FLUX_Shenhe_Lora

Shenhe lora Usage

Model Source

The LoRA model used in this project is sourced from:

TJ Flux Shenhe on CivitAI

Regional Flux Pipeline

The Regional Flux Pipeline utilized in this project is available at:

Regional Prompting FLUX on GitHub

Acknowledgments

We would like to express our sincere gratitude to the creators and contributors of the LoRA model and the Regional Flux Pipeline for their valuable work and resources.

Installtion

pip install -U diffusers transformers torch sentencepiece peft controlnet-aux moviepy protobuf

Demo

import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
pipe.load_lora_weights("svjack/FLUX_Shenhe_Lora")
pipe.enable_sequential_cpu_offload()

prompt = "tj_sthenhe, hair ornament,sliver hair,long hair,braid,"

image = pipe(prompt,
             num_inference_steps=24,
             guidance_scale=3.5,
            ).images[0]
image.save("shenhe.png")

from IPython import display
display.Image("shenhe.png", width=512, height=512)

Shenhe Use Regional Flux Pipeline README (Draw Shenhe in custom rectangle region)

This README provides a guide on how to use the Regional Flux Pipeline, a powerful tool for generating images with regional control using PyTorch. The pipeline allows you to specify different prompts for different regions of the image, enabling fine-grained control over the generated content.

Table of Contents

• Installation
• Usage
  • Step 1: Load the Pipeline
  • Step 2: Configure Attention Processors
  • Step 3: Set General Settings
  • Step 4: Define Regional Prompts and Masks
  • Step 5: Configure Region Control Factors
  • Step 6: Generate the Image
  • Step 7: Display the Image
  • Step 8: Draw a Transparent Rectangle
• Chinese Translations

Installation

Create a New Conda Environment

conda create --name py310 python=3.10 && conda activate py310 && pip install ipykernel && python -m ipykernel install --user --name py310 --display-name "py310"

Install Dependencies

We use a specific commit from the diffusers repository to ensure reproducibility, as newer versions may produce different results.

sudo apt-get update && sudo apt-get install git-lfs ffmpeg cbm

# Install diffusers locally
git clone https://github.com/huggingface/diffusers.git
cd diffusers

# Reset diffusers version to 0.31.dev
git reset --hard d13b0d63c0208f2c4c078c4261caf8bf587beb3b
pip install -e ".[torch]"
cd ..

# Install other dependencies
pip install -U transformers sentencepiece protobuf PEFT

# Clone this repo
git clone https://github.com/svjack/Regional-Prompting-FLUX

# Replace file in diffusers
cd Regional-Prompting-FLUX
cp transformer_flux.py ../diffusers/src/diffusers/models/transformers/transformer_flux.py
huggingface-cli login

Usage

Step 1: Load the Pipeline

First, load the Regional Flux Pipeline from a pretrained model and set the desired data type:

import torch
from pipeline_flux_regional import RegionalFluxPipeline, RegionalFluxAttnProcessor2_0

pipeline = RegionalFluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
pipeline.load_lora_weights("svjack/FLUX_Shenhe_Lora")
pipeline.to("cuda")

Step 2: Configure Attention Processors

Next, configure the attention processors to use the RegionalFluxAttnProcessor2_0 for specific attention layers:

attn_procs = {}
for name in pipeline.transformer.attn_processors.keys():
    if 'transformer_blocks' in name and name.endswith("attn.processor"):
        attn_procs[name] = RegionalFluxAttnProcessor2_0()
    else:
        attn_procs[name] = pipeline.transformer.attn_processors[name]
pipeline.transformer.set_attn_processor(attn_procs)

Step 3: Set General Settings

Define the general settings for the image generation:

image_width = 1280
image_height = 768
num_inference_steps = 24
seed = 124

base_prompt = "A snowy chinese hill in the background, A big sun rises."
background_prompt = "a photo of a snowy chinese hill"

