File size: 3,889 Bytes
d25f2cc e4760c2 c0ea2b5 911cd56 c0ea2b5 e4760c2 d25f2cc cfc6838 a7b6832 fc287b1 bf1f799 c683cba 899e71e c683cba 74f294d 2d6b215 c9a6b2d 3a6ff2d 2d6b215 3a6ff2d c9a6b2d a959e12 74f294d a959e12 a09f287 74f294d a09f287 08a0bb2 ca6299e 1b3cfd6 08a0bb2 74f294d 08a0bb2 ca6299e 1b3cfd6 08a0bb2 74f294d 219b0d6 9b9816a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 |
---
language:
- en
tags:
- stable-diffusion
- stable-diffusion-diffusers
- text-to-image
license: creativeml-openrail-m
---
# !NEW! Starsector Portraits LORA
Added a LORA trained version you can find in the 4_LORA folder, the keyword to use in prompt was changed for the LORA to just **starsector**
Image showing the LORA used with different models for each row with the prompt "starsector photo of a _________" the blank filled with the word above each column.
---
# Starsector Portraits
This is a fine-tuned Stable Diffusion model trained on images from the videogame Starsector in order to provide mod makers with access to more easily creatable portraits in the style of the game.
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"
"an older man wearing a fancy suit with a purple cape, style of starsectorportrait"
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.
**1.5 trained model vs 1.4 trained model and hypernetwork presence**
**2.0 trained model vs 2.1 finetuned model**
**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_v2_1.5_13431.pt:** Hypernetwork finetuned for the 1.5 trained model on the full data set it greatly improves overall accuracy of the the generated portraits. I highly recommend using this along with the 1.5 model for all portrait generation.
**HN_ssportrait_rc_v1_no-helmet_finetune_v1:** Hypernetwork finetuned for the 1.4 trained model using the full data set minus images with faces that are covered, 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. |