Source-linked AI summary
Ring Forcing: Towards Precise Long-Term Memory for Autoregressive Video Diffusion
Bowen Xue, Brandon Y. Feng, Chenguo Lin, Yuchen Lin, Yujia Zeng, Lvmin Zhang, Maneesh Agrawala, Honglei Yan, Panwang Pan
TL;DR
Long-video generation lacks robust long-term memory, requiring both object permanence and the capacity to use distant history. Ring Forcing enforces long-range retrieval and combines compression, timestep composition, and sparse RoPE to expand memory under fixed sequence lengths. The method reports improved object permanence and faithful minutes-long history reuse, while extended rollouts and complex multi-object occlusions remain challenging.
Problem
Long-video generation must preserve object identity across re-entry and use information from ultra-long, distant history, but current autoregressive models struggle with both.
Method
Ring Forcing constructs ring-structured training instances that require distant-history retrieval, then combines compressed timestep-dependent history views with sparse RoPE.
Results
Ring Forcing improves object permanence and enables faithful reuse of minutes-long history, significantly outperforming state-of-the-art methods.
Takeaways & Limitations
The framework supports precise long-term memory construction and utilization for minutes-long autoregressive video generation under fixed sequence-length constraints.
Takeaways & Limitations
Causal autoregressive rollouts can still accumulate subtle appearance, motion, and scene distortions over extremely long horizons.
Abstract
from arXiv · showhide
Scaling video generation to long durations reveals a critical bottleneck: current models lack robust long-term memory. This deficiency can be studied along two critical aspects: object permanence, the ability to precisely reproduce the appearance of objects upon re-entry; and memory capacity, the ability to process ultra-long context and use information from distant history. Robust long-term memory requires both: object permanence without sufficient context handling limits the temporal scope, while long context length without permanence fails to maintain identity. To address this, we present Ring Forcing, an autoregressive video diffusion framework designed to robustly construct and precisely utilize long-term memory. Our ring-structured training strategy enforces retrieval from distant history, effectively reconciling the trade-off between strict historical adherence and generative diversity. To expand memory capacity, we introduce a compression and timestep composition strategy. Under fixed sequence length constraints, this method extends the effective historical span to minutes-long durations and achieves a comprehensive receptive field over the entire history. Furthermore, we present a sparse RoPE mechanism to enable flexible, scalable memory adaptation while fully exploiting pre-trained priors. Extensive experiments demonstrate that Ring Forcing achieves superior minutes-long coherence and object permanence, significantly outperforming state-of-the-art methods.
1 Introduction
Ring Forcing addresses long-video generation’s long-term memory bottlenecks by enforcing distant-history retrieval and expanding effective context under fixed sequence-length constraints. Its experiments report faithful minutes-long history reuse and stronger object permanence.
- Motivation: Long-video generation requires both object permanence and memory capacity to preserve object identity and use information from distant history.Object permanence concerns reproducing an object’s appearance after re-entry, while memory capacity concerns processing ultra-long context.
- Motivation: Standard linear training encourages myopic transitions from the latest frame, leaving distant history effectively unused during inference.The training distribution’s linear-narrative bias provides little incentive to retrieve information from earlier frames.
- Motivation: Frame- or token-level context management faces a trade-off between covering longer histories and controlling sequence length and computational cost.Longer receptive fields require more tokens, whereas shorter sequences shrink historical coverage.
- Ring Forcing: Ring-structured training embeds the ground-truth target clip into distant history, making long-range retrieval necessary for minimizing training error.The construction is described as “the future becomes the history,” anchoring current-target supervision in the past.
- Ring Forcing: Compression, timestep composition, and sparse RoPE expand effective context under fixed sequence-length constraints while using pretrained priors across variable histories.The approach combines complementary history views across diffusion timesteps with flexible positional encoding.
- Results: Experiments show faithful reuse of minutes-long history and improved object permanence, establishing a strong baseline for long-term video memory.The reported outcome is presented as a significant improvement over state-of-the-art methods.
2 Related Work
Related work addresses long-video generation through autoregressive frameworks, training strategies that improve rollout robustness, and context-management methods based on retrieval or compression.
- Long-video generation: Autoregressive video diffusion has become prevalent for long-video generation because bidirectional attention incurs quadratic cost when scaled to long videos.The related methods cited include SkyReels-V2, Magi-1, CausVid, and Self Forcing.
- Training paradigms: Training-based approaches improve long-video robustness through anti-drifting mechanisms, attention sinks, long-horizon training, joint denoising, and local teacher distillation.The passage cites FramePack, LongLive, Rolling Forcing, and Self-Forcing++ as examples.
