acceptodds
Under review as a conference paper at ICLR 2027

TTT2Reel: End-to-End Test-Time Training for Long Video Generation

14 views

Abstract

Generating a long, coherent video requires preserving visual information about previously-generated frames in the window beyond a model's bounded context window. We present TTT2Reel, an end-to-end test-time training framework for chunk-wise autoregressive long video generation that stores evolving visual history in lightweight LoRA parameters. Before generating each new chunk of a video, the model updates these parameters by optimizing a flow-matching loss on the preceding chunk, while keeping the pretrained backbone frozen. We train the model through bilevel optimization, where an inner optimization loop updates the parametric memory through test-time training, and the outer loop optimizes its initial parameters to minimize subsequent prediction losses across the video. To withstand compounding errors in self-generated context at inference, we estimate model-dependent residuals online and replay them into both the memorization and prediction steps during training. Experiments on minute-long video generation show that TTT2Reel achieves the best average rank across six VBench-Long dimensions among the evaluated methods, balancing dynamic content with visual quality and consistency.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Test-Time Training, Long Video Generation, Parameter-Efficient Fine-Tuning, Meta-Learning

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_BYKuOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents TTT2Reel, an end-to-end test-time training framework for long video generation. The authors design an elegant implementation of test-time training for long video generation with LoRA and first-order model-agnostic meta-learning (FOMAML). An error replay mechanism is further proposed to mitigate error accumulation during autoregressive rollouts. Experimental results show that TTT2Reel outperforms several strong baselines after training on UltraVideo-Long, which consists of only 17k videos.

Critical strengths

The implementation of TTT in this paper is elegant and practical. Unlike common designs that insert a new TTT layer (instantiated via an MLP) or replace the pre-trained attention layers of DiT, the approach in this paper uses a lightweight LoRA module. Such a design protects the original performance of base models and delivers better generalizability than prior work (e.g., [1]). And the performance achieved is also impressive, especially the results for ultra-long (3-minute) video rollouts.

[1] Dalal K, Koceja D, Xu J, et al. One-minute video generation with test-time training[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 17702-17711.

Critical weaknesses
  1. The contribution of the proposed TTT memory is unclear, as the main performance gains appear to come from error replay and no direct evidence demonstrates improved long-range memory retention or retrieval.

  2. The computational cost of end-to-end bilevel training is not reported, making the method’s practical efficiency and comparisons with other approaches difficult to assess.

Questions for authors

The experimental results do not fully support the authors’ claims. The reported VBench results and qualitative demonstrations primarily evaluate the visual quality of the generated videos. However, as stated in line 75, TTT2Reel is intended to address both the exposure-bias problem and the capacity problem by representing completed chunks with limited capacity.

The ablation results also leave the contribution of the TTT component unclear. Table 3 suggests that the gains in most visual-quality and consistency metrics are driven more by error replay than by test-time training. For example, removing error replay decreases Aesthetic Quality from 58.90 to 56.80 and Imaging Quality from 67.64 to 64.57, whereas removing test-time training decreases them to 57.90 and 66.12, respectively. Moreover, when end-to-end training and error replay are enabled, adding test-time training slightly decreases Subject Consistency and Background Consistency. As currently presented, the VBench metrics do not provide sufficient evidence for the effectiveness of TTT, since the prompts and camera motions appear relatively simple and may not necessarily require a memory mechanism like TTT.

The ablation also suggests that end-to-end training is important for the reported performance. However, the proposed bilevel training procedure may incur substantially higher training costs than the baselines used for comparison, which should be discussed to ensure a fair comparison.

I would appreciate it if the authors could address the following questions:

  1. Could the authors provide direct evidence that the proposed memory can retain and retrieve information from early chunks over long horizons, rather than evaluating only the final video quality? If not, the authors may wish to soften the described contributions.

  2. Could the authors report the average time per optimization step and the total GPU-hours?

  3. Please provide metrics to measure anti-drifting performance in the ablation study, for example, by computing the difference between the VBench metrics of the last chunk and those of the first chunk.

5 comments

We thank the reviewer for recognizing the elegance and practicality of our LoRA-based TTT framework, particularly its ability to preserve the pretrained backbone, as well as the strong performance demonstrated on long-video generation. We appreciate the constructive feedback and address the concerns below.

(W1 and Q1) Evidence of long-term visual memory retention and retrieval.

As requested, we additionally evaluate memory retention by measuring the flow-matching loss on the first chunk after successive LoRA updates, and add a control in which the first chunk is never memorized (Fig. 11, Appendix F.2). With memorization, the loss remains stable, with relative changes within 1.7%, while the control rises monotonically to +4.0%. The memorized curve stays below the control throughout, indicating that earlier visual information is preserved despite subsequent updates.

