imgailab/flux1-torch-int8

FLUX.1 Schnell - PyTorch INT8

This is a INT8 pytorch version of black-forest-labs/FLUX.1-schnell.

Model Details

• Model Type: FLUX.1 Schnell
• Parameters: 12B
• Backend: PyTorch
• Quantization: INT8
• Memory Usage: ~12.0GB
• Conversion Date: 2025-08-09

Usage

PyTorch INT8

# PyTorch INT8 quantized model
from diffusers import FluxPipeline
import torch

# Load INT8 quantized model
pipe = FluxPipeline.from_pretrained(
    "Mitchins/flux1-torch-int8",
    torch_dtype=torch.qint8,
    use_safetensors=True
)

# For CPU inference
pipe = pipe.to("cpu")

# Generate image
image = pipe("A beautiful landscape", num_inference_steps=20).images[0]
image.save("output.png")

Performance

Backend	Quantization	Memory	Speed (CPU)	Speed (GPU)	Quality
PyTorch	INT8	~12.0GB	Good	Fast	Slightly Reduced

Limitations

• INT8 quantization may slightly reduce image quality
• Best suited for CPU inference or memory-constrained environments

Citation

@misc{flux1-pytorch-int8,
  title = {FLUX.1 Schnell PyTorch INT8}
  author = {ImageAI Server Contributors}
  year = {2024}
  publisher = {HuggingFace}
  url = {https://huggingface.co/Mitchins/flux1-torch-int8}
}

---

Converted using ImageAI Server Model Converter v1.0