metadata

license: other
license_name: faipl-1.0-sd
license_link: https://freedevproject.org/faipl-1.0-sd/
base_model:
  - Laxhar/sdxl_noob

Illumina-NoobVpd

Fine-tuned NoobAI-XL(v-prediction) and merged SPO

Requirements

AUTOMATIC1111 WebUI on dev branch
Latest ComfyUI
ReForge on dev_upstream_experimental branch

Instruction for AUTOMATIC1111

Switch branch to dev
Copy configs/sd_xl_v.yaml to models/Stable-Diffusion/
Rename it to the same as the model name

Instruction for ReForge

Switch branch to dev_upstream_experimental
Find “Advanced Model Sampling for Forge” at the bottom of the page
Enable “Enable Advanced Model Sampling”
Select v_prediction in Discrete Sampling Type

Example Workflow for ComfyUI

Download it from here

Prompt Guidelines

Almost same as the base model

To improve the quality of background, add simple background, transparent background to Negative Prompt.

Recommended Prompt

standard

Positive: None(Works good without masterpiece, best quality)

Negative: worst quality, low quality, bad quality, lowres, jpeg artifacts, unfinished, oldest, old, photoshop \(medium\), abstract

Recommended Settings

Steps: 14-28

Sampler: DPM++ 2M(dpmpp_2m)

Scheduler: Simple

Guidance Scale: 4-9

Hires.fix

Hires upscaler: 4x-UltraSharp or Latent(nearest-exact)

Denoising strength: 0.4-0.5(0.6 for latent)

Merge recipe(Weighted sum)

I made 6 Illustrious-based models and merged them.

Stage 0: finetunes v-pred test model with AI-generated images
Stage 1: finetunes stage 0 model with 300 scenery images from Gelbooru
Stage 2:

*A-F: finetuned stage1(ReLoRA)

A * 0.6 + B * 0.4 = tmp1
tmp1 * 0.6 + C * 0.4 = tmp2
tmp2 * 0.7 + F * 0.3 = tmp3
tmp3 * 0.7 + E * 0.3 = tmp4
tmp4 * 0.5 + D * 0.5 = tmp5
tmp5 * 0.65 + sd15 * 0.35 = tmp6
tmp6 + SPO LoRA = Result

Training scripts:

sd-scripts

Notice

This model is licensed under Fair AI Public License 1.0-SD

If you make modify this model, you must share both your changes and the original license.

You are prohibited from monetizing any close-sourced fine-tuned / merged model, which disallows the public from accessing the model's source code / weights and its usages.