whatmodel.sh / image / CogView2

CogView2

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

CogView2 is an open-weights image model from THUDM, released April 2022. Licence: apache-2.0.

Catalogue identity notes

THUDM/CogView2 (hierarchical 6B-9B-9B transformer per its README). Comparison-only: runs only through the SwissArmyTransformer research code; no single copy-pasteable command or memory figure.

Model information sources

· Parameter count ↗

Maker-stated size of the pretrained CogLM transformer: 'We pretrain a 6B-parameter transformer' (CogView2 paper). The image tokenizer and the super-resolution modules are not separately counted. Observed 2026-10-07.

1.4±4.0#144 of 144 image modelsfrom 4 benchmark results across 3 sources

Score within the LMArena text-to-image comparison set only.

$ how it compares

The image models ranked either side of CogView2, out of 197 with a score.

#modelscorewhereRAMreleased
193LCM Dreamshaper v7 SimianLuo (LCM authors)23±4local~1.7GB2023-10
194InstructPix2Pix UC Berkeley (Brooks, Holynski, Efros)19±3local~2.6GB2023-01
195X-LXMERT UNC Chapel Hill17±4comparison only2020-09
196AttnGAN (COCO) Microsoft / Rutgers / Duke16±4comparison only2017-11
197CogView2 THUDM1±4local~14GB2022-04

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

$ what it's good at

text to image▂▅▅▃▄▄█▃▁▂▁▁ #113 of 126 * · 0.4% of its score · 1 test

Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ), VILA-U 7B (256px) here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
HPS v2 benchmark (4-style average)26.5% ✎▃··▃▅·▅·▅▅█▅ beats 6/14hard · avg 27%▇ 0.4%blunt
other tests3 not in a task · 93% of its score
testresultvs the others in its rundifficultycountswhy
ImageReward human ranking of T2I models (#wins, real user prompts) lab74.0000 wins ✎▄·········█▄ beats 0/3avg 526.557 wins▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 63%sharp
HEIM text-image alignment (aspect win rate, v1.1.0)7.4% ✎▃····█▁·▁··▁ beats 0/9very easy · avg 92%▇▇▇▇▇▇▇ 16%sharp · noisy
MS-COCO 2014 zero-shot FID-30k at 256px ↓ (ImageReward protocol) lab26.2000 FID ✎▄··········█ beats 0/2avg 7.871 FID▇▇▇▇▇ 13%blunt · small run

$ where to run it

Other ways to run it

version

  1. Run it as published · needs ~43.7 GB, NVIDIA only · from github.com/zai-org
    git clone https://github.com/zai-org/CogView2 && cd CogView2 && pip install -r requirements.txt && git clone https://github.com/Sleepychord/Image-Local-Attention && (cd Image-Local-Attention && python setup.py install) && SAT_HOME=./sat_models python cogview2_text2image.py --mode inference --fp16 --input-source input.txt --output-path samples_sat_v0.2 --batch-size 16 --max-inference-batch-size 8

Weights and documentation: github.com/zai-org ↗

$ how the score was worked out · ImageReward: Learning and Eva… 76%, HEIM: Holistic Evaluation of… 16%, tgxs002/HPSv2 repository READ… 0%

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
ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation (arXiv 2304.05977v4) · lab-reported2▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 76%CC-BY-4.0
HEIM: Holistic Evaluation of Text-to-Image Models (arXiv 2311.04287v1)1▇▇▇▇▇ 16%CC-BY-4.0
tgxs002/HPSv2 repository README, HPS v2 benchmark table1▇ 0%Apache-2.0
prior · pull toward the middle▇▇ 7%
weighted average= 1.4

$ 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