acceptodds
Under review as a conference paper at ICLR 2027

SubjectSwitch: Background-Preserving Subject Switching in Interactive Long Video Generation

Abstract

When a new text prompt requests a subject change during interactive long video generation, the video generation model must replace the existing subject while preserving background context that remains compatible with the new instruction. Yet, existing autoregressive models struggle with this balance, since they commonly use a key–value (KV) cache to reuse representations of previously generated frames. To address these limitations, we introduce SubjectSwitch, a training-free framework that selectively updates the recent KV cache and controls the influence of the persistent attention sink. For the recent KV cache, SubjectSwitch first identifies subject changes between consecutive prompts. It then suppresses background responses in subject attention by subtracting background attention, thereby locating the subject-related regions of the KV cache. These regions are selectively reconstructed under the new prompt, while the background regions of the KV cache are preserved. However, selectively updating the recent cache alone is insufficient to prevent subject leakage; the persistent attention sink still continuously broadcasts visual cues from the initial frames, causing the obsolete subject to reappear later in the video. To resolve this, we estimate its influence at selected generation blocks and denoising timesteps and selectively correct predictions in the affected regions, reducing the reappearance of initial-frame content in later frames. We evaluate 60-second generation with six sequential prompts on VidProM-Switch and MotionBind-Switch. SubjectSwitch achieves the best results among the compared methods on Subject Switch Cleanliness, Clean Switch Rate, and Switch Latency Score, improving Clean Switch Rate over LongLive, the strongest baseline on this metric, by 23.43 and 26.60 percentage points, respectively, while keeping the five reported VBench-Long scores within 1.3 points of LongLive.

Primary area:
generative models
Keywords:
Generative models, Video generation, Long video generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EuAgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SubjectSwitch is a training-free method for replacing subjects during interactive long-video generation while preserving compatible backgrounds. SAR suppresses the outgoing subject in historical recaching inputs and selectively replaces KV entries at source-subject locations. CSRS compares clean-latent predictions with and without the initial frame sink, then applies regionally gated corrections to reduce suspected replay. Experiments on two prompt-switching benchmarks report substantial improvements in clean subject replacement over LongLive.

Critical strengths
  1. The method addresses a concrete failure mode with targeted memory updates. Suppressing the outgoing subject before recaching avoids reintroducing its appearance through historical inputs. Selective cache writes also provide a clear mechanism for preserving background context.
  2. The reported overall improvement is substantial. Clean Switch Rate increases from 54.92/61.06 to 78.35/87.66 over LongLive, while background consistency decreases by only 0.03/0.39 points. These results support the practical value of the complete system.
Critical weaknesses
  1. The overall pipeline is understandable, but the replay-score construction, mask recovery/refinement procedures, and several gating settings are insufficiently specified.
  2. The semantic benefit of CSRS remains insufficiently established. Its correction subtracts a gated difference between with-sink and without-sink clean-latent predictions. This difference can contain useful context as well as obsolete content, making the replay detector central to correctness. However, its exact score construction is not provided. In Table 2, CSRS changes CSR by −0.59/+0.85 points, its strongest consistent improvement is in RS, which measures abrupt pixel-similarity increases toward the first frame rather than semantic subject recurrence.
  3. The spatial adaptation mechanism needs more targeted validation. Historical cache updates and initial transition control rely on the outgoing subject’s support. Although the method subsequently localizes the new subject and commits target frames in full, it is unclear how reliably the first target block is generated when the new subject differs substantially in location or extent.
Questions for authors
  1. Please specify the complete replay-score formula, spatial/temporal alignment procedure, gating parameters, and mask refinement rules. How sensitive are the results to these choices?
  2. On manually annotated recurrence events, does SAR+CSRS reduce semantic recurrence relative to SAR alone? How well does RS agree with these annotations?
  3. How does the method perform when the replacement subject changes location or size substantially? Can controlled ablations distinguish the contributions of source suppression, selective cache writes, and transition blending?
  4. What are the measured end-to-end overhead and prompt-switch latency, including recaching and additional prediction branches?
4 comments

We have added the complete formulation and implementation details in Appendix H, with default parameters in Tables 9 and 10.

