language:
- en
thumbnail: https://huggingface.co/jkcarney/source4_v1.0/blob/main/images/woman1.png
license: creativeml-openrail-m
tags:
- stable-diffusion
- stable-diffusion-diffusers
- text-to-image
- safetensors
- diffusers
inference: true
Source4 is a Stable Diffusion 1.5/Dreambooth model trained on portraits with dramatic, multicolored lighting. These portraits are heavily inspired by modern, multicolored theatrical lighting. This model was fine tuned without prior preservation loss for 8000 steps on over 90 high quality 512x512 images.
In the prompt, use activation token source4 style
I recommend experimenting with different colors in your prompt. For example, adding teal tones or red tones to your prompt bring those respective colors out. Also try a smaller weight on the source4 style prompt (ie, (source4 style:0.8)), for a slightly less pronounced but more realistic effect.
Euler a and DPM++ 2S a Karras for 25-30 steps have typically given me the highest quality images. Euler a typically gives me more "dreamy" portraits while DPM++ 2S a Karras gives me more realistic portraits.
If the effect is too pronounced, I recommend decreasing the weight of the activation token or adding words such as harsh to the negative prompt. An alternative is to specify detailed face to ensure it gets added.
Here are some examples with High Res Fix:
Enjoy!