GitHub项目FLUX.2开源了，支持 10 张图片同时参考的开源模型

# 创建虚拟环境 python -m venv flux2_env source flux2_env/bin/activate  # Linux/Mac # 或 flux2_env\Scripts\activate  # Windows  # 安装核心依赖 pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121 pip install diffusers transformers accelerate huggingface_hub  # 安装可选依赖（用于水印等功能） pip install invisible-watermark  

export FLUX2_DEV_PATH="/path/to/your/downloaded/weights" export FLUX2_AE_PATH="/path/to/autoencoder/ae.safetensors"  

huggingface-cli login  

git clone https://github.com/black-forest-labs/flux2.git cd flux2  # 高端 GPU（GB200） python -m flux2 --interactive  # H100 需要 CPU offloading python -m flux2 --interactive --cpu_offloading True  

> A serene lakeside cabin at sunset, with warm orange sky reflecting on calm water  

from flux2 import generate_image  prompt = "A serene lakeside cabin at sunset, with warm orange sky reflecting on calm water" image = generate_image(     prompt=prompt,     num_inference_steps=50,     guidance_scale=4.0,     seed=42 ) image.save("output.png")  

from flux2 import edit_image from PIL import Image  input_image = Image.open("input.jpg") prompt = "Transform this into a watercolor painting style"  edited_image = edit_image(     prompt=prompt,     image=input_image,     num_inference_steps=50,     guidance_scale=4.0 ) edited_image.save("edited_output.png")  

from flux2 import generate_from_multiple_images  reference_images = [     Image.open("style_ref.jpg"),     Image.open("pose_ref.jpg"),     Image.open("lighting_ref.jpg") ]  prompt = "Professional product photography combining these reference elements"  result = generate_from_multiple_images(     prompt=prompt,     reference_images=reference_images,     num_inference_steps=50 ) result.save("multi_ref_output.png")  

import torch from diffusers import Flux2Pipeline from diffusers.utils import load_image from huggingface_hub import get_token import requests import io  # 配置 repo_id = "diffusers/FLUX.2-dev-bnb-4bit" device = "cuda:0" torch_dtype = torch.bfloat16  # 远程文本编码器函数 def remote_text_encoder(prompts):     response = requests.post(         "https://remote-text-encoder-flux-2.huggingface.co/predict",         json={"prompt": prompts},         headers={             "Authorization": f"Bearer {get_token()}",             "Content-Type": "application/json"         }     )     prompt_embeds = torch.load(io.BytesIO(response.content))     return prompt_embeds.to(device)  # 加载量化模型（不包含文本编码器） pipe = Flux2Pipeline.from_pretrained(     repo_id,     text_encoder=None,     torch_dtype=torch_dtype ).to(device)  # 准备提示词 prompt = """Realistic macro photograph of a hermit crab using a soda can 
as its shell, partially emerging from the can, captured with sharp detail 
and natural colors, on a sunlit beach with soft shadows and a shallow 
depth of field, with blurred ocean waves in the background. The can has 
the text `BFL Diffusers` on it and it has a color gradient that starts 
with #FF5733 at the top and transitions to #33FF57 at the bottom."""  # 生成图像 image = pipe(     prompt_embeds=remote_text_encoder(prompt),     # 可选：添加参考图像     # image=load_image("https://example.com/reference.png"),     generator=torch.Generator(device=device).manual_seed(42),     num_inference_steps=50,  # 28 步是速度与质量的良好折衷     guidance_scale=4.0, ).images[0]  # 保存结果 image.save("flux2_output.png")  

from flux2 import generate_image_with_upsampling  simple_prompt = "a cat on a beach"  # 系统会自动将提示词扩展为更详细的描述 image = generate_image_with_upsampling(     prompt=simple_prompt,     use_local_upsampling=True,     num_inference_steps=50 ) image.save("upsampled_output.png")  

import os os.environ["OPENROUTER_API_KEY"] = "your_api_key_here"  image = generate_image_with_upsampling(     prompt=simple_prompt,     use_openrouter=True,     openrouter_model="anthropic/claude-3.5-sonnet",  # 或其他模型     num_inference_steps=50 )  

from flux2 import generate_image from invisible_watermark import WatermarkEncoder  # 生成图像 image = generate_image(prompt="your prompt here")  # 嵌入水印 encoder = WatermarkEncoder() encoder.set_watermark('bytes', 'your_watermark_data'.encode('utf-8')) watermarked_image = encoder.encode(image, 'dwtDct')  watermarked_image.save("watermarked_output.png")  

# 需要安装 c2pa-python # pip install c2pa-python  from c2pa import Builder  builder = Builder() builder.add_ingredient("flux2_generated") builder.add_assertion("stds.schema-org.CreativeWork", {     "author": "FLUX.2 [dev]",     "datePublished": "2024-11-25" })  # 为图像添加 C2PA 清单 builder.sign_file("output.png", "output_with_c2pa.png")  

from flux2 import generate_image  image = generate_image(     prompt="your prompt",     cpu_offloading=True,  # 启用 CPU offloading     num_inference_steps=50 )  

python -m flux2 --interactive --cpu_offloading True  

# 高质量模式 image_hq = generate_image(prompt, num_inference_steps=50)  # 快速模式 image_fast = generate_image(prompt, num_inference_steps=28)  

# 较低的 guidance (2.0-3.0) # - 更有创意和多样性 # - 可能偏离提示词  # 中等的 guidance (3.5-4.5) # - 平衡的表现（推荐） # - 既遵循提示又有创意  # 较高的 guidance (5.0-7.0) # - 严格遵循提示词 # - 可能过于死板  image = generate_image(     prompt="your prompt",     guidance_scale=4.0  # 推荐值 )  

import torch  # 设置固定种子 generator = torch.Generator(device="cuda").manual_seed(42)  # 生成可重复的结果 image1 = generate_image(prompt, generator=generator) image2 = generate_image(prompt, generator=generator)  # 相同的结果  

# 一次生成 4 张不同的图片 images = pipe(     prompt_embeds=remote_text_encoder(prompt),     num_images_per_prompt=4,     num_inference_steps=28,     guidance_scale=4.0 ).images  for i, img in enumerate(images):     img.save(f"output_{i}.png")