Qualitative examples in Fig. 12 (Appendix F.3) and on our demo page further demonstrate long-horizon visual memory retention and retrieval. Specifically, the astronaut example (One Minute Video Rollouts) shows Earth moving out of view and reappearing with the correct spatial relationship, while the fox example (Ultra-Long Video Rollouts) exhibits similar scene-revisitation behavior.

(W2 and Q2) Computational cost of bilevel training.

As requested, we additionally report the training cost of TTT2Reel, including average wall-clock time per optimization step and total GPU-hours. Training TTT2Reel at the 5B scale requires 91.2 seconds per step on 8 NVIDIA H100 GPUs (96GB) and 374 GPU-hours in total. For comparison, standard SFT that randomly crops a 5-second clip from each UltraVideo-Long video and trains for the same number of steps requires 14.0 seconds per step and 57 GPU-hours in total under the same hardware configuration. Although TTT2Reel introduces additional computational overhead during training, it achieves consistent performance gains across 6 dimensions on VBench-Long as shown in the table below, without increasing inference cost. In addition, training the 14B variant of TTT2Reel takes 381.9 seconds per optimization step on 16 NVIDIA H100 GPUs (96GB) and 683 GPU-hours in total.

Dynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
SFT61.56±1.2056.47±0.7961.50±1.5496.14±0.5295.94±0.6397.75±0.69
TTT2Reel73.57±0.7758.90±0.4867.64±0.3196.88±0.0896.77±0.1398.64±0.04

(Q3) Anti-drifting performance.

As requested, we evaluate anti-drifting performance by measuring the absolute difference in VBench-Long metrics between the first and last video chunks. The results below (mean ± std) show that removing TTT, bilevel training, or the error replay strategy increases the gaps in most metrics. Notably, removing bilevel training leads to larger differences in visual quality and consistency metrics than removing the error replay, highlighting the importance of end-to-end meta-learning in mitigating long-horizon drift.

Test-Time TrainingEnd-to-End TrainingError ReplayDynamic Degree↓Aesthetic Quality↓Imaging Quality↓Subject Consistency↓Background Consistency↓Motion Smoothness↓
✔✔✔25.9±4.33.83±0.394.78±0.540.99±0.080.78±0.080.49±0.05
-✔✔26.4±4.45.75±0.537.25±0.731.11±0.120.82±0.080.69±0.12
✔-✔32.9±4.56.81±0.638.47±0.871.45±0.200.90±0.090.88±0.18
--✔31.9±4.84.71±0.415.41±0.591.22±0.110.85±0.070.65±0.08
✔✔-25.0±4.36.53±0.657.97±0.841.34±0.190.84±0.070.71±0.08

The effect of TTT on consistency metrics. We also observe that removing TTT produces more static videos, leading to higher subject and background consistency scores at the expense of motion dynamics, as demonstrated by the lower Dynamic Degree score in Table 3.

Reviewer_BYKuOpenReview ↗

Cost-Benefit of TTT2Reel: Requesting a Controlled Long-Term Memory Benchmark

Thank the authors for the constructive response and for adding the qualitative examples. I agree that the astronaut and fox examples show some scene-revisitation ability, and I appreciate this effort.

However, a remaining concern is whether the additional training overhead is justified by the observed memory benefits. The current evaluation compares models under an equal number of optimization steps; however, given that TTT2Reel requires 91.2 seconds per step compared to 14.0 seconds for standard SFT (roughly a 6.5× increase), this comparison is not compute-matched. Under equivalent GPU-hours, SFT would accrue substantially more optimization steps. Consequently, it would be beneficial to provide either an SFT baseline trained under matched compute budgets (GPU-hours) or a targeted, controlled memory benchmark.

While a full compute-matched evaluation might be computationally demanding, a smaller-scale controlled experiment could still offer valuable insights. For instance, on a curated subset of approximately 16 video clips (each around 30 seconds) where a subject exits the frame for a sustained interval and later re-enters, one could compute the face-embedding cosine similarity (e.g., via ArcFace) between pre-departure and post-return frames. Isolating the memory benefit could be achieved by comparing full TTT with a non-TTT or sliding-window variant under comparable settings, alongside LongLive 2.0 and other TTT baselines if applicable. Reporting quantitative metrics, failure modes, and representative visualizations would significantly strengthen the evaluation. Furthermore, while the first-chunk flow-matching loss serves as a useful auxiliary signal for monitoring forgetting, it does not constitute a direct metric for long-term retrieval.

If such an analysis demonstrates a clear and statistically significant memory advantage, I would be pleased to raise my score. Otherwise, I encourage the authors to moderate their claims regarding long-term visual memory and reposition the core contribution primarily as an efficient LoRA-based TTT framework—a contribution whose value I continue to recognize.

We thank the reviewer for the constructive follow-up and the opportunity to further substantiate the benefits of TTT2Reel through compute-matched comparisons and controlled long-term memory evaluations.

(Q1) Comparison with standard SFT under matched compute budgets.

