Research map / Video / Video generation
PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference
Xiaofeng Mao, Shaohao Rui, Bo Zheng, Kaining Ying, Chuanhao Li, Mingmin Chi, Kaipeng Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sink tokens (N_sink 0 -> 8) (ablation) | PackForcing · 60 sampled prompts | overall CLIP score: 31.24 -> 35.09 | Improved |
| add sink tokens (N_sink 0 -> 8) (ablation) | PackForcing · 60 sampled prompts | Subject Consistency: 74.72 -> 93.11 | Improved |
| increase sink size from 8 to 16 (ablation) | PackForcing · 60 sampled prompts | Dynamic Degree / CLIP: Dynamic Degree 49.84 -> 35.16; CLIP 35.09 -> 35.39 | Mixed |
| use both HR and LR compression branches vs HR only (ablation) | PackForcing · VBench-Long 60s | CLIP: 32.97 -> 33.54 | Improved |
| use both HR and LR compression branches vs LR only (ablation) | PackForcing · VBench-Long 60s | Overall Consistency: 25.18 -> 26.07 | Improved |
| dynamic context selection instead of FIFO eviction (ablation) | PackForcing · VBench-Long 60s | Subject Consistency: 87.82 -> 88.62 | Improved |
| dynamic context selection instead of random selection (ablation) | PackForcing · VBench-Long 60s | Subject Consistency: 86.31 -> 88.62 | Improved |
| add incremental temporal RoPE adjustment (ablation) | PackForcing · VBench-Long 60s | CLIP gap early vs late: 2.53 -> 0.95 | Improved |
| three-partition compressed cache vs Self-Forcing for long generation | PackForcing vs Self-Forcing · VBench-Long 120s | Subject Consistency: 74.40 -> 92.84 | Improved |
| three-partition compressed cache vs LongLive | PackForcing vs LongLive · VBench-Long 60s | Subject Consistency: 92.00 -> 90.49 | Hurt |
| three-partition compressed cache vs LongLive | PackForcing vs LongLive · VBench-Long 60s | Dynamic Degree: 44.53 -> 56.25 | Improved |
| three-partition compressed cache vs Deep Forcing | PackForcing vs Deep Forcing · VBench-Long 120s | Background Consistency: 92.55 -> 91.88 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- Deep Forcing — Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression (ICML 2026) · better on 5 of 6 shared results
- Rolling Forcing — Rolling Forcing: Autoregressive Long Video Diffusion in Real Time (ICLR 2026) · better on 5 of 6 shared results
- LongLive — LongLive: Real-time Interactive Long Video Generation (ICLR 2026) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.89)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.89)
- Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.88)
- VideoMAR: Autoregressive Video Generation with Continuous Tokens (NeurIPS 2025 · similarity 0.88)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
