whatmodel.sh / image / VQ-Diffusion (ITHQ)
VQ-Diffusion (ITHQ)
↗ share
+ contribute
VQ-Diffusion (ITHQ) is an open-weights image model from Microsoft, released November 2022. Its published size is 1.27B parameters. Licence: mit.
Catalogue identity notes
microsoft/vq-diffusion-ithq; released month = Hugging Face release. Comparison-only: VQDiffusionPipeline is deprecated upstream and no memory figure is published.
Model information sources
1.27B = the Improved VQ-Diffusion-L backbone that the makers trained on ITHQ-200M (the released ITHQ checkpoint); not stated whether the VQ-VAE and the CLIP text encoder are included.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of VQ-Diffusion (ITHQ), out of 197 with a score.
| # | model | score | where |
|---|---|---|---|
| 108 |  Step1X-Edit v1.1 | 48±3 | local |
| 109 | FuseDream (BigGAN 512) | 48±15 | comparison only |
| 110 | VQ-Diffusion (ITHQ) | 48±15 | local |
| 111 | LAFITE (HPS v2 checkpoint) | 48±15 | comparison only |
| 112 |  SeedEdit 3.0 | 48±3 | comparison only |
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #118 of 126 * · 6.1% of its score · 1 test
Trails CogView2, Versatile Diffusion v1.0, Dreamlike Photoreal 2.0 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| HPS v2 benchmark (4-style average) | 25.1% ✎ | ▃··▃▅·▅·▅▅█▅ beats 2/14 | 6.1% |
$ where to run it
Other ways to run it
- Run it as published · needs ~4.1 GB, NVIDIA only · from huggingface.co/microsoft (official)
uv run --with torch --with diffusers --with transformers python -c "import torch; from diffusers import VQDiffusionPipeline; p=VQDiffusionPipeline.from_pretrained('microsoft/vq-diffusion-ithq', torch_dtype=torch.float16).to('cuda'); p('teddy bear playing in the pool', truncation_rate=1.0).images[0].save('out.png')"
Official weights and documentation: huggingface.co/microsoft ↗
$ how the score was worked out · tgxs002/HPSv2 repository READ… 6%
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
| source | results | share of the score |
|---|---|---|
 tgxs002/HPSv2 repository README, HPS v2 benchmark table | 1 | 6% |
| prior · pull toward the middle | 94% | |
| weighted average | = 47.8 |
$ improve this page
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