BIFE: Better Interaction, Fewer Errors for Minute-Long Video Generation

1 Introduction

• Sliding-window KV caches discard earlier tokens at capacity, weakening long-range interactions, while accumulated prediction errors degrade quality and consistency over time.The reported effects include subject and background inconsistency, color drift, and visual distortions.Source: §1 · p. 2
• BIFE combines a semantic sparse KV cache that retrieves relevant history with Block Forcing, which regularizes cross-block coherence.The cache stores sparse KV from previous blocks; Block Forcing addresses the long-range training–inference gap.Source: §1 · pp. 2–3
• InterVBench comprises 1,000 minute-long videos annotated every 2–5 seconds and introduces Video Drift Error metrics for long-horizon evaluation.The metrics assess long-horizon interaction, temporal consistency, and visual fidelity.Source: §1 · pp. 2–3
• 22.2% on VDE-Subject and 19.4% on VDE-Clarity are BIFE’s reported improvements over baselines on InterVBench.The introduction reports experiments on InterVBench and VBench.Source: §1 · p. 3

2 Related Work

• Single-shot methods include autoregressive and autoregressive diffusion approaches, while multi-shot systems handle scene transitions and movie-style methods compose blocks into a coherent narrative.Autoregressive diffusion models denoise within blocks and condition causally across blocks; approaches differ in sequential versus symmetric schedules.Source: §2 · p. 3
• Diffusion Transformers use bidirectional attention without KV caching, while autoregressive frameworks reuse KV caches for variable-length decoding but have limited parallelism and long-horizon quality.Minute-scale generation makes accumulated prediction errors particularly consequential.Source: §2 · p. 3
• Autoregressive diffusion extends temporal horizons through block-wise denoising and cached context, but remains constrained by information loss, cache-induced errors, and scarce interactive benchmarks.The paper positions BIFE and InterVBench as responses to these gaps.Source: §2 · p. 3
• BIFE’s architecture comparison frames its design around long-horizon interaction and block-wise error accumulation in minute-long generation.Figure 4 compares AR, diffusion, autoregressive diffusion, and BIFE.Source: Figure 4 · p. 4

3.1 Flow Matching

• Flow Matching replaces DDPM noise prediction with velocity prediction along a trajectory from a real starting frame to a Gaussian endpoint.The predicted velocity provides the denoising direction in latent space for reconstructing the current block.Source: §3.1 · p. 4

3.2 Self Forcing

• Self Forcing generates a full video semi-autoregressively and trains video-level distribution matching against real videos, exposing the model to its own errors.A discriminator evaluates entire videos, and the generator is trained to match the generated-video distribution to the real-video distribution.Source: §3.2 · p. 4
• Training on its own predictions reduces exposure bias and improves temporal consistency.Source: §3.2 · p. 4

4.1 Overview

• BIFE addresses KV-cache information loss, propagated block errors, and training exposure bias with long-range memory conditioning and cross-block regularization.The method combines KV-cache memory design with training objectives for cross-block regularization.Source: §4.1 · p. 5

4.2 Semantic Sparse KV Cache

• The sparse cache probes token importance and retains the smallest token set whose cumulative importance reaches a fraction τ of the total score.The resulting cache contains only the most relevant context tokens.Source: §4.2 · pp. 5–6
• For each new block, BIFE retrieves the top-l past KV caches by prompt-embedding similarity and combines them with the two most recent caches.The aggregated cache conditions generation of the target block alongside its current prompt.Source: §4.2 · p. 6

4.3 Block Forcing

• Self Forcing alone lacks cross-block coherence, and very long generation can still drift in subject identity or scene appearance.Block Forcing is introduced to address this limitation by anchoring the current block to relevant historical context.Source: §4.3 · p. 6
• Block Forcing resamples the top-l relevant past blocks to the current block’s temporal length and averages them into a semantic reference.Its loss penalizes deviation of the predicted velocity from both noise and the semantic reference, with weighting γ; the interactive long-tuning loss combines Self Forcing and Block Forcing.Source: §4.3 · p. 6

Dataset

• InterVBench contains 1,000 videos from DanceTrack, GOT-10k, HD-VILA-100M, and ShareGPT4V, with video blocks at least 50 seconds long.Table 1 overviews the datasets used to construct the benchmark.Source: §5 · p. 6, Table 1 · p. 7
• GPT-4o generates fine-grained captions for every 2–3 seconds of each video.Source: §5 · p. 6
• Human annotators check sourcing, block splitting, and caption accuracy, with at least two annotators participating at each stage.Checks include suitability of source videos, transition errors in blocks, and semantic accuracy and coherence of captions.Source: §5 · pp. 6–7

6.1 Implementation Details

• Training adapts Wan2.1-T2V-1.3B to a four-step causal-attention student, then tunes it interactively with a 14B teacher on 60-second switch prompts.Training extends successive 5-second clips to a maximum of 60 seconds, with one prompt switch sampled uniformly from 5 to 55 seconds; inference uses 8-frame chunks and top-2 semantic retrieval due to GPU-memory limits.Source: §6.1 · pp. 7–8

