ERNIE-8B-NF4 | Debloated
This repository contains the ERNIE-8B transformer weights, debloated and quantized to 4-bit NF4 using bitsandbytes for maximum efficiency on consumer-grade hardware (16GB VRAM limit).
🚀 Model Details
- Architecture: ERNIE Image Transformer (8B Parameters)
- Backbone: Mistral-3 Text Encoder
- Quantization: NF4 (NormalFloat 4-bit)
- Primary Use: High-fidelity image synthesis within the Asset Editor ecosystem.
🛠 Integration
This model is specifically tuned for the Asset Editor pipeline. It supports dynamic Trajectory Shifting (mu) to ensure coherency at various inference step counts.
Usage in Asset Editor:
- Ensure the
models/ernie-bnb-nf4 directory exists.
- The system will automatically detect and strike this manifold upon selection in the UI.
🔗 Links
Note: This is a specialized manifold optimized for the Veetance sovereign resource governor.
