New Forge only.
an adaptation of:
will download ~8GB because it fetches both pytorch model and safetensors model. Don't know why. The 7B model doesn't have this issue.
May need installation of some additional dependencies:
attrdict
I've reworked the UI from the original Huggingface Space. Added automatic saving of generated images, with some infotext included, to webUI outputs
directory. Name will be Janus-Pro_{date and time}{batch number}.png
.
Model is capable of generating variable height (including > 384px), but changing width causes distortion.
Note
Install via Extensions tab; Install from URL sub-tab; use URL of this repo.
Tip
You can edit forge-app.py
line 20 to use_7B = True
to use the larger model instead, quantized to 4 bit (comment out the two quantization_config=
if you want, or change to 8 bit), but t2i is too slow to be usable for me. Multimodal is still OK. Download size is 13.8GB.
I haven't added any form of manual offloading; standard diffusers model offloading or sequential offloading don't work.
prompt | size | image | note |
---|---|---|---|
The image depicts a unique landscape with geothermal pools in the foreground, surrounded by arid terrain and distant mountains under a starry sky. Above the landscape, there is a depiction of the Milky Way galaxy, with red stars and nebulae visible. | 384x512 | ![]() |
low CFG, mid temperature |
a beautiful woman with her face half covered by golden paste, the other half is dark purple. one eye is yellow and the other is green. closeup, professional shot (from this reddit post) | 384x560 | ![]() |
increasing height too far tends to fade/blur the lower part |
digital art shows a person with blue hair styled in a pony tail. They have a black choker around their neck and wear a red velvet jacket. The background is a bustling rustic market. | 384x384 | ![]() |
mid CFG / higher temperature seems better for stylized images |