Replay scoring and gating. Each region u∈{sub,bg}u\in\{\mathrm{sub},\mathrm{bg}\} is scored using appearance, motion, boundary-discontinuity, and causal-alignment evidence:

\qquad g_u=\mathbf{1}[r_u\ge\eta_u]\, \sigma\left(\frac{r_u-\eta_u}{\max(T,10^{-6})}\right).$$ The background and subject gates are independent, with thresholds of 0.15 and 0.06, respectively, and temperature $T=0.05$. H.1 specifies every score term, regional weight, and normalization rule; H.3 specifies gating, correction, and execution scheduling. **Spatial/temporal correspondence.** Replay scoring compares matching within-block frame indices and latent coordinates, without optical-flow registration or temporal search. Reference lengths are matched by repetition/truncation, and spatial sizes by per-frame bilinear interpolation. Historical cache updates use global latent-frame indices to retrieve the corresponding source masks (H.2). **Mask selection and refinement.** Masks are selected from background-subtracted attention using a cumulative attention-mass fraction of 0.72 and initial area bounds of 2%–45%. H.4 specifies the quality checks, bounding-box recovery, and bounded-attention recovery rules. Temporal refinement uses local bidirectional matching of cached self-attention value features, with a three-token search radius, minimum confidence of 0.50 in both directions, and one-token cycle-consistency tolerance. Supports propagate with a factor of 0.75, are binarized at values greater than zero, and are trimmed to the 45% per-frame area cap. All selection and refinement parameters are listed in Table 10. To assess sensitivity, Appendix F evaluates 100 six-prompt sequences randomly sampled from VidProM-Switch and MotionBind-Switch, with 50 sequences from each dataset. All configurations use the same subset, and each configuration applies shared parameter values across all sequences. We change one parameter at a time while keeping the others fixed. The default background threshold, subject threshold, background-attention subtraction strength, and cumulative attention-mass fraction are 0.15, 0.06, 0.75, and 0.72, respectively. The subject-switching and replay results from Table 6 are reproduced below. All scores use a 0–100 scale; higher is better except for RS, for which lower is better. RS measures the cumulative severity of abrupt pixel-similarity increases toward the first frame. | Setting | SSC ↑ | CSR ↑ | SLS ↑ | RS ↓ | |---|---:|---:|---:|---:| | SubjectSwitch (default) | 89.667 | 82.778 | 73.519 | 3.785 | | Background replay threshold = 0.10 | 87.833 | 80.556 | 73.333 | 3.821 | | Background replay threshold = 0.20 | 88.375 | 81.111 | 73.472 | 3.781 | | Subject replay threshold = 0.04 | 87.472 | 80.000 | 72.963 | 3.158 | | Subject replay threshold = 0.08 | 87.819 | 80.000 | 72.963 | 3.683 | | Background-attention subtraction strength = 0.50 | **91.597** | **86.111** | **74.028** | 4.036 | | Background-attention subtraction strength = 1.00 | 88.764 | 81.111 | 73.102 | **1.840** | | Cumulative attention-mass fraction = 0.65 | 85.069 | 75.833 | 70.509 | 2.152 | | Cumulative attention-mass fraction = 0.80 | 89.597 | 83.333 | 73.981 | 3.826 | Across the replay-threshold settings, the maximum decreases from the default configuration are 2.19 points in SSC, 2.78 percentage points in CSR, and 0.56 points in SLS, while RS ranges from 3.158 to 3.821. This indicates limited sensitivity of end-to-end switching performance within the tested threshold ranges. Across all nine configurations, RS ranges from 1.840 to 4.036, with the lowest value obtained at a background-attention subtraction strength of 1.00. Across all nine SubjectSwitch configurations, the ranges of SC, BC, MS, AQ, and IQ are 0.181, 0.038, 0.036, 0.461, and 0.370 points, respectively, each below 0.5 points. These results support using shared settings across the evaluated prompts within the tested parameter ranges. The sensitivity study covers these four parameters; score weights, gate temperature, and correspondence/refinement constants remain fixed.

