File size: 1,261 Bytes
c59da89
ebd4da2
 
 
 
 
 
c59da89
 
 
 
 
 
447ccda
 
 
 
 
ee4ce9e
6385211
447ccda
 
fc26347
0715887
447ccda
6c56669
447ccda
 
 
 
 
 
2a38a3a
 
447ccda
 
 
462e3fb
447ccda
 
 
 
2a38a3a
447ccda
 
c59da89
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
---
tags:
- text-to-image
- flux
- diffusers
- quantization
license: other
language:
- en
base_model:
- black-forest-labs/FLUX.1-dev
pipeline_tag: text-to-image
---

## Model Overview

`Silan10/flux_quantized_half` is a **half-precision (FP16) variant** of the
[`black-forest-labs/FLUX.1-dev`](https://huggingface.co/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

```python
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")