We thank the reviewer for highlighting the discrepancy in training compute. We agree that standard SFT can benefit from substantially more optimization steps under an equivalent compute budget. As requested, we additionally train the SFT baseline on 16 NVIDIA H100 GPUs for 7,300 steps, totaling 374.6 GPU-hours, closely matching the training cost of TTT2Reel (374.1 GPU-hours).

The following table shows that extended SFT improves video generation performance on most metrics, but still lags behind TTT2Reel across all 6 VBench-Long dimensions, which justifies the additional training overhead of our approach.

H100 GPU HoursDynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
SFT57.061.56±1.2056.47±0.7961.50±1.5496.14±0.5295.94±0.6397.75±0.69
Compute-matched SFT374.669.44±0.9655.83±0.6765.69±0.5496.12±0.3396.06±0.4097.83±0.12
TTT2Reel374.173.57±0.7758.90±0.4867.64±0.3196.88±0.0896.77±0.1398.64±0.04

(Q2) A targeted, controlled memory benchmark

We thank the reviewer for suggesting this controlled memory benchmark. As requested, we design 16 prompt streams following the format of Table 7 (Appendix F.3), each featuring a human character with a clearly visible face who exits the scene for a sustained interval before re-entering. Using these prompts, we compare TTT2Reel against LongLive-2.0, two ablation variants, TTT2Reel-Static (without TTT) and TTT2Reel-Reset (resetting LoRA weights to their learned initialization after each chunk), and two TTT video methods, TTTVideo [1] and LaCT [2]. Note that TTTVideo and LaCT released code without training data or checkpoints. So we retrained both using their official implementations, but with our own training data.

Quantitative metrics. As requested, we report the cosine similarity between ArcFace embeddings of pre-departure and post-return frames, alongside an instruction-following metric that measures whether the generated video follows the intended sequence of events (the character enters, leaves, the scene remains empty, and the character returns). TTT2Reel achieves the highest scores on both metrics, demonstrating strong identity consistency and prompt adherence in scenarios requiring long-term visual memory.

MethodsCosine Similarity ↑Instruction Following (%) ↑
LongLive-2.00.447±0.02510.4±1.5
TTT2Reel-Static0.332±0.03878.1±2.6
TTT2Reel-Reset0.341±0.04375.0±2.6
TTTVideo0.268±0.01232.3±1.5
LaCT0.243±0.00938.5±1.5
TTT2Reel0.632±0.01587.5±2.6

Failure mode analysis. Although LongLive-2.0 attains the second-highest cosine similarity, it exhibits the lowest instruction-following rate (10.4%). Inspecting its videos shows that the main character rarely leaves the scene. Its high similarity therefore reflects a person who always stays in the scene rather than a face recalled after a sustained absence, so it does not indicate long-horizon visual memory. Both TTT2Reel-Static and TTT2Reel-Reset generally follow the prompt streams, but they often regenerate a different person instead of recalling the face since they lack long-term memory. TTTVideo and LaCT follow the prompt streams in only 32.3% and 38.5% of the examples, far below TTT2Reel’s 87.5%. In particular, TTTVideo often changes the character’s identity even before departure, while LaCT tends to collapse into flat, saturated colors after ~15 seconds.

Representative visualizations. We have expanded Appendix F.3 of the revised manuscript to include two representative examples (Figure 13), illustrating the advantages of TTT2Reel and the failure modes of baselines.

[1] Dalal, Karan, et al. "One-minute video generation with test-time training." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.
[2] Zhang, Tianyuan, et al. "Test-time training done right." International Conference on Learning Representations. Vol. 2026. 2026.

Reviewer_Y3HnOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TTT2Reel, which uses Test-Time Training to continuously encode visual information from previous video chunks into LoRA parameters. It employs bilevel optimization to learn a LoRA initialization that supports memory updates and subsequent chunk prediction, along with Error Replay to mitigate exposure bias. Experiments show that TTT2Reel achieves a good balance between motion dynamics, visual quality, and consistency in minute-long video generation, with the best average rank on VBench-Long.

Critical strengths
  1. The method stores historical visual information by dynamically updating LoRA parameters, keeping memory overhead independent of video length. This is an interesting approach to long-video generation.
  2. Bilevel optimization learns a LoRA initialization that makes Test-Time Training useful for future chunk prediction, rather than simply fitting previous chunks. Combined with Error Replay to mitigate exposure bias, the overall framework is well designed.
  3. The experiments cover both T2V and I2V generation, including videos up to four minutes long. The paper also provides efficiency comparisons and ablation studies, making the results generally convincing.
Critical weaknesses
  1. The effectiveness of parametric memory is mainly evaluated through video quality metrics, with limited direct evidence that long-term visual information is actually retained.
  2. Error Replay relies on residuals estimated from single-step predictions, but comparisons with other error perturbation strategies and more detailed ablations are missing.
  3. Inference efficiency is mainly reported using normalized compute overhead, without direct comparisons of actual generation time or peak GPU memory usage.
  4. Comparisons with some recent long-video generation methods are missing, making it difficult to fully assess the method's competitiveness.
Questions for authors
  1. In Table 3, removing Test-Time Training slightly improves Subject Consistency and Background Consistency. How can the authors verify that the LoRA parameters actually retain long-term visual information? Could they include experiments involving character reappearance or scene revisitation, or compare resetting the LoRA parameters at each chunk with continuously updating them?
  2. The paper uses FOMAML to learn the LoRA initialization, but the current ablations mainly compare models with and without end-to-end meta-training. How much does bilevel optimization improve over standard supervised fine-tuning (SFT) of the LoRA initialization, followed by the same Test-Time Training procedure?
  3. Table 2 compares efficiency using Normalized Compute across different backbones and inference settings. Could the authors provide comparisons of actual generation time, FPS, and peak GPU memory usage?
  4. Could the authors include comparisons with recent long-video generation methods such as Self Gradient Forcing, Self-Forcing++, and Context Forcing? These methods also focus on long-term consistency and error accumulation, so comparisons under a unified long-video evaluation setting would be useful.
  5. Since the LoRA memory has a fixed capacity, could newly generated content gradually overwrite earlier information as the video grows longer? Could the authors evaluate different LoRA ranks or longer generation horizons to better understand the memory capacity and potential forgetting issues?
3 comments

We thank the reviewer for recognizing the novelty and thoughtful design of TTT2Reel, as well as its comprehensive experimental evaluation and convincing results. We also appreciate the constructive feedback and address the concerns in detail below.

(W1 and Q1) The reviewer is interested in results on long-term visual memory retention.

Effect of TTT on consistency-related metrics. We thank the reviewer for pointing out this observation. Our inspection of the generated videos reveals that removing TTT produces more static videos than full TTT2Reel, leading to higher subject and background consistency scores at the expense of motion dynamics, as demonstrated by the lower Dynamic Degree score.

Evidence of long-horizon visual memory retention. We evaluate memory retention by measuring the flow-matching loss on the first chunk after successive LoRA updates, and add a control in which the first chunk is never memorized (Fig. 11, Appendix F.2). With memorization, the loss remains stable, with relative changes within 1.7%, while the control rises monotonically to +4.0%. The memorized curve stays below the control throughout, indicating that earlier visual information is preserved despite subsequent updates.

Qualitative examples in Fig. 12 (Appendix F.3) and on our demo page further demonstrate long-horizon visual memory retention. Specifically, the astronaut example (One Minute Video Rollouts) shows Earth moving out of view and reappearing with the correct spatial relationship, while the fox example (Ultra-Long Video Rollouts) exhibits similar scene-revisitation behavior.

Continuous vs. Reset LoRA Updates. We thank the reviewer for suggesting this comparison. We extend our ablation study with a baseline that resets the LoRA parameters to their learned initialization at each chunk, preventing the accumulation of parametric memory across chunks. As shown in the following table, continuous updates outperform the resetting baseline across four of six metrics, particularly in motion dynamics and visual quality. Resetting slightly improves subject and background consistency due to reduced motion, consistent with our observations when disabling TTT. We further provide qualitative comparisons on two 60-second rollouts in Figure 13 (Appendix F.7). This additional comparison isolates the benefit of accumulating LoRA updates across chunks, providing further evidence that persistent parametric memory contributes to long-horizon video generation beyond independent per-chunk adaptation.

Dynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
Reset Updates72.35±0.3858.10±0.2766.56±0.3496.92±0.0396.93±0.0898.53±0.03
Continuous Updates73.57±0.7758.90±0.4867.64±0.3196.88±0.0896.77±0.1398.64±0.04

(W2 and Q4) Error perturbation strategies and additional baselines.

Error perturbation strategies. We compare the error replay strategy against the synthetic noise, blur, and saturation perturbations adopted in Helios. The results below show that the error replay consistently outperforms these synthetic perturbations across all six VBench-Long metrics, demonstrating its leading effectiveness in mitigating accumulated generation errors.

Dynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
Noise/Blur/Saturation67.97±1.2555.21±0.5762.21±0.7396.33±0.1096.27±0.1198.18±0.16
Error replay73.57±0.7758.90±0.4867.64±0.3196.88±0.0896.77±0.1398.64±0.04

Additional long-video baselines. We evaluate TTT2Reel against Context Forcing. TTT2Reel outperforms Context Forcing across all six metrics, with particularly large gains in Dynamic Degree and consistency-related metrics. Our original evaluation in Table 1 also includes Rolling Forcing, LongLive, and LongLive-2.0, which focus on long-horizon consistency and error accumulation.

Dynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
Context Forcing49.51±0.9858.20±0.1967.58±0.1594.88±0.0995.34±0.0698.57±0.04
TTT2Reel73.57±0.7758.90±0.4867.64±0.3196.88±0.0896.77±0.1398.64±0.04

(W3 and Q3) Computational efficiency.

We additionally report the actual wall-clock time, FPS, and peak memory cost per GPU of TTT2Reel and baselines. For a fair comparison, we compare TTT2Reel (5B) against the feature-based memory method, LongLive-2.0 (5B), and compare TTT2Reel (14B) with two frame-based memory methods, FramePack (13B), Helios-Base (14B). Apart from similar model sizes, all baselines retain the full history in their context and use the same sampling configuration (50 steps without CFG). We evaluate 25-second video generation, as longer videos exceed the 96 GB memory capacity of a single H100 GPU for some baselines.

MethodModel SizeWall-Clock Time (s) ↓FPS ↑Peak Memory/GPU (GB) ↓Dynamic Degree ↑Subject Consistency ↑Motion Smoothness ↑
LongLive-2.05B732.00.8377.866.38±0.3597.29±0.0698.14±0.09
TTT2Reel5B1093.00.5638.270.62±0.5697.72±0.0798.65±0.05
MethodModel SizeWall-Clock Time (s) ↓FPS ↑Peak Memory/GPU (GB) ↓
FramePack13B12600.00.0366.5
Helios-Base14B9742.00.0472.0
TTT2Reel14B2609.50.1563.1

Despite test-time backpropagation, TTT2Reel achieves the lowest peak per-GPU memory usage on both 5B and 14B and outperforms both frame-based memory methods in inference speed, although it remains slower than the feature-based memory methods. Importantly, its fixed-size parametric memory state ensures that memory consumption does not grow with the video length.

(Q2) Bilevel optimization vs. standard SFT.

As requested, we compare FOMAML-based bilevel optimization against standard supervised fine-tuning (SFT), both followed by the same test-time training procedure. For bilevel optimization, we reuse the results of TTT2Reel without the error replay from our ablation study (Table 3). For SFT, we randomly crop a 5-second clip from each UltraVideo-Long video and train the model for the same number of steps. The results below on 1-minute video generation (mean ± std) show that bilevel optimization consistently outperforms SFT across all six VBench-Long metrics under the TTT setting, demonstrating its effectiveness in learning a better initialization for TTT than the standard SFT.

Dynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
SFT61.56±1.2056.47±0.7961.50±1.5496.14±0.5295.94±0.6397.75±0.69
Bilevel Optimization72.70±0.5156.80±0.4464.57±0.4296.69±0.1196.66±0.0998.46±0.07

(Q5) Memory capacity.

Memory retention. As discussed in W1, we measure the flow-matching loss on the first video chunk after successive LoRA updates on subsequent chunks (Fig. 11, Appendix F.2). The results suggest that memorizing chunks with similar content can reinforce earlier memories, while continued updates gradually introduce forgetting. Nevertheless, the percentage change in inner-loop loss remains small, up to 2.1% as shown in Figure 11.

LoRA rank. We additionally evaluate LoRA rank 256 on 1-min video generations. The results show that increasing LoRA capacity improves all six VBench-Long metrics, particularly Aesthetic Quality and Imaging Quality.

LoRA RankDynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
12873.57±0.7758.90±0.4867.64±0.3196.88±0.0896.77±0.1398.64±0.04
25673.60±0.5859.70±0.3068.35±0.2897.04±0.0696.89±0.0898.77±0.05

Longer generation horizons. We further extend the evaluation to 6-minute videos using VisionReward prompts and Gemini-3.7-Flash as the judge. TTT2Reel maintains the highest pass rate across all evaluated durations, although its advantage narrows as video length increases.

Methods6 min
FramePack22.4±1.5
LongLive-2.026.6±0.6
Helios-Base36.8±1.1
TTT2Reel37.9±1.2
Reviewer_boJoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles minute-long video generation with chunk-wise autoregressive diffusion models. Instead of storing past frames or a KV cache, TTT2Reel writes video history into LoRA weights: before generating each chunk, it takes a few gradient steps on the previous chunk with a flow-matching loss, while the backbone stays frozen. The LoRA initialization is meta-learned with first-order MAML so that these test-time updates actually help next-chunk prediction. To reduce drift from self-generated context, the authors also replay model-estimated errors into both the memorization and prediction steps during training. On Wan2.2-5B and Wan2.1-14B, the method gets the best average rank on VBench-Long and wins Gemini-judged head-to-head comparisons against several recent long-video baselines, with modest compute overhead.

