‎
Blaze.1 Portrait Image Fidelity [ Base Flux.1 Merged ]
Blaze.1 Portrait is a flux-based adapter designed to generate high-quality, realism-driven super-portraits. This advanced model underwent rigorous training using a diverse set of images to ensure exceptional output in portrait generation. With 72 variations and a total of 1080 images per variation, the training process involved meticulous curation and processing of 14-bit raw Camera Uncompressed Format (CUF) images. For comprehensive insights, refer to the detailed documentation on training procedures, data curation, and parameter optimization.
Image Processing Parameters
Parameter | Value | Parameter | Value |
---|---|---|---|
LR Scheduler | Constant | Noise Offset | 0.03 |
Optimizer | AdamW | Multires Noise Discount | 0.1 |
Network Dim | 64 | Multires Noise Iterations | 10 |
Network Alpha | 32 | Repeat & Steps | 30 & 38,000 |
Epoch | 45 | Save Every N Epochs | 1 |
The model was trained with an overall image count of 1080 images, repeated across variations to create a total dataset of 1100 repetition per image.
Training was conducted for 4 epochs per set, resulting in a cumulative total of 38,000 steps, with the image labeler Florence2-en (natural language & English).
Dataset Details
The dataset consists of highly realistic images, meticulously curated for optimal model performance. Smaller adapter versions were initially trained on batches, and high-resolution images were later generated from these adapters for further training.
Pre-training upscaling techniques were employed to enhance pixel clarity, ensuring superior image fidelity during model inference.
Image compression details:
- Minimum compression: 1.4 MB
- Maximum compression: 10.1 MB
- Average compression: 3.2 MB
Training Infrastructure
Blaze.1 Portrait was trained using state-of-the-art hardware:
- GPU Configuration: NVIDIA A100 SXM (2x)
- vCPU: 32 cores
- VRAM: 160 GB
- RAM: 250 GB
- Disk Space: 400 GB
- Network Memory: High-speed interconnects for efficient data transfer
Training Runtime
Total machine runtime: 8 hours 14 minutes
Trigger Words
To generate images using Blaze.1 Portrait, utilize the trigger phrase:
- b1 portrait
Incorporating this trigger in your prompts ensures optimal model performance and high-quality portrait generation.
Download Model
Weights for Blaze.1 Portrait are available in Safetensors format. You can download them from the Files & Versions tab using the link below:
Download Blaze.1 Portrait Weights
@misc{strangerzonehf2024blaze1,
title={Blaze.1 Portrait: High-Quality Realism-Driven Portrait Generation with Flux-Based LoRA Adapter},
author={StrangerzoneHF},
year={2024},
note={public model repository on Hugging Face},
url={https://huggingface.co/strangerzonehf/Blaze.1-Portrait},
model_trainer={@prithivMLmods},
dataset_curator={@prithivMLmods}
- Downloads last month
- 2
Model tree for strangerzonehf/Blaze.1-Portrait
Base model
prithivMLmods/Flux.1-Merged