The detailed construction of RS, including its thresholds and aggregation procedure, is provided in Appendix G.1, “Definition and Interpretation of RS.” Appendix G.2, “Human Reversion Assessment and RS,” reports the direct comparison between SAR alone and SAR+CSRS, with all other generation settings unchanged. The results are reproduced below; each row contains 200 videos.

DatasetMethodHuman return events ↓Videos with eventsRS ↓
Dataset1SAR only277182/200 (91.0%)34.06
Dataset1SAR+CSRS (Ours)1313/200 (6.5%)0.45
Dataset2SAR only298191/200 (95.5%)53.36
Dataset2SAR+CSRS (Ours)87/200 (3.5%)6.24

Adding CSRS reduces manually annotated opening-content return events from 277 to 13 and from 298 to 8, corresponding to reductions of 95.31% and 97.32%, respectively. The proportions of videos with an annotated return decrease from 91.0% to 6.5% and from 95.5% to 3.5%. These observations provide evidence beyond RS that adding CSRS to SAR reduces perceptible opening-content replay. Regarding agreement with human annotations, RS decreases from 34.06 to 0.45 and from 53.36 to 6.24, matching the aggregate direction of the human results.

To examine performance under substantial spatial changes, Appendix E.1 evaluates LongLive and SubjectSwitch on 100 six-prompt sequences targeting substantial changes in subject position or spatial extent. The results from Table 4 are reproduced below. All scores use a 0–100 scale; higher is better.

MethodSSC ↑CSR ↑SLS ↑
LongLive73.2655.7954.91
SubjectSwitch87.8975.7974.39

SubjectSwitch improves SSC and SLS by 14.63 and 19.48 points, respectively, and CSR by 20.00 percentage points. These results support improved end-to-end switching performance on the evaluated spatial-change benchmark.

We acknowledge that the first target block may not reliably realize a substantially different subject location or extent, because the initial transition still uses the outgoing subject’s support. In our configuration, each prompt covers approximately 40 latent frames on average, while one block contains three latent frames, accounting for approximately 7.5% of a prompt segment. The source-derived initial transition therefore occupies a relatively small portion of the segment.

After the first target block is generated, we localize the target from the actual generated latents and incorporate its support into subsequent transition blending. Newly generated target frames are committed across the full spatial grid, while updates to pre-switch history remain restricted to the source support. This provides a mechanism for subsequent adaptation to the target’s observed location and extent. Nevertheless, we recognize that reliably generating the first target block remains an issue that cannot be overlooked.

To distinguish the three components’ contributions, Appendix I additionally reports controlled leave-one-component-out ablations on a separate set of 100 six-prompt sequences, with 50 sampled from each of our two prompt datasets. All variants use matched per-sequence random seeds and the same remaining generation settings; the full method is regenerated using the same generation code. Scores are averaged over the same 360 subject-changing boundaries. Table 11 is reproduced below.

MethodSSC ↑CSR ↑SLS ↑
SubjectSwitch (full)89.0681.9473.43
w/o source suppression71.1450.8341.99
w/o selective cache writes83.6075.0071.06
w/o transition blending96.4393.6180.56

Removing source suppression disables masked source-latent suppression during both transition and target recaching, while retaining selective writes and blending. SSC and SLS decrease by 17.92 and 31.44 points, respectively, and CSR decreases by 31.11 percentage points. This supports its role in reducing old-subject residue and enabling clean, timely replacement.

Removing selective cache writes instead applies uniform K/V writes within the same recaching window, while retaining source suppression and blending. SSC and SLS decrease by 5.46 and 2.36 points, respectively, and CSR decreases by 6.94 percentage points, supporting the contribution of spatially selective cache updates.

Removing transition blending directly adopts the target-branch prediction while retaining the branch-computation and cache-control paths. This improves all three switching metrics, so we do not claim that transition blending improves switching cleanliness or latency. The transition blending module is designed to make subject changes appear more natural and gradual, mitigating the abruptness of directly switching from the old subject to the new one. Specifically, it combines predictions from the old- and new-prompt branches using attention-derived spatial support and soft weights, progressively increasing the target branch’s weight as generation proceeds. This is intended to let the subject and surrounding regions complete their visual transformation gradually.

