What Helps Denoising Can Hurt Memory: Rethinking History in Long Video Generation
Organizations: Georgia Institute of Technology · Peking University · Zhejiang University · Tsinghua University ·
Abstract
Autoregressive video generators produce long videos block by block, using historical context to maintain temporal consistency. After denoising, a committing pass encodes each clean block into keys and values (KV) that later blocks read, and the same clean block can be encoded into different KV depending on the history it reads. We therefore study historical context not only as a condition for denoising but also as an input to committing. Older historical context that supports denoising can harm subsequent generation when read during committing. At fixed context size and temporal positions, older historical context lowers mean clip-to-clip Subject Consistency. Across 5 generator configurations supporting longer contexts, consistency rises, plateaus, and then declines as committing context expands. Motivated by these findings, we propose Commit Forcing, a training-free rule that organizes committing into a chain. During committing, each block reads only the preceding block's stored KV, so older historical context reaches new entries only through the chain. We apply this rule to 6 pretrained generators. On 60-second VBench, Commit Forcing improves Quality and Total on all 6 generators, by up to 1.27 Quality points and 1.10 Total points. Long-range consistency gains extend to 240-second videos. In a blinded comparison on 120-second LongLive videos, Commit Forcing achieves win rates of 83.5% for subject and scene consistency and 82.9% for overall preference against the base generator, excluding ties.
Figures & tables
| Quality | Semantic | Total | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Configuration | Base | Commit Forcing | Base | Commit Forcing | Base | Commit Forcing | |||
| CF | 84.79 | 85.25 | ∗ | 72.03 | 72.41 | 82.24 | 82.68 | ∗ | |
| Rolling Sink | 84.84 | 85.25 | ∗ | 75.13 | 75.31 | 82.90 | 83.26 | ∗ | |
| RF-TF | 84.77 | 85.42 | ∗ | 72.10 | 75.00 | ∗ | 82.24 | 83.34 | ∗ |
| LongLive | 84.70 | 84.99 | ∗ | 76.43 | 76.84 | 83.05 | 83.36 | ∗ | |
| SGF+ | 86.32 | 86.63 | 76.22 | 76.66 | 84.30 | 84.64 | |||
| Generator | Window | Subject c2c | Dynamic Degree | Generator | Window | Subject c2c | Dynamic Degree |
| SGF | Base | 86.20 | 58.57 | LongLive | Base | 84.27 | 47.27 |
| 89.96 | 72.47 | 84.60 | 56.53 | ||||
| 90.45 | 68.40 | 85.52 | 54.13 | ||||
| 90.25 | 70.60 | 85.49 | 53.83 | ||||
| 85.25 | 53.23 | ||||||
| RF-TF | Base | 75.46 | 64.80 | Rolling Sink | Base | 85.07 | 45.93 |
Appendix figures & tables48 assets
Supplementary material from the paper’s appendix.
Appendix
| Tail edit | DINOv2 | CLIP |
|---|---|---|
| Saturation , brightness | ||
| Saturation , brightness | ∗ | |
| Frozen tail | ∗ | |
| Central 80% crop, resized | ∗ | ∗ |