Critical strengths
  1. Simple and practical. LoRA on a frozen backbone, no architectural changes, and fixed-size memory. The method scales to 14B models at moderate training cost, with small overhead relative to the base model’s chunk-wise generation.
  2. Well-motivated components. The ablation shows naive TTT hurts without meta-learning, which makes the meta-learned initialization a useful finding. Error replay further improves robustness to self-generated context.
  3. Broad evaluation and clear presentation. Experiments cover T2V and I2V, two backbones, and length extrapolation up to 4 minutes, with favorable results on the reported benchmarks. The paper is clear and gives full implementation details.
Critical weaknesses
  1. Missing related work and baselines. The paper does not compare against prior TTT video methods (TTTVideo[1]; LaCT[2]; SlowFast-VGen[3]), and it ignores SSM-based memory (VideoSSM[4]), which also keeps a fixed-size state but updates it with a single forward pass. Its related work also omits several recent lines on long-video memory, such as retrieval or memory-bank methods (WorldMem; Mixture of Contexts), and long-context or rollout-based training (Diffusion Forcing; FAR; Self-Forcing++; Context Forcing) Novelty is limited, since the method mainly combines ideas from TTT-E2E, PERK, and SVI.
  2. Unclear efficiency. No wall-clock time or FPS is reported. The 1.17× overhead looks small only because the base model uses 50 denoising steps; with few-step distilled models, test-time backprop would be far more costly. Table 2 normalizes each method to its own base model, so the numbers cannot be compared across methods. Backprop at inference also blocks streaming, and its memory cost is not fully reported.
  3. Weak evaluation protocol. Table 1 compares a 50-step 5B model against 1.3B real-time models, and model sizes are mixed throughoutput. The I2V results rely on a single LLM judge with an uncalibrated pass rate and no human study. No variance is reported, though many gaps are below one point. Core claims are under-supported. Removing TTT slightly improves consistency (Table 3), so LoRA may act as a better adapter rather than a memory. The anchor and local frames confound the results, and there is no retrieval-style test (e.g., revisiting a scene).

[1] Dalal, Karan, et al. "One-minute video generation with test-time training." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

[2] Zhang, Tianyuan, et al. "Test-time training done right." International Conference on Learning Representations. Vol. 2026. 2026.

[3] Hong, Yining, et al. "Slowfast-vgen: Slow-fast learning for action-driven long video generation." International Conference on Learning Representations. Vol. 2025. 2025.

[4] Yu, Yifei, et al. "Videossm: Autoregressive long video generation with hybrid state-space memory." arXiv preprint arXiv:2512.04519 (2025).

Questions for authors

See weekness.

7 comments

We thank the reviewer for recognizing the simplicity and practicality of TTT2Reel, particularly its approach of consolidating video history into LoRA parameters for subsequent video chunk generation, as well as its favorable performance against existing long-video generation baselines. We also appreciate the reviewer’s constructive feedback and address the concerns in detail below.

(W1) Additional baselines, related work, and novelty.

Prior TTT baselines. We thank the reviewer for suggesting these comparisons. TTTVideo and LaCT released code without training data or checkpoints. So, we retrained both using their official implementations, but our own training date. We would also be happy to compare with SlowFast-VGen and VideoSSM once their codebases become publicly available.

The results on VBench-Long (mean ± std over three random seeds) are reported below.

Dynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
TTTVideo71.54±0.6254.17±0.3460.11±0.1795.54±0.0495.38±0.0898.33±0.05
LaCT63.68±0.6348.67±0.2356.36±0.1293.34±0.0394.85±0.1195.96±0.05
TTT2Reel73.57±0.7758.90±0.4867.64±0.3196.88±0.0896.77±0.1398.64±0.04

TTT2Reel outperforms both methods across all six metrics. Beyond these performance gains, TTT2Reel preserves the pretrained backbone and updates only lightweight LoRA parameters. In contrast, TTTVideo and LaCT replace attention layers with TTT layers, introducing architectural modifications that may disrupt pretrained capabilities.

Related work. We have expanded the related work section to discuss the suggested long-video memory and long-context or rollout-based training methods in the revised manuscript, with all changes highlighted in blue.

Novelty. While TTT2Reel draws inspiration from these methods, our contribution goes beyond simply combining these techniques. We introduce a unified framework for autoregressive video diffusion that meta-learns to consolidate generated video history into parametric memory while mitigating accumulated errors. Unlike prior TTT video methods and long-video generation systems, TTT2Reel enables long-horizon video generation with a frozen backbone and fixed-size memory, as validated by our experiments.

(W2) The reviewer is interested in additional analysis on computational efficiency and wonders whether backprop at test time blocks streaming.

Comparisons of wall-clock time, FPS, and peak memory cost. We report the wall-clock time, FPS, and peak memory cost per GPU of TTT2Reel and baselines. For a fair comparison, we compare TTT2Reel (5B) against the feature-based memory method, LongLive-2.0 (5B), and compare TTT2Reel (14B) with two frame-based memory methods, FramePack (13B) and Helios-Base (14B). Apart from similar model sizes, all baselines retain the full history in their context and use the same sampling configuration (50 steps without CFG). We evaluate 25-second video generation, as longer videos exceed the 96 GB memory capacity of a single H100 GPU for some baselines (though not for TTT2Reel).