This design also explains the ablation results: SSC, CSR, and SLS reward earlier clean target emergence and penalize intermediate mixed appearances, but they do not directly measure the perceptual naturalness of the transformation. Intermediate appearances retained for a gradual transformation can therefore lower these scores. Figure 15 illustrates this distinction: the full method exhibits intermediate appearance during the transition, whereas the variant without blending produces the complete target subject earlier. Transition blending thus focuses on controlling how the subject changes over time to produce a gradual visual transition.

The measured end-to-end video-generation runtime overhead relative to LongLive is 16.23%. Prompt-switch latency increases from 4.785 s to 9.964 s, an additional 5.179 s per switch. These measurements include recaching and the additional prediction branches.

Reviewer_R3ypOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SubjectSwitch, a training-free method for replacing subjects during interactive long video generation while preserving compatible background context. Evaluation shows substantial improvements in subject switching over LongLive.

Critical strengths
  1. I think the main idea technically makes sense. Recomputing a cache under a new prompt can still preserve the old subject because the historical latents have not changed. Replacing the historical latents in that region with the background mean, then reconstructing and replacing the local KV under a new prompt, makes sense.

  2. The switching results are strong, with large improvements over LongLive.

Critical weaknesses
  1. CSRS makes sense to me, but it seems to depend on threshold parameters. This method depends a lot on heuristics, which makes it less practical.

  2. The practical cost of the full pipeline is unclear. It uses offline LLM parsing. The paper claims negligible overhead but does not report inference throughput or wall-clock latency.

  3. See Question 1.

  4. The related-work discussion could be more complete. For example it missed related works like FramePack, Context Forcing, Context as Memory, and RELIC

Questions for authors
  1. I have a question about the practicality of the task this paper is solving: isn't it very rare to see people want to change objects during real-time generation, unless they are generating some special animated shots? And doesn't it look very unrealistic? In long video generation, we keep objects consistent, and if you want to switch objects more naturally, an object can walk out of the camera's view and then a new object can come in.
4 comments

We acknowledge that CSRS includes threshold parameters. To examine its dependence on their precise values, we have added a sensitivity analysis in Appendix F using 100 six-prompt sequences randomly sampled from VidProM-Switch and MotionBind-Switch, with 50 sequences from each dataset. All configurations use the same subset, and each configuration applies one shared parameter setting across all 100 sequences.

Specifically, we test background replay thresholds of 0.10, 0.15, and 0.20, and subject replay thresholds of 0.04, 0.06, and 0.08, changing one parameter at a time while keeping the others fixed. We additionally vary the background-attention subtraction strength and cumulative attention-mass fraction. The default values for these four parameters are 0.15, 0.06, 0.75, and 0.72, respectively. The subject-switching and replay results from Appendix F, Table 6, are reproduced below. All scores use a 0–100 scale; higher is better except for RS, for which lower is better. RS measures the cumulative severity of abrupt pixel-similarity increases toward the first frame.

SettingSSC ↑CSR ↑SLS ↑RS ↓
SubjectSwitch (default)89.66782.77873.5193.785
Background replay threshold = 0.1087.83380.55673.3333.821
Background replay threshold = 0.2088.37581.11173.4723.781
Subject replay threshold = 0.0487.47280.00072.9633.158
Subject replay threshold = 0.0887.81980.00072.9633.683
Background-attention subtraction strength = 0.5091.59786.11174.0284.036
Background-attention subtraction strength = 1.0088.76481.11173.1021.840
Cumulative attention-mass fraction = 0.6585.06975.83370.5092.152
Cumulative attention-mass fraction = 0.8089.59783.33373.9813.826

Across the replay-threshold settings, including the default, SSC, CSR, and SLS remain within 87.47–89.67, 80.00–82.78, and 72.96–73.52, respectively, while RS ranges from 3.158 to 3.821. Relative to the default configuration, the maximum decreases are 2.19 points in SSC, 2.78 percentage points in CSR, and 0.56 points in SLS. These results indicate limited sensitivity of end-to-end switching performance within the tested threshold ranges.

