flux-quantized-half

Model Overview

Silan10/flux-quantized-half is a half-precision (FP16) variant of the black-forest-labs/FLUX.1-dev text-to-image model. In this version, the transformers, text_encoder and text_encoder_2 folders have been converted to FP16.

This repository only changes the numerical precision of the weights to torch.float16 using PyTorch. This is not real quantization (like int8/int4). Still, converting the model to float16 saves memory, reduces RAM usage and speeds up loading times.

Usage

import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "Silan10/flux-quantized-half",
    torch_dtype=torch.float16
)
pipe.to("cuda")  # or pipe.enable_model_cpu_offload() for low VRAM

prompt = "Close-up portrait photo of a standing 30 year old female with twin braids hairstyle."
image = pipe(
    prompt,
    guidance_scale=3.5,
    num_inference_steps=20,
    generator=torch.Generator("cpu").manual_seed(0)
).images[0]

image.save("flux_half_sample.png")