| Tail from another seed of the same prompt | ∗ | ∗ |
| Tail from another prompt | ∗ | ∗ |
| Contrast | DINOv2 | CLIP | Adjusted DINOv2 |
|---|---|---|---|
| SGF: sink excluded from committing | ∗ | ∗ | |
| SGF: sink excluded from denoising | ∗ | ∗ | ∗ |
| SGF: sink excluded from both stages | ∗ | ∗ | ∗ |
| SGF: committing sink retained, denoising sink excluded | |||
| LongLive: sink excluded from committing | |||
| LongLive: sink excluded from denoising | ∗ | ∗ |
| Reads sink | Consistency [95% CI] | Drift [95% CI] | |
|---|---|---|---|
| Denoising | Committing | ||
| LongLive: 12-frame reads | |||
| — | — | ||
| Generator | Readout | Sink | Recent | Reduction |
|---|---|---|---|---|
| SGF | Drift | 0.034 | 0.010 | ∗ |
| Brightness drift | 0.045 | 0.018 | ∗ | |
| Fraction with declining motion | 0.25 | 0.05 | ∗ | |
| LongLive | Drift | 0.051 | 0.033 | ∗ |
| Brightness drift | 0.035 | 0.032 | ||
| Fraction with declining motion | 0.28 | 0.05 | ∗ |
| Intervention | Consistency | Own | Other |
|---|---|---|---|
| LongLive: no sink when committing | ∗ | ∗ | |
| LongLive: no sink when denoising | ∗ | ∗ | ∗ |
| LongLive: no sink in either stage | ∗ | ∗ | ∗ |
| Rolling Sink: recent-context committing | ∗ | ∗ | |
| SF: committing with remote context throughout | ∗ | ∗ | ∗ |
| SF: committing with remote context before 24 s | ∗ | ∗ |
| policy | policy | Q | DD | Subject c2c |
|---|---|---|---|---|
| ∗ | ∗ | |||
| ∗ | ∗ | |||
| ∗ | ∗ | |||
| ∗ |
| Contrast | Q | DD | Subject c2c | IQ | AQ | |
|---|---|---|---|---|---|---|
| 200 | ∗ | ∗ | ||||
| 100 | ∗ | ∗ | ||||
| 100 | ∗ | ∗ | ∗ | |||
| 100 | ∗ | ∗ | ∗ |
| KV source | Position | Consistency | Drift | Brightness drift | log motion ratio |
|---|---|---|---|---|---|
| Recent | Recent | — | — | — | — |
| Recent | Initial | ||||
| Sink | Initial | ∗ | ∗ | ||
| Sink | Recent | ∗ | ∗ | ∗ |
| Contrast | Quality | Subject c2c | DD | MS | TF |
|---|---|---|---|---|---|
| B A | ∗ | ∗ | ∗ | ∗ | ∗ |
| B C | ∗ | ∗ | ∗ | ||
| B D | ∗ | ∗ | ∗ | ∗ | ∗ |
| Generator | History | Quality | DD | Subject c2c | Background c2c |
|---|---|---|---|---|---|
| SGF | 84.11 | 66.27 | 90.61 | 92.85 | |
| 84.21 | 68.37 | 90.53 | 92.74 | ||
| 84.19 | 68.60 | 90.30 | 92.55 | ||
| 84.24 | 69.67 | 90.02 | 92.42 | ||
| 84.34 | 71.30 | 89.92 | 92.34 | ||
| Sink | 83.14 | 72.53 | 78.62 | 83.52 |
| Generator | History | SC | BC | Subject in-clip | IQ | AQ | MS | TF |
|---|---|---|---|---|---|---|---|---|
| SGF | 98.26 | 96.84 | 97.66 | 70.58 | 62.34 | 98.42 | 96.87 | |
| 98.22 | 96.80 | 97.57 | 70.61 | 62.36 | 98.38 | 96.80 | ||
| 98.19 | 96.79 | 97.55 | 70.58 | 62.24 | 98.37 | 96.78 | ||
| 98.17 | 96.76 | 97.53 | 70.61 | 62.22 | 98.36 | 96.75 | ||
| 98.16 | 96.73 | 97.50 | 70.65 | 62.22 | 98.34 | 96.71 | ||
| Sink | 97.15 | 96.00 | 96.61 | 69.71 | 59.78 | 97.99 | 96.09 |
| Metric | SGF | LongLive |
|---|---|---|
| Clip-to-clip Subject Consistency | ||
