whatmodel.sh / video / CogVideo 9B (original)

CogVideo 9B (original)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

CogVideo 9B (original) is an open-weights video model from Tsinghua / Zhipu, released May 2022. Its published size is 9.4B parameters. Licence: cogvideo-license.

Catalogue identity notes

Original 2022 CogVideo, distinct from CogVideoX 2024.

Model information sources

· Weights licence ↗ · Parameter count ↗

9.4B = distinct parameters of the two CogVideo stage transformers (7.7B each, sharing 6B frozen CogView2 weights), per the makers' paper; the abstract rounds to 9B. The separately downloaded CogView2 super-resolution model and the icetk tokenizer are not included.

23.2±6.8#93 of 93 video modelsfrom 1 benchmark result across 1 source* thinly tested, treat the rank as provisional

$ how it compares

The video models ranked either side of CogVideo 9B (original), out of 139 with a score.

#modelscorewhereRAMreleased
135LVDM (text-to-video) HKUST / Tencent AI Lab25±3local~6.1GB2023-05
136ModelScope T2V 1.7B Alibaba DAMO Academy25±2local~4.6GB2023-03
137Floor33 Pictures (August 2023) Floor33 Pictures24±2comparison only2023-08
138ZeroScope v2 XL (576w → XL pipeline) cerspense24±2local~15GB2023-06
139CogVideo 9B (original) Tsinghua / Zhipu23±7local~52GB2022-05

see all video models · find the ones that fit your machine

$ what it's good at

text to video▆▄▃▄▆▅█▃▂▂▂▄ #68 of 68 * · 81% of its score · 1 test
testresultvs the others in its rundifficultycountswhy
FETV manual video quality (mean of static and temporal quality)3.27 / 5 ✎█·····█····█ beats 0/2very easy · avg 4.63 / 5▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 81%at its level · small run

$ where to run it

Other ways to run it

  1. Run it as published · needs ~51.7 GB, NVIDIA only · from github.com/zai-org
    git clone -b CogVideo https://github.com/zai-org/CogVideo && cd CogVideo && pip install -r requirements.txt && pip install git+https://github.com/Sleepychord/Image-Local-Attention && MASTER_PORT=$(shuf -n 1 -i 10000-65535) SAT_HOME=./sat_models python cogvideo_pipeline.py --input-source interactive --output-path ./output --parallel-size 1 --both-stages --use-guidance-stage1 --guidance-alpha 3.0 --generate-frame-num 5 --tokenizer-type fake --mode inference --distributed-backend nccl --fp16 --model-parallel-size 1 --temperature 1.05 --coglm-temperature2 0.89 --top_k 12 --sandwich-ln --seed 1234 --num-workers 0 --batch-size 4 --max-inference-batch-size 8

Official weights and documentation: huggingface.co/THUDM ↗

$ how the score was worked out · FETV: A Benchmark for Fine-Gr… 81%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
FETV: A Benchmark for Fine-Grained Evaluation of Open-Domain Text-to-Video Generation (arXiv:2311.01813v3)1▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 81%CC BY 4.0
prior · pull toward the middle▇▇▇▇▇▇ 19%
weighted average= 23.2

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json