MethodModel SizeWall-Clock Time (s) ↓FPS ↑Peak Memory/GPU (GB) ↓Dynamic Degree ↑Subject Consistency ↑Motion Smoothness ↑
LongLive-2.05B732.00.8377.866.38±0.3597.29±0.0698.14±0.09
TTT2Reel5B1093.00.5638.270.62±0.5697.72±0.0798.65±0.05
MethodModel SizeWall-Clock Time (s) ↓FPS ↑Peak Memory/GPU (GB) ↓
FramePack13B12600.00.0366.5
Helios-Base14B9742.00.0472.0
TTT2Reel14B2609.50.1563.1

Despite test-time backpropagation, TTT2Reel achieves the lowest peak per-GPU memory usage on both 5B and 14B and outperforms both frame-based memory methods in inference speed. Importantly, its fixed-size parametric memory state ensures that memory consumption does not grow with the video length.

Backprop blocks streaming. We acknowledge that sequential memorization and prediction introduce additional latency. To address this, we implement a parallel pipeline using two GPUs. While one GPU memorizes chunk $k$, the other generates chunk $k+1$ using the parametric memory up to chunk $k-1$ and the local context from chunk $k$. Crucially, this reduces wall-clock time from 184.1s to 122.6s (1.50x speedup) for TTT2Reel-5B and from 318.0s to 182.3s (1.74x speedup) for TTT2Reel-14B, without increasing peak per-GPU memory usage. Notably, the parallelized TTT2Reel-5B achieved the second-highest FPS among all compared methods, demonstrating that test-time gradient backpropagation does not necessarily block streaming video generation.

(W3) Evaluation protocol and evidence of parametric memory.

Baseline comparability. We acknowledge that the compared methods differ in model size and sampling configurations. We evaluate baselines in the setting in which they perform the best as aligning the sampling configuration actually hurts certain methods’ performance (e.g., LongLive performs worse with 50 denoising steps than with 4 steps). We prioritize baselines with comparable settings while covering diverse memory mechanisms. However, most prior methods release only a single model size, making strictly matched comparisons difficult.

Dynamic Degree ↑Aesthetic Quality ↑Imaging Quality ↑Subject Consistency ↑Background Consistency ↑Motion Smoothness ↑
LongLive (4 steps)44.5658.1767.5697.1395.8998.61
LongLive (50 steps)54.0754.9953.1795.3295.8898.09

LLM judges and human evaluation. The original VisionReward benchmark uses 29 binary questions with aggregation weights calibrated to human preferences based on CogVLM2’s outputs. Since these weights cannot easily transfer to other judges, we report the unweighted pass rate. Importantly, all 29 weights are positive, meaning that each “yes” contributes positively to the original quality score, supporting the use of the unweighted pass rate.

To improve evaluation reliability, we additionally evaluate with Qwen3.8-Max and Gemini 4 Argon. The results based on the majority vote over 3 judges confirm TTT2Reel's leading performance, consistent with the original Gemini-3.7-Flash evaluation.

Pass Rate ↑Instruction-Following ↑
MAGI-1-4.5b46.3%±0.946%±4.4
Wan2.146.6%±1.054%±3.1
LongLive2-5b57.9%±1.240%±5.2
FramePack67.0%±1.084%±2.0
SkyReels68.0%±1.084%±2.3
Helios-Base71.0%±0.986%±2.1
TTT2Reel74.2%±0.988%±1.7

We further conduct a human study with five annotators, each evaluating 200 video pairs (50 per baseline). Each annotator spent a full day completing the annotations. Ties were split equally between wins and losses when computing preference rates. TTT2Reel is preferred over all four baselines, with preference rates ranging from 77.2% to 94.6%.

ComparisonTTT2Reel PreferredBaseline Preferred
vs. FramePack88.6%11.4%
vs. Helios-Base77.2%22.8%
vs. SkyReels-V284.8%15.2%
vs. LongLive-2.094.6%5.4%

Variance. We have added the standard deviation to the ablation study (marked in blue) in the revised manuscript (Table 3).

Effect of TTT on consistency-related metrics and persistent or local anchors. We find that removing TTT at inference produces mostly static videos, leading to higher subject and background consistency scores at the expense of motion dynamics, as demonstrated by its lower Dynamic Degree score. To isolate the contribution of parametric memory from the persistent anchor and local context, we compare against Wan2.2-TI2V$^\dagger$ (Table 1), which retains the same anchor and local frames as TTT2Reel but without test-time training. TTT2Reel outperforms this controlled baseline across all six metrics (Table 1), with qualitative comparisons provided on the demo page (Same Prompt, Four Rollout Strategies).