Across all nine SubjectSwitch configurations, RS ranges from 1.840 to 4.036, with the lowest value obtained at a background-attention subtraction strength of 1.00. The ranges of SC, BC, MS, AQ, and IQ are 0.181, 0.038, 0.036, 0.461, and 0.370 points, respectively, each below 0.5 points. The complete video-quality results are provided in Table 6.

While the method retains heuristic design choices, these results provide evidence of tolerance to parameter changes within the tested ranges and support the practical use of shared replay-threshold settings across the evaluated prompts.

The measured video-generation runtime is 16.23% higher than LongLive, excluding offline LLM preprocessing, while inference throughput is 13.99% lower. Offline LLM parsing is benchmarked separately in a serial setting, averaging 2.208 s per successfully parsed prompt pair.

We agree that subject identity should remain consistent when the user intends to preserve the same subject, and that having one subject leave the view before another enters is a natural transition. We consider explicitly requested subject changes an additional editing capability for interactive generation. One possible use case is interactive storyboarding, where a user may redirect an ongoing scene toward a new character while retaining the established environment. The task does not prescribe an in-place transformation as the required transition.

The exit-and-entry transition you suggest is compatible with this task. It concerns how the change unfolds visually, while our work addresses how historical memory should adapt to an explicit replacement instruction. A subject leaving the current view does not necessarily remove its representations from the historical KV cache. When the user requests a replacement, these cached cues may conflict with the updated instruction and continue to influence subsequent generation. Our method addresses this potential conflict by selectively updating subject-related memory while preserving compatible background context.

We also evaluate perceived transition naturalness in our user study. In the comparison with LongLive, 36 of 50 responses (72%) favor SubjectSwitch on transition naturalness (Sec. 4.4, Fig. 7). This provides comparative evidence of improved perceived naturalness under our evaluation setting, although it does not establish that every transition is physically realistic.

We therefore view natural transition design and historical memory adaptation as complementary aspects of interactive subject changes. We have not systematically evaluated explicit exit-and-entry control; coordinating these motions with memory adaptation would be a valuable extension.

We have added discussion of all four works to Section 2 (Related Work, p. 3) of the revised manuscript. FramePack and Context Forcing are discussed under “Streaming and Autoregressive Video Generation,” while Context as Memory and RELIC are discussed under “Interactive Generation and Memory Adaptation.” The added passages are reproduced below:

FramePack compresses historical frame context according to importance and introduces anti-drifting sampling to mitigate long-horizon quality degradation (Zhang et al., 2025). Context Forcing addresses the supervision mismatch between long-context students and short-context teachers through long-context teacher supervision and Slow-Fast Memory (Chen et al., 2026b).

Context as Memory conditions generation on historical frames retrieved using camera field-of-view overlap to maintain scene consistency (Yu et al., 2025). RELIC supports interactive scene exploration through compressed historical latent tokens encoded with relative actions and absolute camera poses within the KV cache (Hong et al., 2025).

These additions are highlighted in red in the Related Work section of the revised manuscript.

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

This paper proposes SubjectSwitch, a training-free framework for subject switching in interactive long-video generation. Built on LongLive, the method addresses two failure modes: incomplete subject replacement caused by stale information in the recent KV cache, and the reappearance of obsolete subjects due to the persistent attention sink. It introduces subject-aware KV cache routing using subject–background attention contrast and counterfactual sink replay suppression to selectively correct predictions influenced by the initial frames. Experiments on VidProM-Switch and MotionBind-Switch show improved subject-switching metrics over several baselines, with additional user-study evidence supporting the perceptual benefits.