| Clip-to-clip Background Consistency | ||
| Quality | ||
| Dynamic Degree | ||
| In-clip Subject Consistency | ||
| Motion Smoothness |
| Contrast | Metric | SGF | LongLive |
|---|---|---|---|
| minus | Clip-to-clip Subject | ||
| minus | Clip-to-clip Subject | ||
| minus | Clip-to-clip Subject | ||
| minus | Clip-to-clip Subject | ||
| minus | Dynamic Degree | ||
| Sink minus | Clip-to-clip Subject |
| Condition means | ||||||
|---|---|---|---|---|---|---|
| Metric | Sink | |||||
| Opening similarity | 0.9347 | 0.9351 | 0.9344 | 0.9339 | 0.9338 | 0.9295 |
| Absolute saturation change | 0.0388 | 0.0394 | 0.0389 | 0.0389 | 0.0387 | 0.0378 |
| Absolute brightness change | 0.0256 | 0.0269 | 0.0271 | 0.0264 | 0.0270 | 0.0240 |
| Latent distance | 0.4511 | 0.4531 | 0.4557 | 0.4546 | 0.4563 | 0.4823 |
| Luma boundary ratio | 1.7313 | 1.7558 | 1.7704 | 1.7431 | 1.7439 | 1.3866 |
| Committing | Intervened KV stored | Later K | Later V | |
|---|---|---|---|---|
| Recent | 1 | Yes | ||
| Recent | 2 | Yes | ||
| Recent | 3 | No | ||
| Recent | 5 | No | ||
| Recent | 8 | No | ||
| Recent | 11 | No |
| Free-generation blocks | saturation | brightness | Pixel MAE | Latent rel. L2 |
|---|---|---|---|---|
| 16–31 (exposure) | 28.52 | 0.548 | ||
| 32–35 | 29.50 | 0.574 | ||
| 36–43 | 21.69 | 0.444 | ||
| 44–51 | 20.04 | 0.408 | ||
| 52–63 | 19.61 | 0.404 |
| Implementation | Weights | Stored frames | Denoising | Committing | |
|---|---|---|---|---|---|
| Base | Recent-context | ||||
| RF-TF, main comparison | SF EMA | 24 | |||
| RF-TF, context control | SF EMA | 24 | |||
| RF denoising on SGF | SGF EMA | 21 | |||
| Alternative RF on SF | SF EMA | 21 | |||
| Configuration | Quality | Total |
|---|---|---|
| CF | ||
| Rolling Sink | ||
| RF-TF | ||
| LongLive | ||
| SGF+ | ||
| SGF |
| Generator | Quality | SC | BC | TF | MS | DD | AQ | IQ |
|---|---|---|---|---|---|---|---|---|
| CF | ||||||||
| Rolling Sink | ||||||||
| RF-TF | ||||||||
| LongLive | ||||||||
| SGF+ | ||||||||
| SGF |
| Configuration | Variant | SC | BC | TF | MS | DD | AQ | IQ |
|---|---|---|---|---|---|---|---|---|
| CF | Base | 97.75 | 96.63 | 99.30 | 98.55 | 58.89 | 65.41 | 69.47 |
| Commit Forcing | 97.96 | 96.92 | 99.21 | 98.43 | 62.31 | 65.82 | 70.32 | |
| Rolling Sink | Base | 98.09 | 97.07 | 99.11 | 98.77 | 53.94 | 65.53 | 70.88 |
| Commit Forcing | 98.28 | 97.34 | 99.06 | 98.66 | 57.41 | 66.02 | 71.28 | |
| RF-TF | Base | 96.73 | 95.92 | 98.88 | 98.37 | 70.28 | 62.65 | 70.34 |
| Commit Forcing | 97.36 | 96.70 | 98.96 | 98.31 | 67.36 | 65.03 | 71.82 |
| Configuration | Variant | OC | MO | HA | Col | SR | Sc | AS | TS | OCon |
|---|---|---|---|---|---|---|---|---|---|---|
| CF | Base | 88.63 | 70.79 | 91.33 | 84.31 | 69.19 | 40.20 | 20.88 | 20.83 | 23.60 |
| Commit Forcing | 87.23 | 73.34 | 93.67 | 85.09 | 67.40 | 39.47 | 20.97 | 21.10 | 23.89 | |
