whatmodel.sh / image / DALL·E Mega (minDALL-E)

DALL·E Mega (minDALL-E)

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

DALL·E Mega (minDALL-E) is an open-weights image model from DALL·E Mini / minDALL-E, released July 2022.

Catalogue identity notes

TIFA Appendix E explicitly identifies minDALL-E as DALL·E Mini with Mega weights. HPS Appendix D names the same min-dalle implementation, not original OpenAI DALL·E. Local command uses CPU; 8GB memory is an estimate including runtime headroom, not measured speed or capacity.

28.7±3.8#134 of 144 image modelsfrom 3 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 DALL·E Mega (minDALL-E), out of 197 with a score.

#modelscorewhereRAMreleased
181Stable Diffusion v2.0 Stability AI29±3comparison only2022-11
182Versatile Diffusion v1.0 SHI Labs29±4local~7.2GB2022-11
183DALL·E Mega (minDALL-E) DALL·E Mini / minDALL-E29±4local~8GB2022-07
184MagicBrush (InstructPix2Pix fine-tune, jul7) Ohio State University (OSU NLP)29±7comparison only2023-07
185Openjourney PromptHero28±2local~2.1GB2022-11

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

$ what it's good at

text to image▂▅▅▃▄▄█▃▁▂▁▁ #119 of 126 * · 64% of its score · 2 tests

Trails CogView2 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
TIFA v1.0 (overall, mPLUG VQA)79.4% ✎█·······▄▄▄▄ beats 4/5very easy · avg 95%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 64%typical
HPS v2 benchmark (4-style average)25.8% ✎▃··▃▅·▅·▅▅█▅ beats 5/14hard · avg 27%▇ 0.4%blunt · at its level
other tests1 not in a task · 30% of its score
testresultvs the others in its rundifficultycountswhy
HEIM text-image alignment (aspect win rate, v1.1.0)58.9% ✎▁····█▁·▁··▁ beats 5/9very easy · avg 92%▇▇▇▇▇▇▇▇▇▇▇▇ 30%sharp · at its level · noisy

$ where to run it

Other ways to run it

  1. Run it as published · needs ~8 GB · from github.com/kuprel
    uv run --with torch --with min-dalle python -c "import torch; from min_dalle import MinDalle; m=MinDalle(models_root='./pretrained', dtype=torch.float32, device='cpu', is_mega=True, is_reusable=True); m.generate_image(text='A lighthouse at sunset', seed=1, grid_size=1, is_seamless=False, temperature=1, top_k=256, supercondition_factor=32, is_verbose=False).save('image.png')"

Weights and documentation: github.com/kuprel ↗

$ how the score was worked out · TIFA: Accurate and Interpreta… 64%, HEIM: Holistic Evaluation of… 30%, 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
TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering (arXiv v3)1▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 64%CC-BY-4.0
HEIM: Holistic Evaluation of Text-to-Image Models (arXiv 2311.04287v1)1▇▇▇▇▇▇▇▇▇ 30%CC-BY-4.0
tgxs002/HPSv2 repository README, HPS v2 benchmark table1▇ 0%Apache-2.0
prior · pull toward the middle▇▇ 6%
weighted average= 28.7

$ 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