6.2 Main Results

• BIFE-1.3B achieves the best VBench-Long scores for subject consistency (0.9410), motion smoothness (0.9870), dynamic degree (0.7720), and image quality (0.6527).It maintains competitive background consistency (0.9650) and aesthetic quality (0.5839).Source: §6.2 · p. 8
• On InterVBench, BIFE achieves the lowest error on all five VDE metrics and the highest subject consistency (0.9597), background consistency (0.9588), and motion smoothness (0.9956).Table 3 compares methods on five VDE metrics and five complementary VBench metrics; BIFE outperforms baselines on most.Source: §6.2 · pp. 8–9, Table 3 · p. 8
• For 60-second generation, the semantic sparse KV cache retains 56% of the original cache and yields lower latency, higher throughput, and lower GPU memory use than all baselines.Semantic retrieval and KV sparsification account for 25.66% and 31.00% of total computation time, respectively.Source: Table 4 · p. 8, §6.2 · p. 9

6.3 Ablation Study

• Block Forcing improves all VDE metrics and achieves the highest InterVBench consistency scores, while token entropy remains lower during inference and falls faster during training than with Self Forcing.The authors attribute its target to explicitly modeling inter-block dependencies; γ = 0.5 balances consistency and generation diversity.Source: §6.3 · p. 9, Table 5 · p. 9
• Semantic sparse KV with τ = 0.98 gives the best trade-off among tested cache settings; results change only slightly above that threshold, with some metrics improving and others degrading.The comparison includes rolling KV, dynamic sparse KV, and semantic sparse KV.Source: §6.3 · pp. 9–10, Table 6 · p. 10
• With CogVideoX-2B 480p on VBench-Long extended to 60 seconds, BIFE outperforms Self Forcing and LongLive under the same setting.The passage reports lower overall performance for CogVideoX-2B models than Wan 2.1 1.3B, while BIFE remains ahead of both baselines.Source: §6.3 · p. 10, Table 7 · p. 10

7 Human Evaluation for InterVBench

• Twenty-five human experts labeled preferences on the test set of generated videos.The evaluation followed the VBench protocol.Source: §7 · p. 10
• For each pair, the preferred model receives 1, the other 0, and a tie gives both models 0.5; win ratio is total score divided by pairwise comparisons.Source: §7 · p. 10
• Table 8 reports strong alignment between InterVBench scores and human evaluations.Source: §7 · p. 10

8 Conclusion

• BIFE's semantic sparse KV cache retrieves historical context, while Block Forcing reduces cross-block drift and stabilizes long-term generation.InterVBench provides fine-grained annotations and VDE metrics for systematic evaluation of long-horizon consistency and interaction quality.Source: §8 · p. 11
• Experiments on InterVBench and VBench show state-of-the-art performance while maintaining strong efficiency.Source: §8 · p. 11

A Algorithm: Semantic Sparse KV Cache

• The semantic sparse cache combines recent sequential context with prior blocks selected by semantic similarity to the current prompt.The algorithm embeds the current and prior prompts, selects the top-K matches, and concatenates their cached states with sequential context.Source: A Algorithm: Semantic Sparse KV Cache · p. 15

B.1 Prompts for InterVBench’s Data Engine

• InterVBench’s data-engine prompt asks for one coherent paragraph describing the frame’s subject, actions, setting, lighting, and cinematic qualities, using the previous frame description as context.Source: §B.1 · p. 15

C Metrics

• Video Drift Error (VDE), inspired by MAPE and WMAPE, measures changes in video quality through five long-video metrics.Source: C Metrics · p. 15
• VDE Clarity tracks temporal drift in sharpness, VDE Motion tracks smoothness drift, and VDE Aesthetic tracks drift in visual appeal; lower scores indicate more stable quality.Source: C Metrics · p. 16
• VDE Background measures setting stability and VDE Subject tracks identity drift; the evaluation also includes five complementary VBench metrics.Source: C Metrics · p. 16

• MAPE measures the average relative deviation between predicted and ground-truth values as a percentage.Source: §C.1.1 · p. 16
• WMAPE normalizes absolute error by the sum of actual values, making it scale-invariant and more robust in practice.Source: §C.1.1 · p. 16
• Both metrics can quantify deviations across frames or segments in video tasks.Source: §C.1.1 · p. 16

C.1.2 Video Drift Error (VDE)

• VDE compares quality scores across segments with the first segment; high values indicate fluctuation or degradation, while low values indicate stable quality over time.Source: §C.1.2 · p. 17
• The method divides a video into equal-duration segments, scores each segment, computes later segments’ relative changes from the first, and weights absolute changes into the final VDE.Source: §C.1.2 · p. 17

C.1.3 VDE Metrics

• Metric-specific VDEs track drift in subject identity, sharpness, motion, aesthetics, and background stability.Low VDE values indicate stable clarity, consistent dynamics, sustained aesthetics, and a stable background; low VDEsubj indicates a recognizably consistent protagonist.Source: §C.1.3 · pp. 17–18

C.1.4 Complementary Metrics

• Complementary VBench measures cover imaging quality, motion smoothness, aesthetic quality, background consistency, and subject consistency.Source: §C.1.4 · p. 18

D Limitations and Future Work

• Multi-shot composition remains unexplored, particularly coherence across scene transitions.The authors also propose exploring a larger InterVBench and 3D-aware modeling to assess and broaden the method’s applicability.Source: D Limitations and Future Work · p. 19

D.1 Visualization Comparison

• BIFE maintains subject and background consistency, image quality, and color across minute-long videos, unlike baselines that show accumulation errors.MAGI-I, Self-Forcing, and PAVDM degrade after around 12 seconds; SkyReel-V2 develops color drift after 12 seconds, while FramePack has poor dynamics and limited content diversity.Source: §D.1 · p. 19
• In an interactive long-video example, BIFE shows stronger subject consistency, smoother motion transitions, and more stable scene structure than existing approaches.Source: §D.1 · p. 20