| | --- |
| | license: other |
| | library_name: diffusers |
| | tags: |
| | - text-to-image |
| | - stable-diffusion-xl |
| | - diffusers |
| | - slerp |
| | - merge |
| | base_model: |
| | - NoobAI-XL-Vpred |
| | - IllumiYume-XL-Vpred |
| | --- |
| | |
| | # SpacesFusion_V2 |
| | |
| | **SpacesFusion_V2** is the evolved iteration of the SpacesFusion project. It addresses the "Knowledge Gap" found in V1 while refining high-resolution stability. |
| | |
| | ## Changlog from V1 |
| | * **Fix:** Significantly improved recognition of 2024-2025 concepts (e.g., Marvel Rivals, newer anime characters) by injecting IllumiYume's Text Encoder weights. |
| | * **Fix:** Reduced anatomical mutations at 1536px+ resolutions by damping the Output Block ratios at high frequencies. |
| | * **Feature:** Retains the superior NoobAI skeletal anatomy. |
| | |
| | ## Composition Strategy (The "Cerebral Transplant") |
| | |
| | * **Text Encoders (CLIP):** **65% IllumiYume** (Primary change). This allows the model to "understand" newer tags. |
| | * **Input Blocks (Structure):** **10-25% IllumiYume**. Kept low to preserve NoobAI's superior anatomy. |
| | * **Middle Block (Concept):** **48% IllumiYume**. Increased significantly to allow concept transfer. |
| | * **Output Blocks (Texture):** **45% -> 30% Damping Curve**. Starts high for style, drops lower at the end to prevent high-res artifacting. |
| | |
| | ## Recommended Settings |
| | * **Resolution:** 832x1216 (Portrait), 1024x1024 (Square), 1536x1536 (High Res Supported) |
| | * **CFG:** 5-7 |
| | * **Sampler:** Euler a / DPM++ 2M SDE Karras |
| | * **V-Prediction:** **ENABLED** (Critical) |
| | |
| | Created by **Bl4ckSpaces** using Automated Evolutionary SLERP V2 Protocol. |
| | |