- Context management: Context-management methods extend effective memory horizons either by retrieving relevant history or by compressing historical information.Examples include Field-of-View-based retrieval, semantic retrieval, and compression-based approaches in world models and video generation.
3 Method
Ring Forcing combines ring-structured training with timestep-composed compression and sparse RoPE to retrieve distant history precisely while extending effective context under fixed sequence-length constraints.
- Ring-Structured Training Strategy: Ring-structured sequences embed the ground-truth target in distant synthetic history, making long-range retrieval necessary during training.The raw video is concatenated with its reversed counterpart, then unrolled into conditioning context.
- Ring-Structured Training Strategy: Random head cropping removes boundary leakage while preserving the reversed target as a retrievable distant answer.The crop prevents the model from inferring the target from an adjacent frame at the history head.
- Ring-Structured Training Strategy: Random context drop alternates ring and standard conditioning, balancing precise historical consistency with unconstrained progression and diversity.Ring mode uses the cropped synthetic history; standard mode uses the original history alone.
- Compression and Timestep Composition: Timestep-composed compression maps long history into a bounded sequence under token budget Bmax while preserving global structure and local details.The compression operator downsamples history spatially and temporally, and timestep composition uses complementary history representations.
- Compression and Timestep Composition: High-noise timesteps use temporally dense context for global structure, whereas low-noise timesteps use temporally sparse sampling for local details.Cyclic temporal shifts vary sampling grids to provide full coverage over long histories.
- Sparse RoPE Strategy: Sparse RoPE maps compressed-history tokens to original continuous spatiotemporal coordinates, preserving relative scales under variable compression.Spatial coordinates center tokens within patches, while history time coordinates remain in the negative relative-time domain.
4 Experiments
Ring Forcing is evaluated against long-video baselines using object-permanence, 1-minute generation, ablation, and human-perception protocols. Results report strong distant-history retrieval, coherent long-duration generation, and favorable reconstruction and diversity trade-offs.
- Comparison with Baselines: The evaluation compares Ring Forcing with LongLive, LongCat, and FramePack on long-term memory and video-generation tasks.
- Comparison with Baselines: The A-D-R benchmark uses 64 cases to test whether subjects reappear consistently after disappearance, with segmentation and feature metrics isolating identity consistency.Appear and reappear segments last 5 seconds, and the first frame excludes the target subject to prevent shortcut use.
- Comparison with Baselines: Ring Forcing retrieves distant history for robust object permanence, whereas baselines often catastrophically forget temporarily disappeared targets.The reported comparison covers both A-D-R results and qualitative visualizations.
- Comparison with Baselines: Ring Forcing maintains rich motion alongside spatiotemporal consistency, avoiding the static or repetitive outputs associated with LongLive’s initial-frame attention sink.
- Comparison with Baselines: On the General Benchmark, Ring Forcing generates 60-second videos while preserving the backbone’s priors and achieves the best reported aesthetics, naturalness, and overall human preference among long-video baselines.The benchmark uses 64 persistent-subject prompts, while human evaluation includes 21 participants rating 10 blind-tested video pairs on a 0-5 scale.
- Ablation Studies: Timestep composition consistently outperforms standalone global and detail streams across compression settings without increasing sequence length.The global stream generally yields higher reconstruction SSIM than the detail-only stream, while composition combines global structure with fine-grained refinement.
- Ablation Studies: 140s of history is virtually identical to the original Wan model’s 5.4s generation, enabling efficient long-video synthesis without increasing hardware requirements.
- Ablation Studies: pdrop = 0.4 provides the reported balance between reconstruction fidelity and generative diversity, avoiding mode dropping.
5 Conclusion
Ring Forcing targets object permanence and memory capacity through unified long-term memory construction and utilization. Its training, compression, timestep composition, and sparse RoPE mechanisms support minutes-long history while balancing adherence and diversity.
- Ring Forcing addresses object permanence and memory capacity as the central bottlenecks in autoregressive long-video generation.
- Ring-structured training retrieves rich information from long-term history while reconciling historical adherence with generative diversity.
- Compression and timestep composition expand the effective historical span to minutes under fixed sequence-length constraints while covering the entire history.
- Sparse RoPE supports precise utilization and reproduction of information spanning minutes-long history.
Ring-Structured Data Construction Details
The implementation details combine randomized context cropping, probabilistic ring dropping, and dual-stream timestep composition. These choices prevent leakage, balance adherence and diversity, and process minutes-long history within fixed sequence limits.