Evidence of parametric memory and the retrieval-style test. We additionally evaluate memory retention by measuring the flow-matching loss on the first chunk after successive LoRA updates, and add a control in which the first chunk is never memorized (Fig. 11, Appendix F.2). With memorization, the loss remains stable, with relative changes within 1.7%, while the control rises monotonically to +4.0%. The memorized curve stays below the control throughout, indicating that earlier visual information is preserved despite subsequent updates.

Qualitative examples in Fig. 12 (Appendix F.3) and on our demo page further demonstrate long-horizon visual memory retention. In particular, the astronaut example (One Minute Video Rollouts) shows Earth moving out of view and reappearing with the correct spatial relationship, while the fox example (Ultra-Long Video Rollouts) exhibits similar scene-revisitation behavior. Together, these results provide additional evidence that the LoRA parameters serve as persistent video memory rather than merely a static adapter.

Comments on the paper

1 comment
AuthorsOpenReview ↗

AC Summary

General response

We thank all three reviewers for their detailed and constructive feedback, and for recognizing the novelty, simplicity, and practicality of TTT2Reel, as well as its strong performance on long-video generation.

During the rebuttal and discussion period, we conducted extensive additional experiments to address the reviewers' concerns. Here, we summarize our main findings, with detailed results provided in our individual responses to the reviewers.

1. Evidence of long-term visual memory retention and retrieval

To further substantiate that the continuously updated LoRA parameters serve as persistent visual memory, we conducted several additional controlled experiments:

  • Memory retention after successive updates (all reviewers). We track the memorization loss of the first video chunk throughout successive memory updates, alongside a control in which the first chunk is never memorized. The results demonstrate that earlier information is retained in LoRA weights despite continued updates.
  • Continuous vs. reset LoRA updates (Reviewer Y3Hn). We demonstrate the benefits of accumulating parametric memory across chunks over independent per-chunk adaptation.
  • Controlled memory benchmark (Reviewer BYKu). We constructed 16 prompt streams featuring human characters leaving and subsequently returning to the scene. TTT2Reel substantially outperforms all baselines, including its static and reset variants, in both appearance preservation and instruction following.

Together with the controlled ablations in our original submission and the qualitative scene-revisitation examples, these results provide strong evidence of long-term visual memory retention and retrieval.

2. Effectiveness and computational cost of bilevel training

We conducted additional experiments to investigate the necessity of our end-to-end training framework and whether its benefits justify the additional training cost.

  • Bilevel optimization vs. standard SFT (Reviewer Y3Hn and Reviewer BYKu). We compare bilevel optimization against standard SFT under both matched optimization steps and matched training compute budgets. TTT2Reel consistently outperforms SFT across all six VBench-Long metrics, demonstrating that its performance gains cannot simply be attributed to increased training compute.

3. Computational efficiency at inference

To address the concerns of Reviewer Y3Hn and Reviewer boJo regarding test-time training, we conducted the following additional studies:

  • Computational efficiency comparison. We benchmark actual wall-clock time, FPS, and peak per-GPU memory usage against representative frame-based and feature-based memory methods, demonstrating the compute and memory efficiency of TTT2Reel despite test-time backpropagation.
  • Parallelization of memorization and prediction. We implement a two-GPU parallelized pipeline that substantially reduces generation latency without increasing peak per-GPU memory usage.
  • Efficiency under few-step sampling. We further demonstrate that, under 4-step sampling, parallelized TTT2Reel achieves comparable wall-clock time and DiT FPS to an optimized LongLive-2.0, with better generation quality on four of six metrics, at the cost of an additional GPU.

These results highlight the potential of gradient-based parametric memory for computationally efficient long-horizon video generation.

4. Novelty, additional baselines, and evaluation reliability

Following the suggestions of Reviewer Y3Hn and Reviewer boJo, we further expanded our experimental evaluation and clarified our contributions.

  • Additional baselines. TTT2Reel outperforms two TTT-based video generation methods (TTTVideo and LaCT), a Temp-LoRA baseline adopted by SlowFast-VGen, and a recent long-video generation method (Context Forcing) across all six VBench-Long metrics.
  • Novelty. We clarify the distinction between TTT2Reel and existing methods, highlighting its advantages of preserving the pretrained backbone, adopting gradient-based memory updates, and explicitly meta-learning the LoRA initialization for both memory consolidation and subsequent chunk prediction.
  • Multiple VLM judges. We extend our VisionReward evaluation with two additional VLM judges, confirming our original findings.
  • Human evaluation. We collected 1,000 human preference judgments. TTT2Reel is consistently preferred over all four competing methods, with preference rates ranging from 77.2% to 94.6%.

Takeaway

Overall, we believe that these extensive additional experiments substantially strengthen the contributions of our work, providing compelling evidence of long-term visual memory retention and retrieval, the effectiveness of end-to-end meta-learning under matched compute budgets, and the competitive performance and underestimated efficiency of TTT2Reel.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.