| Rolling Sink | Base | 89.97 | 74.51 | 91.17 | 91.58 | 72.81 | 42.81 | 20.77 | 22.85 | 24.98 |
| Commit Forcing | 89.85 | 74.31 | 90.67 | 92.98 | 73.46 | 43.63 | 20.74 | 22.71 | 24.92 | |
| RF-TF | Base | 88.05 | 63.58 | 92.83 | 87.96 | 73.15 | 35.47 | 20.66 | 21.96 | 24.59 |
| Commit Forcing | 90.49 | 70.46 | 94.67 | 89.78 | 75.76 | 43.01 | 20.60 | 22.36 | 24.77 |
| Configuration | Base | Commit Forcing | Semantic |
|---|---|---|---|
| CF | 72.03 | 72.41 | |
| Rolling Sink | 75.13 | 75.31 | |
| RF-TF | 72.10 | 75.00 | |
| LongLive | 76.43 | 76.84 | |
| SGF+ | 76.22 | 76.66 | |
| SGF | 74.99 | 74.78 |
| Base generator | Subject c2c [95% CI] | Background c2c [95% CI] |
|---|---|---|
| SGF | +2.71 [+1.28,+4.08] | +4.82 [+4.05,+5.64] |
| RF-TF | +4.44 [+3.14,+5.67] | +4.37 [+3.78,+4.97] |
| CF | +2.35 [+1.22,+3.49] | +3.35 [+2.56,+4.18] |
| Quality | Subject c2c | DD | |||||
| Configuration | Base | Modified | Base | Modified | |||
| Generation seed 0 | |||||||
| SGF (6) | 83.58 | 84.51 | ∗ | 86.54 | 91.30 | ∗ | ∗ |
| RF-TF (6) | 82.01 | 82.43 | ∗ | 75.02 | 81.38 | ∗ | ∗ |
| LongLive (6 frames) | 82.23 | 82.73 | ∗ | 86.01 | 87.21 | ∗ | ∗ |
| CF, full rule | 82.78 | 82.91 | 78.72 | 81.95 | ∗ | ||
| Generator | DD | Subject c2c | DD-adjusted Subject c2c |
|---|---|---|---|
| SGF | |||
| SGF+ | |||
| RF-TF | |||
| Rolling Sink | |||
| LongLive | |||
| CF |
| Comparison | Duration | Criterion | Wins | Losses | Ties |
|---|---|---|---|---|---|
| SGF: 9-frame reference vs. Base | 60 s | Consistency | 39 | 13 | 2 |
| Overall preference | 35 | 12 | 7 | ||
| LongLive: Commit Forcing vs. Base | 60 s | Consistency | 19 | 25 | 10 |
| Overall preference | 20 | 24 | 10 | ||
| LongLive: Commit Forcing vs. Base | 120 s | Consistency | 111 | 22 | 17 |
| Overall preference | 107 | 22 | 21 |
| Duration | Contrast | Quality | Subject c2c | Background c2c | DD | MS | TF |
|---|---|---|---|---|---|---|---|
| 120 s | 6 frames base | ∗ | ∗ | ∗ | ∗ | ∗ | ∗ |
| 6 frames 12 frames | ∗ | ∗ | ∗ | ∗ | ∗ | ||
| 6 frames ID-Forcing | ∗ | ∗ | ∗ | ∗ | ∗ | ||
| 240 s | 6 frames base | ∗ | ∗ | ∗ | ∗ | ∗ | ∗ |
| 6 frames 12 frames | ∗ | ∗ | ∗ | ∗ | |||
| 6 frames ID-Forcing | ∗ | ∗ | ∗ | ∗ | ∗ |
| Duration | Context / contrast | Quality | DD | Subject c2c | Background c2c |
|---|---|---|---|---|---|
| Self-only absolute scores | |||||
| 120 s | Self-only | 82.22 | 52.68 | 84.73 | 88.05 |
| 240 s | Self-only | 82.05 | 51.45 | 84.65 | 88.00 |
| Paired changes [95% CI] | |||||
| 120 s | Self-only base | ∗ | ∗ | ∗ | |
| 120 s | 6 frames self-only | ∗ | ∗ | ∗ | ∗ |
| Duration | Base | Configuration | Quality | Subject c2c | Color drift |
| 120 s | LongLive | Base | 81.72 | 84.16 | 60.15 |
| ID-Forcing | 82.27 | 78.72 | 70.35 | ||
| Commit Forcing (6 frames) | 82.05 | 86.03 | 74.45 | ||
| 12-frame ablation | 81.93 | 85.25 | 74.71 | ||
| SF | Base | 77.63 | 64.02 | 23.96 | |
| RF-TF + Recent-context | 81.78 | 76.27 | 65.33 |