Critical strengths
  1. The motivation is well defined. The paper identifies two distinct failure modes in prompt-driven subject switching: residual source-subject information in the recent KV cache and unwanted recurrence caused by the persistent attention sink. It proposes two targeted solutions: subject-aware cache routing and counterfactual sink correction accordingly.
  2. Extensive experiments convincingly demonstrate the effectiveness of the approach.
Critical weaknesses
  1. The methodological novelty is somewhat incremental. What fundamentally distinguishes its subject–background attention contrast from existing attention-based localization and selective memory-update techniques is not clear yet.
  2. The paper covers a relatively narrow range of prompt transitions. It would be valuable to extend the evaluation to other interactive editing scenarios, such as changing the background while preserving the subject, modifying subject attributes, or combining multiple editing instructions.
  3. The paper does not report inference latency, throughput, or peak GPU memory usage compared with baselines. Quantifying these costs is important for assessing the practical value of the training-free approach in interactive generation.
Questions for authors
  1. Could the authors clarify the contribution of subject–background attention contrast? In particular, how does background suppression improve subject localization over existing attention-based methods? A controlled comparison with simpler alternatives, such as subject-attention masks without background suppression, would help isolate its benefit.
  2. Can the authors evaluate transitions that modify the background while preserving the subject, change subject attributes, or combine multiple edits? Such results would clarify whether the proposed framework generalizes beyond subject identity or category replacement.
3 comments

We compare LongLive + attention control with LongLive + SAR, without CSRS. The results are reported in Appendix D, Table 3. The subject-switching metrics are reproduced below.

V = VidProM-Switch; M = MotionBind-Switch. All scores use a 0–100 scale; higher is better.

MethodSSC ↑ (V / M)CSR ↑ (V / M)SLS ↑ (V / M)
LongLive70.43 / 76.7254.92 / 61.0656.50 / 53.67
LongLive + attention control76.78 / 83.2671.43 / 80.0966.61 / 72.01
LongLive + SAR86.84 / 92.2078.94 / 86.8174.02 / 74.17

On VidProM-Switch and MotionBind-Switch, SAR improves CSR over the attention-control baseline by 7.51 and 6.72 percentage points, respectively. SSC improves by 10.06 and 8.94 points, while SLS improves by 7.41 and 2.16 points. The complete comparison, including the five VBench-Long metrics, is provided in Appendix D.

Figure 5 further examines the localization mechanism. Panels (b) and (d) visualize subject attention and its positive difference after subtracting weighted background attention, using a common color scale. Panels (e) and (f) show the corresponding diagnostic masks obtained with the same threshold. In this example, directly thresholding subject attention retains scattered background responses, whereas background suppression produces a mask more concentrated on the subject. This illustrates how subject-attention-only masking can include irrelevant background regions and provide less focused guidance for subject-related KV cache reconstruction.

Together, these results support the additional switching benefits of SAR as a whole over attention control alone, while Figure 5 illustrates the localization benefit of background suppression. SAR couples prompt-guided localization with subject-aware reconstruction of historical KV memory. Specifically, stable-background attention provides an explicit reference for suppressing background responses and identifying historical regions requiring reconstruction (Sec. 3.2, Eq. 4). SAR then suppresses outgoing-subject cues in the historical latents before re-encoding them under the new prompt. The reconstructed subject-related KV entries are written back while the original background entries are retained (Eq. 5). This addresses a limitation of selecting update regions alone: historical latents can still carry outgoing-subject appearance into the reconstructed cache. Thus, SAR addresses both where historical memory should be updated and which visual information should be used to reconstruct it.

We have expanded the evaluation and application scenarios in Appendix E, adding quantitative results for subject replacement with substantial spatial changes and background switching, together with qualitative examples of background switching and subject attribute editing.

1. Subject replacement with substantial spatial changes. We evaluate LongLive and SubjectSwitch on a dataset of 100 six-prompt sequences targeting substantial changes in subject position or spatial extent. The results from Appendix E.1, Table 4, are reproduced below. All scores use a 0–100 scale; higher is better.

MethodSSC ↑CSR ↑SLS ↑
LongLive73.2655.7954.91
SubjectSwitch87.8975.7974.39

SubjectSwitch improves SSC and SLS by 14.63 and 19.48 points, respectively, and CSR by 20.00 percentage points. These results support cleaner and more timely subject replacement in the evaluated spatial-change scenarios.

2. Background switching with subject preservation. We extend cache editing to the background by treating the complement of the localized subject region as the background region and directing cache updates to this region.

