What is fine-tuning and how does Astria support it?
Fine-tuning adapts pre-trained generative image models like Flux or Stable Diffusion to create personalized results for specific subjects or styles. Astria's API allows generating studio-quality images, including professional headshots, cartoons, and avatars, using various fine-tuning types like Checkpoint and LoRAs for SDXL and SD1.5.
What are the alternatives to fine-tuning offered by Astria?
Astria provides a FaceID-like adapter approach alongside Dreambooth fine-tuning. This alternative offers lower resemblance but requires significantly less processing time, making it suitable for creating lower fidelity results. It's a faster option for certain use cases where high fidelity isn't the primary goal.
What happens to my model and data after fine-tuning?
The entire fine-tune, including the CKPT file, model, training pictures, generated images, and prompts, is kept for one month. After this period, all data is automatically deleted. Users have the option to delete their data at any time or extend the model storage beyond 30 days through the billing page.
Can I download the CKPT file from Astria?
Yes, once the fine-tune processing is complete, you can download the CKPT file. This file can then be used with other software like Automatic1111 and ComfyUI, providing flexibility for further use of your custom-trained models outside of the Astria platform.
What image size should I upload for training?
While there's no hard limit, a reasonable upload limit is 3MB per image. Images should be at least 512x512 pixels, as models train on this size. For better results and to accommodate face-cropping, 1024x1024 or 2048x2048 are recommended. Ensure low compression to avoid artifacts.