Step 4: Define Regional Prompts and Masks

Specify the regional prompts and corresponding masks for different parts of the image:

regional_prompt_mask_pairs = {
    "0": {
        "description": "A dignified woman stands in the foreground, her sliver hair and long braid adorned with a hair ornament, her face illuminated by the cold light of the snow. Her expression is one of determination and sorrow, her clothing and appearance reflecting the historical period. The snow casts a serene yet dramatic light across her features, its cold embrace enveloping her in a world of ice and frost. tj_sthenhe, hair ornament, sliver hair, long hair, braid.",
        "mask": [128, 128, 640, 768]
    }
}

Step 5: Configure Region Control Factors

Set the control factors for region-specific attention injection:

mask_inject_steps = 10
double_inject_blocks_interval = 1
single_inject_blocks_interval = 1
base_ratio = 0.2

Step 6: Generate the Image

Generate the image using the specified prompts and masks:

regional_prompts = []
regional_masks = []
background_mask = torch.ones((image_height, image_width))

for region_idx, region in regional_prompt_mask_pairs.items():
    description = region['description']
    mask = region['mask']
    x1, y1, x2, y2 = mask
    mask = torch.zeros((image_height, image_width))
    mask[y1:y2, x1:x2] = 1.0
    background_mask -= mask
    regional_prompts.append(description)
    regional_masks.append(mask)

if background_mask.sum() > 0:
    regional_prompts.append(background_prompt)
    regional_masks.append(background_mask)

image = pipeline(
    prompt=base_prompt,
    width=image_width, height=image_height,
    mask_inject_steps=mask_inject_steps,
    num_inference_steps=num_inference_steps,
    generator=torch.Generator("cuda").manual_seed(seed),
    joint_attention_kwargs={
        "regional_prompts": regional_prompts,
        "regional_masks": regional_masks,
        "double_inject_blocks_interval": double_inject_blocks_interval,
        "single_inject_blocks_interval": single_inject_blocks_interval,
        "base_ratio": base_ratio
    },
).images[0]

image.save(f"shenhe_in_snow_hill.jpg")

Step 7: Display the Image

Display the generated image:

from IPython import display
display.Image("shenhe_in_snow_hill.jpg", width=512, height=512)

Step 8: Draw a Transparent Rectangle

Optionally, draw a transparent rectangle on the generated image to highlight a specific region:

from PIL import Image, ImageDraw

def draw_transparent_rectangle(image_path, bbox, color, alpha=128, output_path=None):
    """
    在指定区域绘制一个半透明的矩形，并将修改后的图片保存到本地新路径。

    :param image_path: 图片路径
    :param bbox: 长度为4的列表，表示矩形的边界框 [x1, y1, x2, y2]
    :param color: 颜色，格式为 (R, G, B)
    :param alpha: 透明度，范围为 0（完全透明）到 255（完全不透明），默认值为 128
    :param output_path: 保存修改后图片的路径，如果为 None，则覆盖原图
    :return: 修改后的图片对象
    """
    image = Image.open(image_path).convert("RGBA")
    overlay = Image.new('RGBA', image.size, (0, 0, 0, 0))
    draw = ImageDraw.Draw(overlay)

    x1, y1, x2, y2 = bbox
    draw.rectangle([x1, y1, x2, y2], fill=(*color, alpha))

    image = Image.alpha_composite(image, overlay)

    if output_path is None:
        output_path = image_path

    image.save(output_path)
    return image

draw_transparent_rectangle("shenhe_in_snow_hill.jpg", [128, 128, 640, 768], (255, 0, 0), alpha=128, output_path="shenhe_in_snow_hill_rec.png")
display.Image("shenhe_in_snow_hill_rec.png", width=512, height=512)

Chinese Translations

• base_prompt: "背景是雪中的中国山丘，一轮大太阳正在升起。"
• background_prompt: "一张雪中的中国山丘的照片"

regional_prompt_mask_pairs 中的内容翻译如下：

{
    "0": {
        "description": "一位端庄的女子站在前景中，她的银发和长辫子上装饰着发饰，她的脸被雪的冷光照亮。她的表情既坚定又悲伤，她的服装和外貌反映了历史时期。雪花在她脸上投下宁静而戏剧性的光线，它的寒冷拥抱将她包裹在冰雪世界中。tj_sthenhe，发饰，银发，长发，辫子。",
        "mask": [128, 128, 640, 768]
    }
}