(This feature is still experimental, but we're working on improving it. Multiple concepts: You can combine multiple trained concepts in a single image.This makes them easier to share, store, and re-use. Smaller outputs: Trained LoRA outputs are much smaller than DreamBooth outputs.Faster training: Training a new concept with LoRA takes just a few minutes.LoRA has a few differences from DreamBooth that make it especially appealing as an alternative: Unlike DreamBooth where you had to wait for a model to push and boot up, LoRA predictions run instantly with no cold boots. You can now train LoRA models in the cloud with a single API call. We've been collaborating with Simo to get LoRA up on Replicate. Check out the README for Simo's inference model on GitHub and the paper on arXiv to learn more about how it works. LoRA was developed by researchers at Microsoft, and Simo has applied it to Stable Diffusion. You can think of it like creating a diff of the model, instead of saving the whole thing.
LoRA stands for Low-Rank Adaptation, a mathematical technique to reduce the number of parameters that are trained. Unlike DreamBooth, LoRA is fast: While DreamBooth takes around twenty minutes to run and produces models that are several gigabytes, LoRA trains in as little as eight minutes and produces models that are around 5MB. Similar to DreamBooth, LoRA lets you train Stable Diffusion using just a few images, and it generates new output images with those objects or styles. It was a way to train Stable Diffusion on your own objects or styles.Ī few short months later, Simo Ryu has created a new image generation model that applies a technique called LoRA to Stable Diffusion.
0 kommentar(er)
