AI Consultant fast-stable-diffusion DM for LoRA, notebooks, workflow or optimization
Joined July 2009
- Tweets543
- Following333
- Followers1.4K
- Likes72K
Manipulating the dataset to save flux lora training time, and potentially increase the quality, this is like batch size but more efficient and it's pretty simple, by stacking 4 512px images inside each 1024px canvas, it is definitely a promising method to optimize resources.
For powerful models like Flux, training LoRAs shouldn't be straightforward, taking advantage of the power of the model gives a lot of room for optimization and fine-grained LoRA training, that includes targeting specific blocks at varied strengths, it saves time, memory & quality
Existing tokens greatly help reduce the training time, the joker and the hound were the training tokens for these flux LoRAs, between 500-700 steps, and about 10-20 minutes training time on A100-80G
Davy Jones, No LoRA vs 350 steps LoRA, Flux.
Extremely low quality 7 pics dataset resized (not upscaled) up to 1024 for more stable training.
AI in 2024 is Bitcoin in 2013, whatever you're doing keep doing it and you'll have a chance to stand with the giants in the future, ups and downs are just noise, bitcoin crashing from 1$ to 75 cents was just noise, same with AI, don't listen to the noise.
SORA is the most dangerous AI yet, ironically, it's coming from a company that pretends caring about safety from the dangers of AI or the so called AGI.
Once SORA is released to the public, the internet will be flooded with fake videos, but that's not the issue ...🧵
... actually, only one feature, video to video at a safe denoise level, 50% would be good, it will protect the integrity of the physics involved and allow style transfer.
Full access should be only allowed in a limited way to artists, and companies with specific use cases ...