DraftKV_Assets / README.md
hp-l33's picture
Add LingBot1.3B Table2 (ORBIA 480, H100) samples: 5 cases x 6 methods
7567669 verified
|
Raw History Blame Contribute Delete
3.42 kB
---
license: other
tags: [video-generation, kv-cache, draftkv]
---
# DraftKV experiment assets
Sample outputs archived from the DraftKV experiments. Full metrics, per-case scores and
code live in the GitHub repository (TillyEndless/DraftKV):
- results: branch `results/bench1-h100-longlive1.3b`
- code and cluster scripts: branch `exp/dfw-h100-bench1`
## bench1/longlive1.3b_h100
Table1 (Bench1 T2V98, official VBench-Long), LongLive 1.3B on NVIDIA H100 (SP1), 957 frames,
16 fps, 832x480, seed 31415. Four cases (PVB_004, PVB_148, PVB_352, PVB_582), each with
Native, DraftKV, ShadowKV, InfiniGen and InfLLM. Each MP4 is the exact file scored by
VBench-Long. `index.json` lists prompts and the six Table1 scores per video.
Generation commits: Native 70a78df (PVB_004 at b08ed41), DraftKV 3309c86 (current Chrono-Compact +
Spatial 8x16 implementation; replaced the earlier absolute-time eb4e4a3 videos),
ShadowKV/InfiniGen/InfLLM 70a78df.
## bench1/lingbot14b_h100
Table1 (Bench1 TI98, official VBench-Long), LingBot 14B on NVIDIA H100, **SP8 for every method**
(including Native), native 832x464 decode, 957 frames, 16 fps. Case PVB_004 with Native, DraftKV
(Chrono-Compact + Spatial 8x16 + P3/z3), ShadowKV, InfiniGen and InfLLM. `index.json` lists the
six Table1 scores per video and the generation run id.
## bench2/lingbot1.3b_h100_subset48
Table2 (Bench2 ORBIA, preliminary 48-case subset), LingBot 1.3B on NVIDIA H100 (SP1), 357 frames
at 16 fps. Case case_0002 with Native, DraftKV, ShadowKV, InfiniGen, InfLLM and WorldKV; each MP4
has the official ORBIA per-case `report.json` it was scored from. ORBIA ran with a Comfy-Org FP16
substitute for the gated facebook/sam3.1 checkpoint (see the GitHub results branch for details).
## bench2/lingbot14b_h100_subset48
Table2 (Bench2 ORBIA, preliminary 48-case subset), LingBot 14B on NVIDIA H100, **SP8 for every
method**, 357 frames at 16 fps. Five cases, one per camera template (case_0001, case_0002,
case_0003, case_0005, case_0013), each with Native, DraftKV (Chrono-Compact + Spatial 8x16 +
P3/z3), ShadowKV, InfiniGen (token-level, author-faithful), InfLLM and WorldKV. Every MP4 has the
official ORBIA per-case `report.json` it was scored from; `index.json` also lists the WorldAtlas
memory metrics per video (Return Gate = ViPE-trajectory gate; the fixed-GT gate is
`gt_return_gate`). These videos predate main's unified LingBot14B BF16-streaming/FULL_SHARD
loader (main e3f2213); the full ORBIA480 rerun uses the new loader. ORBIA used the Comfy-Org FP16
substitute for the gated facebook/sam3.1 checkpoint.
## bench2/lingbot1.3b_h100
Table2 (Bench2 ORBIA, **all 480 cases**), LingBot 1.3B on NVIDIA H100 (SP1), 357 frames at 16 fps.
Five cases, one per camera template (case_0001 peek_occlude_retract, case_0003 lateral_out_and_back,
case_0005 yaw_return, case_0007 forward_backward, case_0013 arc_return), each with Native, DraftKV
(Chrono-Compact + Spatial 8x16 + P3/z3), ShadowKV, InfiniGen (token-level), InfLLM and WorldKV. Every
MP4 has the official ORBIA per-case `report.json` it was scored from; `index.json` also lists the
WorldAtlas memory metrics per video (Return Gate = ViPE-trajectory gate; fixed-GT gate is
`gt_return_gate`). Generated at branch commits 16823a5 / e4478fe (see the GitHub results branch,
`results/bench2-h100-lingbot1.3b`). Complements the earlier single case in
`bench2/lingbot1.3b_h100_subset48/case_0002`.