| Committing configuration | Quality | Consistency | DD | |
|---|---|---|---|---|
| SC | BC | |||
| CF | ||||
| Base | 83.60 | 98.03 | 96.78 | 58.46 |
| Exclude slow memory only | 83.41 | 98.23 | 97.00 | 48.83 |
| Exclude sink only | 84.10 | 97.97 | 96.70 | 65.52 |
| Commit Forcing | 84.04 | 98.21 | 96.87 | 59.64 |
| Metric | Base | Commit Forcing | |
|---|---|---|---|
| RF-TF | |||
| Quality (reconstructed) | 82.29 | 82.86 | +0.57 [+0.35, +0.80] |
| Subject Consistency | 96.62 | 97.17 | +0.55 [+0.44, +0.67] |
| Background Consistency | 95.87 | 96.40 | +0.53 [+0.46, +0.59] |
| Temporal Flickering | 96.44 | 96.49 | +0.05 [ 0.08, +0.17] |
| Motion Smoothness | 98.18 | 98.11 | 0.06 [ 0.15, +0.01] |
| Generator | Committing context | Quality | Dynamic Degree | Fused Subject | Subject c2c |
|---|---|---|---|---|---|
| SF + sink | Sink + recent | 81.33 | 57.52 | 96.72 | 75.18 |
| SF + sink | Self-only | 81.28 | 67.80 | 96.56 | 74.69 |
| SF + sink | Recent trajectory (21) | 81.63 | 59.58 | 96.87 | 78.56 |
| LongLive | Sink + recent | 82.23 | 42.98 | 97.76 | 86.01 |
| LongLive | Self-only | 82.73 | 53.03 | 97.69 | 86.21 |
| LongLive | Recent trajectory | 82.65 | 49.38 | 97.70 | 86.93 |
| Metric | 6 frames | Self-only | |
| Quality | 84.99 | 85.17 | |
| Semantic | 76.84 | 76.49 | |
| Total | 83.36 | 83.43 | |
| Subject Consistency | 98.00 | 97.84 | |
| Background Consistency | 97.23 | 97.05 | |
| Temporal Flickering | 99.28 | 99.21 |
| Framework | Score | Base | 9-frame adaptation | |
|---|---|---|---|---|
| IDF-SF | Quality | 85.26 | 85.21 | |
| Semantic | 73.76 | 74.90 | ||
| Total | 82.96 | 83.15 | ||
| IDF-LL | Quality | 84.91 | 84.64 | |
| Semantic | 74.93 | 75.98 | ||
| Total | 82.92 | 82.91 |
| Configuration | Variant | SC | BC | TF | MS | DD | AQ | IQ |
|---|---|---|---|---|---|---|---|---|
| IDF-LL | Base | 97.80 | 96.89 | 99.21 | 98.51 | 56.85 | 66.08 | 70.58 |
| 9-frame adaptation | 97.90 | 97.05 | 99.29 | 98.64 | 50.69 | 66.26 | 70.72 | |
| IDF-SF | Base | 97.32 | 96.63 | 98.91 | 98.32 | 68.01 | 64.23 | 71.48 |
| 9-frame adaptation | 97.65 | 96.77 | 99.00 | 98.37 | 64.21 | 64.70 | 71.65 |
| Configuration | Variant | OC | MO | HA | Col | SR | Sc | AS | TS | OCon |
|---|---|---|---|---|---|---|---|---|---|---|
| IDF-LL | Base | 86.00 | 75.55 | 94.17 | 91.45 | 71.69 | 43.05 | 20.81 | 22.44 | 25.00 |
| 9-frame adaptation | 87.30 | 76.99 | 94.50 | 92.37 | 74.47 | 45.54 | 20.67 | 22.43 | 25.05 | |
| IDF-SF | Base | 87.73 | 67.96 | 92.83 | 87.30 | 74.71 | 42.26 | 20.55 | 22.26 | 25.06 |
| 9-frame adaptation | 88.66 | 73.12 | 92.33 | 89.66 | 76.41 | 42.78 | 20.40 | 22.41 | 25.10 |
| Base generator | Configuration | Quality | Dynamic Degree | Fused Subject | Subject c2c |
|---|---|---|---|---|---|
| SF | Base | 80.22 | 34.58 | 96.99 | 74.22 |
| SF | Complete ID | 82.20 | 61.95 | 97.03 | 79.07 |
| SF | ID framework + recent | 82.44 | 60.78 | 97.22 | 81.60 |
| LongLive | Base | 82.23 | 42.98 | 97.76 | 86.01 |
| LongLive | Complete ID | 82.54 | 52.52 | 97.69 | 82.83 |
| LongLive | ID framework + recent | 82.45 | 47.05 | 97.84 | 84.88 |
| Contrast | Quality | Semantic | Total |
|---|---|---|---|
| 12 frames Base | |||
| 6 12 frames |
| Generator | Contrast | Quality | Semantic | Total |
|---|---|---|---|---|
| SGF | 6 9 frames | |||
| RF-TF | 6 21 frames | |||
| SGF+ | 6 9 frames | |||
| Rolling Sink | 6 21 frames |
| Quality | DD | Subject c2c | |||||
|---|---|---|---|---|---|---|---|
| Seed | Mean | Mean | Mean | ||||
| 0 | 3 | 82.73 | — | 53.03 | — | 86.21 | — |
| 6 | 82.73 | 49.40 | 87.21 | ||||
| 9 | 82.72 | 49.47 | 87.23 | ||||
| 12 | 82.65 | 49.38 | 86.93 | ||||
| 1 | 3 | 82.83 | — | 54.18 | — | 85.91 | — |
| Generator (reference ) | New context | Quality | DD | Subject c2c | Background c2c | MS | TF |
|---|---|---|---|---|---|---|---|
| \reporthighlight SGF (9) | 6 | ||||||
| SGF (9) | 3 (self-only) | ∗ | |||||
| \reporthighlight SGF+ (9) | 6 | ||||||
| SGF+ (9) | 3 (self-only) | ∗ | ∗ | ||||
| \reporthighlight RF-TF (21) | 6 | ∗ | ∗ | ∗ | |||
| RF-TF (21) | 3 (self-only) | ∗ | ∗ |
| Generator | Quality | DD | Subject c2c | |
|---|---|---|---|---|
| RF-TF | 3 | |||
| \reporthighlight | 6 | — | — | — |
| 9 | ||||
| 12 | ||||
| 21 | ||||
| Rolling Sink | 3 |
| Seed | Context | Quality | DD | Subject c2c | Quality | Subject c2c |
|---|---|---|---|---|---|---|
| 0 | 81.28 | 67.80 | 74.69 | |||
| \reporthighlight | 81.64 | 56.58 | 79.62 | — | — | |
| 81.54 | 56.68 | 80.09 | ||||
| 81.77 | 58.08 | 79.97 | ||||
| 81.63 | 59.58 | 78.56 | ||||
| \reporthighlight 1 | 81.57 | 58.90 | 78.87 | — | — |
| Window contrast | Quality | DD | Subject c2c |
|---|---|---|---|
| 9 minus 6 | |||
| 12 minus 6 | |||
| 12 minus 9 | |||
| 21 minus 9 | |||
| 21 minus 6 |
| Committing context | Quality | DD | Subject c2c | Background c2c | Quality | Subject c2c |
| Base (21, with sink) | 83.55 | 64.70 | 84.90 | 88.55 | ||
| (self-only) | 83.51 | 76.50 | 90.37 | 92.98 | ||
| 83.64 | 67.50 | 91.22 | 93.57 | |||
| 83.71 | 64.53 | 91.76 | 94.04 | |||
| 83.73 | 64.50 | 91.86 | 94.03 | |||
| \reporthighlight | 83.71 | 63.97 | 92.02 | 94.14 | — | — |
| Committing window | Quality | Subject c2c | Background c2c | DD |
|---|---|---|---|---|
| Adaptive 3 | ∗ | ∗ | ∗ | |
| Adaptive 6 | ∗ | ∗ | ∗ | |
| Fixed 6 | ∗ | ∗ | ∗ | |
| Adaptive 6 fixed 6 | ∗ | ∗ |
| Committing context | Q | DD | Subject c2c | Background c2c |
|---|---|---|---|---|
| Global 3 + recent 18 | ∗ | ∗ | ∗ | |
| Recent 6 | ∗ | ∗ |
| Metric | at 20 frames | at 40 frames |
|---|---|---|
| Final-return similarity | ∗ | |
| Color drift | ||
| Seam ratio | ∗ | |
| 1-second similarity | ∗ |
Long-Horizon Video Generation
Long-horizon video generation creates coherent, controllable videos over extended durations despite limited context and accumulating errors.
See all 181 papers