StreamDiffusion v0.1.3 TOX 🖼️🚀
Txt2img, offline support, working models menu and more!
( 0.1.4 ) Fixes Config error which might happen when first loading 0.1.3 due to not removing my custom VAE path.
Txt2img now available ! 'Sdmode' menu added in Settings 2 page
- This can be updated live when Streaming.
- If Sdmode is set to txt2img, then the stream will not be effected by the input image
- All settings still have an effect on the output, including the step sliders.
Added 'My Models' menu which display a list of models that is working for your setup (inspired by suggestion from Function Store - check out his TouchDesigner tools)
- If you select a Model from this list, it will populate the Modelid parameter with the working path / huggingface id
- Syncs with a json file stored in your virtual environment. Must edit json to remove model if no longer working.
Prompt block deletion will now update prompt settings instantly
Added 'Debug (CMD stays open)' toggle parameter, which will cause the Stream cmd window to stay open if it disappears instantly. (for debugging)
Cuda 12.1 Installation Fixes:
- Added more info in popup if cuda version is not found / supported
- Fixed typo that caused issue when installing with cuda12.1 (only cuda version supported are 11.8 and 12.1)
- Moved nvidia-pyindex to start of installation to ensure NVIDIA package list is used
Offline model support (needs diffuser format) - The way that the repo loads the models currently will check huggingface api for Model information if it is not in the correct format that it expects. And it also has used the huggingface id for VAE + LCM model so it also checks for those. This may have caused issues when starting stream, such as long pauses or error on download.
OFFLINE ADDITIONS:
- 'Get local LCM / VAE models' pulse - added Install pulse to download these two models and set correct parameters in Lora parameter page
- Custom LCM and Custom VAE and associated toggles in the Lora parameter page
- Improved handling of local model paths (.safetensors / 1.5 models may load more consistently)
HOW TO OFFLINE: To get the Model ID/Path to work locally, you need to manually download a huggingface repo/model in the diffusers format and then link the repo folder. example 'F:/StreamDiffusion/models/Model/sd-turbo' is what I have the parameter set as. and I ran this cmd 'git clone https://huggingface.co/stabilityai/sd-turbo' in the models\Model folder inside StreamDiffusion base folder.
Here are some working models from huggingface that are in diffuser format: (thanks @ forkni for the model repo list)
stabilityai/sd-turbo
darkstorm2150/Protogen_x5.8_Official_Release
runwayml/stable-diffusion-v1-5
prompthero/openjourney
Lykon/dreamshaper-7
( You do not need to reinstall the venv after 0.1.2 unless you had issues with cuda 12.1! )
Comments (3)
same here with 11.8
Hi mate, thanks again for the great stuff! I have an issue and I don't know why it's happening. (I had it since the beginning of streamdiffusion TOX) https://drive.google.com/file/d/1-SmdH13TBjXtjAllS6-kHMbm34qFexoo/view?usp=sharing sometimes the stream works just fine (6-7 fps on a 3060 laptop with 6gb vram, which is incredible btw) however, sometimes when the stream starts -no errors, everything looks as it should- it just renders one frame, and then disconnects, and then connects again, then disconnects again, like in the video. Anybody else having the same trouble? any ideas what may be causing it?
Im guessing its hitting ram or gpu vram limit depending on what else is open. I have a timer that is being used to check if the Stream is active. When the fps gets low, it will end up doing what you are seeing. If you go inside and change the timer_osc 's length parameter to a higher value, you should be good.