of our work. We are glad that the new experiments addressed your main concerns. Your feedback helped us strengthen our evaluation, and we will include these results in the final version.

Thank you again for your time and thoughtful review. Please let us know if you have any further questions before the discussion period ends.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.