language:
- en
tags:
- stable-diffusion
- stable-diffusion-diffusers
- text-to-image
license: creativeml-openrail-m
Starsector Portraits
This is a fine-tuned Stable Diffusion model trained on images from the videogame Starsector
Use the token starsectorportrait in your prompts for the effect.
Example prompts & tips
"a starsectorportrait of a person wearing a green balaclava and armored spacesuit"
"a starsectorportrait of an older man wearing a fancy suit with a purple cape"
Even with the training I found it was still useful to include some other phrases and tokens that can be helpful to more accurately match the base art style:
sabattier effect to get a more accurate lighting effect
brush strokes to achieve the more painterly feel of the portraits
conte if brush strokes is not doing enough or adding colorful artifacts to images this is another option
Sample images from the model:
File Versions and Comparison Details
All model versions are trained using a hand labeled image set of 556 images with black/white backgrounds and flipped versions of the same images
Any model versions with pploss in the name were trained with prior preservation loss enabled and used a regularization set of 8000 portrait images comprised of photos/paintings/digitalart of humans in modern day attire.
CFG and SAMPLE STEP comparisons between rc_v1 model files
(prompt: "a starsectorportrait of a person" sampler: euler_a)
Adaptability comparisons between rc_v1 model files
(prompt: "a starsectorportrait of an alien cat" sampler: euler_a)
Hypernetworks
Fine tuning hypernetworks to help with various aspects of image generation.
HN_ssportrait_rc_v1_no-helmet_finetune_v1: Hypernetwork finetuned on the full data set minus portraits with face covering helmets, should promote much cleaner facial detail and image quality while still providing for a decent range of helmet generation. Negative prompting for helmet can also help if you find helmets are still generated too frequently.