Add Readme
Browse files
README.md
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: diffusers
|
| 3 |
+
pipeline_tag: text-to-image
|
| 4 |
+
tags:
|
| 5 |
+
- stable-diffusion-xl
|
| 6 |
+
- v-prediction
|
| 7 |
+
- anime
|
| 8 |
+
- sparse-merge
|
| 9 |
+
- epsm
|
| 10 |
+
license: openrail++
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# Project-Cosmos-V1.2 (EPSM Edition)
|
| 14 |
+
|
| 15 |
+
**Project Cosmos V1.2** introduces **Energy-Preserving Sparse Merge (EPSM)**. It retains the clean, noise-free anatomy of V1.1 but restores the "Style Volume" of NoobAI that was lost during filtration.
|
| 16 |
+
|
| 17 |
+
## The V1.2 Upgrade
|
| 18 |
+
* **Smart Boost (Output Blocks):** In V1.1, filtering noise also reduced the overall intensity of the NoobAI style. In V1.2, we calculate the energy lost during filtering and scientifically boost the remaining clean signal to match the original intensity.
|
| 19 |
+
* **Safety Clamp:** Unlike the experimental V1.0, the boost is mathematically clamped (Max 3.5x) to prevent any possibility of artifact explosion.
|
| 20 |
+
* **Anatomy Protection:** Input and Middle blocks are NOT boosted. They remain in strict "Clean Mode" to ensure perfect fingers and poses.
|
| 21 |
+
|
| 22 |
+
## Recommended Settings
|
| 23 |
+
* **Sampler:** Euler a / DPM++ 2M SDE
|
| 24 |
+
* **CFG:** 5.0 - 8.0 (Colors will be more vibrant than V1.1)
|
| 25 |
+
* **Steps:** 28+
|