We evaluate LongLive and this background-editing extension on 50 paired six-prompt sequences, with two requested background switches per video. Five frames are sampled within each 10-second post-switch window, yielding 100 switching windows and 500 sampled frames per method.

Appendix E.2 introduces Background Matching Score (BMS). Qwen2.5-VL-7B-Instruct first describes each sampled frame without access to the generation prompt or method label; a subsequent text-only judgment compares this description with the requested target background. The results from Table 5 are reproduced below.

MethodBMS ↑
LongLive37.61
SubjectSwitch (background-editing extension)60.73

The 23.12-point improvement indicates greater agreement with the requested backgrounds at the sampled post-switch times. In addition, Figure 13 provides complementary qualitative examples from ten generated videos, illustrating background changes alongside broadly consistent subject appearance.

3. Subject attribute editing. We also explore modifying selected attributes of the same subject, including color, hairstyle, facial hair, and size. Figure 14 in Appendix E.3 presents examples from ten generated videos covering human, animal, robot, and everyday object subjects. These examples illustrate attribute changes alongside broadly consistent scene content, although some outputs remain imperfect. Since our current subject-switching metrics do not directly assess attribute-editing accuracy and preservation of unedited content, we report qualitative results for this setting.

Compared with LongLive, SubjectSwitch increases inference time by 16.26%, corresponding to 13.99% lower inference throughput, and increases peak GPU memory usage by 1.584 GB (3.83%).

Comments on the paper

1 comment

Summary of Revisions

We thank the reviewers for their constructive feedback. The revised manuscript retains the original method and main experimental results while adding further validation and implementation details. The major additions are highlighted in red.

  1. Additional attention-control comparison (Appendix D, Table 3). We compare LongLive, LongLive with attention control alone, and LongLive with SAR. SAR improves CSR over attention control alone by 7.51 and 6.72 percentage points on VidProM-Switch and MotionBind-Switch, respectively.

  2. Subject switching with spatial changes (Appendix E.1, Table 4). We add an evaluation on 100 six-prompt sequences involving substantial changes in subject position or spatial extent. SubjectSwitch achieves a CSR of 75.79, compared with 55.79 for LongLive, and improves SSC and SLS.

  3. Extended editing scenarios (Appendices E.2–E.3, Table 5, Figures 13–14). We add an exploratory background-switching comparison on 50 paired sequences, obtaining a Background Matching Score of 60.73 versus 37.61 for LongLive, together with qualitative examples of background switching and subject attribute editing.

  4. Parameter sensitivity (Appendix F, Table 6). We evaluate nine configurations on a fixed subset of 100 sequences, varying four parameters one at a time. We report switching metrics, RS, and video-quality metrics, showing modest changes under the tested replay-threshold variations and larger effects from some localization settings.

  5. Human assessment and RS interpretation (Appendix G, Tables 7–8). We add human opening-content reversion annotations for 800 videos comparing SAR alone with SAR+CSRS. Annotated events decrease from 277 to 13 and from 298 to 8 on the two datasets. We distinguish the post-hoc pixel-based RS metric from the internal regional CSRS score and clarify their interpretation and limitations.

  6. Complete implementation specification (Appendix H, Tables 9–10). We provide the full replay-score formula and regional weights, spatial/temporal correspondence and frame-index routing, gating and correction parameters, execution schedule, and mask selection, recovery, and temporal refinement rules.

  7. Controlled component ablations (Appendix I, Table 11, Figure 15). We add matched-seed ablations on 100 six-prompt sequences, evaluated over the same 360 subject-changing boundaries. Removing source suppression or selective cache writes reduces CSR from 81.94 to 50.83 or 75.00, respectively. Removing transition blending increases CSR to 93.61 and also improves SSC and SLS. We clarify that blending is intended to encourage more natural, gradual visual transformations, with Figure 15 illustrating its effect on intermediate appearances during the transition.

  8. Expanded related work (Section 2). We extend the discussion of historical-context compression, long-context supervision, retrieval-based memory, and interactive scene exploration, including FramePack, Context Forcing, Context as Memory, and RELIC.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.