Pinned Tweet
It is October 20, 2024, and this is stable diffusion turbo image-to-image running live from my webcam in Chrome CLIENT-SIDE at 3 FPS on a 3070 Ti laptop using ONNX Runtime Web + WebNN + DirectML. Check out my im2im fork of Microsoft's WebNN preview: github.com/eyaler/webnn-deve…
#WebAI
The new A100 High-RAM on @GoogleColab allows running @Alibaba_Qwen Image Edit Plus, and with the 8-step ligthning lora it takes ~9 sec/frame, so we can do the @genekogan-@radamar nano-banana-loop style e.g. "Show the next scene one second later" locally: github.com/eyaler/Qwen-Image…
1
1
324
I did find Qwen takes the word "future" too literally (and cyan-purply)... That's why I went with the alternative "next/later" wording as I just wanted the very near future. Also, it seems to really like overlaying diagrams. Nano banana did much better, but no skipped frames here
1
84
i have been using @Bitly until recently for almost as decade as my goto method to distribute links, and i am not happy.
on apr 5 2022 all j[.]mp shortlinks created during jan 17 to apr 5 and tested and working up to that day, suddenly stopped working. there was no public [🧵]
1
122
without paying. while i could parse my bitly links page myself and extract the data, with no usage analytics i would have to set up hundreds of new shortlinks. now this all feels to me like rivermind plus lux vip level shit. so i created nay[.]ad - "mobile" in hebrew and also...
1
97
If you missed my recent IRL tutorial on combinatorial optimization with Google OR-Tools, I am giving it again online this Wednesday 17:30 (Israel time) at @PyData Global 2024
global2024.pydata.org/cfp/ta…
1
163
I like #SANA's safety filter that draws a heart in response to NSFW prompts. But even better is it's "not-safety" filter that draws NSFW images in response to "not-NSFW" prompts!
Giving a short tutorial next Monday on combinatorial optimization with OR-Tools @PyDataTLV (open loudspeaker!)
1
108
Girlfriend said I can only use this one photo. OK but did you hear about sd-turbo image-to-image? Well it can run in your browser! Almost. You do need to download 2.5GB of models + GPU + some ops currently only implemented for Windows/DirectML + Chormium (+WebNN is behind a flag)
1
161
Oh, and I may have forgot to mention that I removed the model safety checks ;)
LinkedIn write-up (part 1): linkedin.com/posts/eyalgruss…
The full 11-part video: youtu.be/iZyR1EFC2WU
(11/11)
72








