Instructions to use lodestones/Chroma1-HD with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use lodestones/Chroma1-HD with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("lodestones/Chroma1-HD", torch_dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
Trainning from scratch
Is it possible to train Chroma from scratch? If so, where can i find the scripts that download the images, and start the training?
I don't have the answer to your question, but I'm really curious as to where you're going with this? This project took months and likely cost a lot of money, why would you want to re-do the training that has already been done?
Mostly to learn and try things. I can access some computer time at work for personal projects when its not used :)
Would be a fun thing if you document everything for reproducability. Just so you know, it takes a lot of resources. Like I just thought of making an upscaler model for Chroma, just a thought that maybe training it on good and bad examples from its outputs would make it act as an upscaler plus refiner of sorts. A small 16M parameter model was taking about 16GB vram, that's with some parts kept in fp8 and using gradient accumulation. Granted that I was shooting in the dark, but I learned training is very intensive as compared to inference. I wouldn't be surprised if training chroma on a batch size of 1 takes more than 80Gb vram.