|
|
--- |
|
|
library_name: diffusers |
|
|
pipeline_tag: text-to-image |
|
|
tags: |
|
|
- stable-diffusion-xl |
|
|
- v-prediction |
|
|
- anime |
|
|
- surgical-merge |
|
|
- high-resolution |
|
|
license: openrail++ |
|
|
--- |
|
|
|
|
|
# Project-Cosmos-V2.1 (Resolution Transplant) |
|
|
|
|
|
**Cosmos V2.1** marks a strategic pivot. We identified that **NoobAI XL** has superior anatomy and style compared to previous bases. However, NoobAI sometimes struggles with very high resolutions (1536px+). |
|
|
|
|
|
## The Method: Surgical Attention Transplant |
|
|
* **Base:** **NoobAI XL v1.0** (100% of Anatomy, ResNets, Style, and Text Encoders). |
|
|
* **Donor:** **IllumiYume v3.5** (Only providing Spatial Awareness). |
|
|
* **Technique:** We applied **SLERP (50%)** *only* to the **Attention Layers** (`attn`/`transformer`). These layers control how the model handles pixel coherence at scale. |
|
|
|
|
|
## The Result |
|
|
* **Anatomy:** Pure NoobAI (Perfect fingers, good proportions). |
|
|
* **Style:** Pure NoobAI (Soft lighting, no burns). |
|
|
* **Resolution:** Inherited IllumiYume's ability to render native **1536x1536** without duplicating heads or breaking bodies. |
|
|
|
|
|
## Recommended Settings |
|
|
* **Sampler:** Euler a / DPM++ 2M SDE |
|
|
* **CFG:** 5.0 - 8.0 |
|
|
* **Resolution:** Tested up to 1536x1536 (Native). |
|
|
|
