SpacesFusion_V2

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.