|
|
--- |
|
|
library_name: diffusers |
|
|
pipeline_tag: text-to-image |
|
|
tags: |
|
|
- stable-diffusion-xl |
|
|
- v-prediction |
|
|
- anime |
|
|
- slerp-merge |
|
|
- masterpiece |
|
|
license: openrail++ |
|
|
--- |
|
|
|
|
|
# Project-Cosmos-V2.0 (The Masterpiece) |
|
|
|
|
|
**Cosmos V2.0** is the culmination of our research. It moves away from destructive pruning (DARE/TIES) and utilizes **Segmented Spherical Linear Interpolation (SLERP)** to create a perfect hybrid. |
|
|
|
|
|
## The "Cosmos" Formula |
|
|
|
|
|
1. **Resolution Lock (1536px Support):** |
|
|
* We identified the **Attention Layers (`attn`)** responsible for spatial resolution. |
|
|
* We skewed these layers **65% towards IllumiYume**. This preserves IllumiYume's capability to generate high-resolution images without artifacts. |
|
|
|
|
|
2. **Super Anatomy (50/50 SLERP):** |
|
|
* Since both NoobAI and IllumiYume have excellent anatomy, we didn't discard either. |
|
|
* We used **SLERP (0.5)** on the Input & Middle blocks. This creates a "perfect average" of both skeletal structures, resulting in robust poses and fingers that are better than either model individually. |
|
|
|
|
|
3. **Anti-Burn Style Engine (90% NoobAI):** |
|
|
* The Output Blocks (Style/Color) are forced to **90% NoobAI** using Linear Interpolation. |
|
|
* This completely eliminates the high-contrast "burning" issue of IllumiYume at high CFG. |
|
|
* **Result:** You can use CFG 7-10 safely with clean, soft lighting. |
|
|
|
|
|
## Recommended Settings |
|
|
* **Sampler:** Euler a / DPM++ 2M SDE |
|
|
* **CFG:** 5.0 - 9.0 (Very stable!) |
|
|
* **Steps:** 28+ |
|
|
* **Resolution:** Supports up to 1536x1536 natively. |
|
|
|
