Pretty much my entire feed on OpenAI Sora 2 is now just videos saying "Double tap on Sora videos to see an emoji of the vibe of the video. Leave a comment if you didn't know". And there are thousands of these engagement farming videos polluting the feed.

Plans for the weekend? ☀️🕶️ how about fine tuning a Qwen Image, Wan or FLUX LoRA for free? Our LoRA Frenzi event is still taking place until Monday!

I implemented a modified version of lodestone-rock 's RamTorch into AI-Toolkit. Currently training a Qwen-Image LoRA on my laptop with a 3060 6GB GPU!!! It is only using ~4GB of VRAM. Crazy!

for a bit of the details how this works it's just really clever offloading strats where the comms are masked and interleaved perfectly with computation. creating virtually zero latency transfer

Thank you for participating in the LoRA Frenzi! 1560 LoRAs were trained during this week 🤯 Missed it? If you have a cool LoRA idea, DM me :)

My collection of Qwen-Image LoRAs 🧵 done thanks to apolinario 🌐, Hugging Face, Ostris and Qwen. It consists in these LoRAs (examples and links on following tweets) huggingface.co/collections/li…

Added a queuing system to the AI Toolkit UI. There is a seperate queue for each GPU. You can add numerous training jobs and they will train one after the other. Queues can be paused and resumed. This is a heavy rework, so let me know if you run into any issues.

I'm currently making a tutorial for training qwen-image-edit-2509 LoRA on <24GB of VRAM. The fact that it can put this QR code on her shirt accurately enough to scan in 1,000 steps is mind boggling. I love this model.

I created a little shop where you can purchase merch to support the development of AI Toolkit. Currently, there is only a shirt with the Ostris logo, but I will add more things soon. Just testing the waters, so open to ideas and feedback. shop.ostris.com

If you run the exact same prompt through conditional and unconditional (positive/negative), it produces the same result as not doing CFG at all, leading to a bad image. Overfitting makes a model generate the newly trained subject both prompted or unprompted, which means it