Dreambooth fforiver-river
Quick Training Report for the Hugging Face DreamBooth fine-tuning hackathon.
Concept : fforiver : forest river
Created on January 23|Last edited on January 26
Comment
Dataset :
- 12 images of forest rivers
Selected pretrained models:
- runwayml/stable-diffusion-v1-5
- prompthero/openjourney
- nitrosocke/elden-ring-diffusion
Config tested :
- learning rate : 1e-6 & 2e-6
- gradient accumulation step : 1 & 2
- nbr_steps : 700 - 500 - 400 - 384 - 312
Run set
3
- Seems to overfit at step 600
- Elding ring and openjourney seems to be more promising for landscapes
Run set
3
- Gradient accumulation step of 2 seems to produce noisier images -> overfits too quickly
- step 312 and 324 seems to give the best results, it seems to overfit quickly afterwards
At step 324 our model seems to overfit a little when producing other objects.
Chosen model :
- prompthero/openjourney : fine-tuned - step 312
- nitrosocke/elden-ring-diffusion : fine-tuned - step 312
Add a comment