- Random Head Cropping: Random head cropping removes a uniformly sampled prefix from the constructed context while preserving the embedded target and preventing information leakage.The crop length l is sampled from Uniform(0, s), where s is the reversed target’s starting index.
- Ring Drop Probability: The ring context is dropped with default probability pdrop = 0.4, alternating between unconstrained progression from original history and explicit target retrieval.
- Compression and Timestep Composition: Dual-stream composition uses diffusion timestep t to process minutes-long history without quadratic computational explosion.The timestep represents noise level, with t = 1 as pure noise and t = 0 as the clean data distribution.
- Compression and Timestep Composition: At high noise, the global stream retains temporal density for semantic structure and motion; at low noise, the detail stream handles high-frequency texture and local detail.The transition threshold is τ = 0.417, matching the native Wan2.2-A14B pre-training configuration.
Training Hyperparameters
The models are fine-tuned with LoRA and trained on distributed H800 infrastructure using AdamW. The reported setup uses rank 128, a 1 × 10−4 learning rate, and 7,000 optimization steps.
- Wan2.1-1.3B and Wan2.2-A14B are fine-tuned with LoRA rank r = 128 in Diffusion Transformer attention layers and FFNs.
- Training uses 32 NVIDIA H800 GPUs across four nodes with AdamW and a base learning rate of 1 × 10−4.
- Optimization runs for 7,000 steps and takes approximately 30 hours.
Computational Overhead
Ring Forcing maintains the base model’s standard token-budget constraint despite incorporating exceptionally long historical context.
- The historical context is compressed with downsampling-factor products fixed at 32, constraining its sequence length.Examples include 2 × 2 × 8 = 32 and 1 × 1 × 32 = 32.
- Under the 5.4-second generation setting at 480 × 832 resolution, total sequence length matches Wan’s standard budget of approximately 32,760 tokens.
- This design preserves computational efficiency while integrating an exceptionally long historical context.
B Limitations and Future Work
The authors identify remaining limitations in minutes-long autoregressive generation, extreme multi-object interactions, and fixed-rate memory compression. Future work targets more flexible compression and broader evaluation.
- Limitations: Minutes-long generation remains challenging because causal autoregressive rollouts can compound small appearance, motion, or scene-detail inaccuracies.Subtle local distortions may still emerge across extended temporal spans despite reduced identity drift.
- Limitations: Extreme scenes with multiple visually similar objects and dense intersecting occlusions require finer-grained memory extraction than the ring-based proxy task provides.
- Limitations: Fixed-rate compression can discard high-frequency information, occasionally affecting extremely small objects or fast transient motions.The fixed strategy bounds VRAM usage identically to the base model but lacks flexibility.
- Future Work: Future work will explore content-adaptive compression and minute-scale, multi-shot evaluation protocols.The proposed direction allocates token budgets according to spatial complexity and computational budgets.
C Additional Results
Additional qualitative results evaluate object permanence across short and extreme disappearance gaps, alongside one-minute unconstrained generation. Ring Forcing retrieves fine-grained historical details while maintaining coherent long-horizon videos.
- A-D-R Benchmark: Additional A-D-R comparisons at 1-second and 5-second gaps show Ring Forcing maintaining object permanence after subjects leave the camera frustum.The examples contrast its behavior with identity drift in autoregressive baselines.
- Extreme Gaps: At 15-second, 30-second, and 60-second gaps, Ring Forcing reliably retrieves specific fine-grained details from distant history.These stress tests examine prolonged visual disconnects without relying on short-term attention biases.
- A-D-R Benchmark: A 1-second gap comparison shows Ring Forcing recovering the identity of a hot dog more accurately than baseline models.
- A-D-R Benchmark: A 1-second gap comparison shows Ring Forcing preserving the structural and textural details of a wooden pallet after occlusion.
- A-D-R Benchmark: A 5-second gap comparison shows Ring Forcing preserving the visual features of a yellow shirt while baselines experience catastrophic forgetting.
- A-D-R Benchmark: Across a 5-second occlusion, Ring Forcing reconstructs complex human attributes while baselines hallucinate new identities.
- Extreme Gaps: At a 15-second disappearance, Ring Forcing recovers diverse objects including pizza, watermelon, and acoustic guitar beyond baseline context limits.
- Extreme Gaps: At a 30-second gap, Ring Forcing preserves temporal consistency and semantic adherence in dynamic scenes involving a burning log, basketball, and tractor.