|
|
--- |
|
|
license: creativeml-openrail-m |
|
|
tags: |
|
|
- stable-diffusion |
|
|
- sdxl |
|
|
- stable-diffusion-xl |
|
|
- text-to-image |
|
|
- illustration |
|
|
- stylized |
|
|
- merged-model |
|
|
- checkpoint |
|
|
pipeline_tag: text-to-image |
|
|
--- |
|
|
|
|
|
# Model Card |
|
|
|
|
|
This is a **custom SDXL checkpoint mix** built on top of **Hoseki’s _hosekiLustrousmix_illusNoobaiEPS11V2_**, enhanced with a curated set of LoRAs (self-trained + community-sourced) and merged using a **custom block-weight merging method**. |
|
|
|
|
|
The goal of this mix is to preserve the base model’s strong illustration qualities while selectively injecting specific stylistic influences from multiple LoRAs in a controlled, stable way. |
|
|
|
|
|
--- |
|
|
|
|
|
## Base Model |
|
|
|
|
|
- **Base:** Hoseki – _hosekiLustrousmix_illusNoobaiEPS11V2_ |
|
|
- **Model Type:** Stable Diffusion XL (SDXL) |
|
|
- **Role:** Primary illustration and style foundation |
|
|
|
|
|
--- |
|
|
|
|
|
## LoRA Influences |
|
|
|
|
|
This checkpoint includes embedded LoRA influence from the following sources: |
|
|
|
|
|
### Self-trained LoRAs |
|
|
- korukomo |
|
|
- old west pulp novel art |
|
|
- art of mathew |
|
|
- azyas |
|
|
- chesshir |
|
|
- datcr |
|
|
- dhsv2 |
|
|
- edh |
|
|
- higashino |
|
|
|
|
|
### Community / Downloaded LoRAs |
|
|
Some LoRAs were sourced from public repositories and merged into the final checkpoint as style contributors. |
|
|
|
|
|
--- |
|
|
|
|
|
## Merge Methodology (Custom Block-Weight Merge) |
|
|
|
|
|
This model was created using a **custom SDXL block-weight merging process**. |
|
|
|
|
|
Rather than applying each LoRA uniformly across the network, each contributor was merged using **individually tuned SDXL block weights**, selected based on how that contributor improved the final result. This helps: |
|
|
|
|
|
- preserve base-model structure and consistency |
|
|
- avoid over-dominant LoRA influence |
|
|
- reduce artifacts and instability |
|
|
- target certain influences more strongly in specific regions of the model (composition vs texture vs style, etc.) |
|
|
|
|
|
This results in a **single consolidated SDXL checkpoint**, rather than requiring the LoRAs at inference time. |
|
|
|
|
|
--- |
|
|
|
|
|
## Intended Use |
|
|
|
|
|
This model is best suited for: |
|
|
|
|
|
✅ stylized illustration |
|
|
✅ graphic / painterly rendering |
|
|
✅ expressive linework |
|
|
✅ pulp-inspired / genre aesthetics |
|
|
✅ creative experimentation |
|
|
|
|
|
It is **not designed for photorealism**. |
|
|
|
|
|
--- |
