Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

11 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

World-model demo comparison on sample demo_29

Input: SANA_WM_dev/playground/dataset/dataset/demo_29.png + demo_29.txt prompt. Shared camera trajectory (~60s): forward, pan right, forward, backward — equal quarters. All runs drive the unmodified baselines/demo server through its HTTP API (one GPU per run unless noted). clip-by-clip = one action per segment, chained on the last frame; whole-video = a single action string / long rollout in one generation call.

video model variant temporal protocol clips x frames fps res duration_s gen_time_s notes
infworld_clip.mp4 Infinite-World default causal (chunk-autoregressive) clip-by-clip 20 x 49 16 896x448 61.26 770.0 whole-video mode not supported (fixed 49-frame chunks)
sana_wm_bidirectional_clip.mp4 Sana-WM bidirectional 1600M 720p + refiner bidirectional clip-by-clip 20 x 49 16 1280x704 60.0 1110.0 refiner on; Pi3X-estimated intrinsics; flow_shift 8
sana_wm_bidirectional_whole.mp4 Sana-WM bidirectional 1600M 720p + refiner bidirectional whole-video 1 x 961 16 1280x704 60.0 927.0 961 frames in one shot
hy_ar_distill_clip.mp4 HY-WorldPlay ar_distill causal (autoregressive) clip-by-clip 12 x 125 24 832x480 62.51 627.0 chunk_latent_frames=4
hy_ar_clip.mp4 HY-WorldPlay ar causal (autoregressive) clip-by-clip 12 x 125 24 832x480 62.51 3679.0 chunk_latent_frames=4
hy_ar_rl_clip.mp4 HY-WorldPlay ar_rl causal (autoregressive) clip-by-clip 12 x 125 24 832x480 62.51 3683.0 chunk_latent_frames=4
hy_bi_clip.mp4 HY-WorldPlay bi bidirectional clip-by-clip 12 x 125 24 832x480 62.51 5731.0 chunk_latent_frames=16
hy_ar_distill_whole.mp4 HY-WorldPlay ar_distill causal (autoregressive) whole-video 1 x 1441 24 832x480 60.04 824.0 1441 frames in one rollout
hy_ar_whole.mp4 HY-WorldPlay ar causal (autoregressive) whole-video 1 x 1441 24 832x480 60.04 4510.0 1441 frames in one rollout
hy_ar_rl_whole.mp4 HY-WorldPlay ar_rl causal (autoregressive) whole-video 1 x 1441 24 832x480 60.04 4506.0 1441 frames in one rollout
hy_bi_whole.mp4 HY-WorldPlay bi bidirectional whole-video 1 x 1441 24 832x480 60.04 10419.0 1441 frames in one rollout
dreamx_ar_clip.mp4 DreamX-World-5B ar (long-horizon autoregressive) causal (autoregressive) clip-by-clip 12 x 81 16 1280x704 60.76 488.0 21 latent frames per clip
dreamx_ar_whole.mp4 DreamX-World-5B ar (long-horizon autoregressive) causal (autoregressive) whole-video 1 x 969 16 1280x704 60.56 307.0 243 latent frames in one rollout
dreamx_cam_clip.mp4 DreamX-World-5B-Cam cam bidirectional clip-by-clip 12 x 81 16 1280x704 60.76 2095.0 shift 3.0
dreamx_cam_whole.mp4 DreamX-World-5B-Cam cam bidirectional whole-video 1 x 961 16 1280x704 60.06 9880.0 961 frames in one bidirectional shot (~12x training clip length); required memory workarounds on one H200
lingbot_small_clip.mp4 LingBot-World small/fast (distilled streaming) causal (chunked streaming) clip-by-clip 20 x 57 16 832x464 71.26 570.0 i2v-A14B; 49-frame request snapped to 57 (3-latent chunking)
lingbot_large_clip.mp4 LingBot-World large/base (act2cam) bidirectional per segment clip-by-clip 20 x 49 16 832x464 61.26 2282.0 i2v-A14B
lingbot_small_whole.mp4 LingBot-World small/fast (distilled streaming) causal (chunked streaming) whole-video 1 x 961 16 832x464 60.56 418.0 i2v-A14B; KV cache bounded to a 24-latent-frame rolling window (LINGBOT_DEMO_LOCAL_ATTN_SIZE=24) — full-sequence cache (~300GB) cannot fit one GPU
lingbot_large_whole.mp4 LingBot-World large/base (act2cam) bidirectional per segment whole-video 1 x 961 16 832x464 60.06 3376.0 i2v-A14B; run distributed on 4 GPUs (FSDP+SP) — 961-frame attention OOMs on 1 and 2 H200s
lingbot_v2_fast_clip.mp4 LingBot-World-v2 14B causal-fast (distilled streaming) causal (chunked streaming, KV window) clip-by-clip 20 x 57 16 832x464 71.26 660.0 LingBot-World-Infinity (arXiv:2607.07534); i2v-A14B single transformer; 4-step distilled, no CFG; local_attn_size=18, sink_size=6 (official run_fast.sh); 49-frame request snapped to 57 (3-latent chunking)
lingbot_v2_fast_whole.mp4 LingBot-World-v2 14B causal-fast (distilled streaming) causal (chunked streaming, KV window) whole-video 1 x 969 16 832x464 60.56 346.0 single 969-frame stream (unbounded-horizon design); local_attn_size=18, sink_size=6
mg3_efficiency_clip.mp4 Matrix-Game-3.0 efficiency (distilled + int8) causal (streaming interactive) clip-by-clip 26 x 57 first, then 40 17 1280x704 62.18 446.0 run distributed on 2 GPUs (single-GPU interactive path has an upstream device bug); whole-video mode not supported (streaming only)
mg3_quality_clip.mp4 Matrix-Game-3.0 quality (base model) causal (streaming interactive) clip-by-clip 26 x 57 first, then 40 17 1280x704 62.18 5908.0 run distributed on 2 GPUs (single-GPU quality path has an upstream device bug); whole-video mode not supported

Latency detail: latency.csv (model load, first-segment, avg per-segment, total).

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages