【Diffusers】いろいろな IP-Adapter を組み合わせてみる

はじめに

こちらの拡張です。
touch-sp.hatenablog.com
「FacaID」が使えるようになったので3つ同時に使用してみました。

命題

この女性をイラスト化して下さい。

結果

左から
「Plus」と「Plus Face」の組み合わせ
「Plus」と「FaceID」の組み合わせ
「Plus」と「Plus Face」と「FaceID」の組み合わせ

方法

最初にそれぞれのImage Embeddingsを計算して保存しました。

Image Embeddingsの作成、保存

「Plus」のImage Embeddings

import torch
from diffusers import AutoPipelineForText2Image
from diffusers.utils import load_image
from pathlib import Path

pipeline = AutoPipelineForText2Image.from_pretrained(
    "model/modernDisneyXL_v3",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipeline.load_ip_adapter(
    "IP-Adapter",
    subfolder="sdxl_models",
    weight_name="ip-adapter-plus_sdxl_vit-h.safetensors",
    image_encoder_folder="models/image_encoder"
)
pipeline.to("cuda")

style_images = [load_image(x.as_posix()) for x in Path("style_ziggy").glob("*.png")]

image_embeds = pipeline.prepare_ip_adapter_image_embeds(
    ip_adapter_image=[style_images],
    ip_adapter_image_embeds=None,
    device="cuda",
    num_images_per_prompt=1,
    do_classifier_free_guidance=True
)
torch.save(image_embeds, "xl_plus.ipadpt")

「Plus Face」のImage Embeddings

import torch
from diffusers import AutoPipelineForText2Image
from diffusers.utils import load_image
from pathlib import Path

pipeline = AutoPipelineForText2Image.from_pretrained(
    "model/modernDisneyXL_v3",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipeline.load_ip_adapter(
    "IP-Adapter",
    subfolder="sdxl_models",
    weight_name="ip-adapter-plus-face_sdxl_vit-h.safetensors",
    image_encoder_folder="models/image_encoder"
)
pipeline.to("cuda")

face_images = [load_image(x.as_posix()) for x in Path("woman").glob("*.jpg")]

image_embeds = pipeline.prepare_ip_adapter_image_embeds(
    ip_adapter_image=[face_images],
    ip_adapter_image_embeds=None,
    device="cuda",
    num_images_per_prompt=1,
    do_classifier_free_guidance=True
)
torch.save(image_embeds, "xl_plusface.ipadpt")

「FaceID」のImage Embeddings

import torch
from diffusers.utils import load_image
from pathlib import Path
from insightface.app import FaceAnalysis
import cv2
import numpy as np

face_images = [load_image(x.as_posix()) for x in Path("woman").glob("*.jpg")]

ref_images_embeds = []
app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
app.prepare(ctx_id=0, det_size=(640, 640))
for im in face_images:
    image = cv2.cvtColor(np.asarray(im), cv2.COLOR_BGR2RGB)
    faces = app.get(image)
    if len(faces) != 0:
        image = torch.from_numpy(faces[0].normed_embedding)
        ref_images_embeds.append(image.unsqueeze(0))

print(f"{len(face_images)} face images, {len(ref_images_embeds)} faces detected")
assert len(ref_images_embeds) > 0, "face detection failed."

ref_images_embeds = torch.stack(ref_images_embeds, dim=0).unsqueeze(0)
neg_ref_images_embeds = torch.zeros_like(ref_images_embeds)
id_embeds = torch.cat([neg_ref_images_embeds, ref_images_embeds]).to(dtype=torch.float16, device="cuda")

torch.save([id_embeds], "xl_faceid.ipadpt")

実行

「Plus」と「Plus Face」の組み合わせ

import torch
from diffusers.utils import load_image
from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler

pipeline = AutoPipelineForText2Image.from_pretrained(
    "model/modernDisneyXL_v3",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipeline.scheduler = DPMSolverMultistepScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    steps_offset=1,
    algorithm_type="sde-dpmsolver++",
    use_karras_sigmas=True
)
pipeline.to("cuda")

pipeline.load_ip_adapter(
    "IP-Adapter",
    subfolder="sdxl_models",
    weight_name=[
        "ip-adapter-plus_sdxl_vit-h.safetensors",
        "ip-adapter-plus-face_sdxl_vit-h.safetensors"
    ],
    image_encoder_folder=None
)
pipeline.set_ip_adapter_scale([0.7, 0.8])

t1 = torch.load("xl_plus.ipadpt")
t2 = torch.load("xl_plusface.ipadpt")
image_embeds = [t1[0], t2[0]]

image = pipeline(
    prompt="a woman",
    ip_adapter_image_embeds=image_embeds,
    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
    num_inference_steps=50,
    num_images_per_prompt=1,
    guidance_scale = 7.5,
    width=1024,
    height=1024, 
    generator=torch.Generator(device="cpu").manual_seed(0)
).images[0]

image.save("plusface.png")

「Plus」と「FaceID」の組み合わせ

import torch
from diffusers.utils import load_image
from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler

pipeline = AutoPipelineForText2Image.from_pretrained(
    "model/modernDisneyXL_v3",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipeline.scheduler = DPMSolverMultistepScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    steps_offset=1,
    algorithm_type="sde-dpmsolver++",
    use_karras_sigmas=True
)
pipeline.to("cuda")

pipeline.load_ip_adapter(
    ["IP-Adapter", "IP-Adapter-FaceID"],
    subfolder=["sdxl_models", None],
    weight_name=[
        "ip-adapter-plus_sdxl_vit-h.safetensors",
        "ip-adapter-faceid_sdxl.bin"
    ],
    image_encoder_folder=None
)
pipeline.set_ip_adapter_scale([0.3, 0.8])

t1 = torch.load("xl_plus.ipadpt")
t2 = torch.load("xl_faceid.ipadpt")
image_embeds = [t1[0], t2[0]]

image = pipeline(
    prompt="a woman",
    ip_adapter_image_embeds=image_embeds,
    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
    num_inference_steps=50,
    num_images_per_prompt=1,
    guidance_scale = 7.5,
    width=1024,
    height=1024, 
    generator=torch.Generator(device="cpu").manual_seed(0)
).images[0]

image.save("faceid.png")

「Plus」と「Plus Face」と「FaceID」の組み合わせ

import torch
from diffusers.utils import load_image
from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler

pipeline = AutoPipelineForText2Image.from_pretrained(
    "model/modernDisneyXL_v3",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipeline.scheduler = DPMSolverMultistepScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    steps_offset=1,
    algorithm_type="sde-dpmsolver++",
    use_karras_sigmas=True
)
pipeline.to("cuda")

pipeline.load_ip_adapter(
    ["IP-Adapter", "IP-Adapter", "IP-Adapter-FaceID"],
    subfolder=["sdxl_models", "sdxl_models", None],
    weight_name=[
        "ip-adapter-plus_sdxl_vit-h.safetensors",
        "ip-adapter-plus-face_sdxl_vit-h.safetensors",
        "ip-adapter-faceid_sdxl.bin"
    ],
    image_encoder_folder=None
)
pipeline.set_ip_adapter_scale([0.5, 0.5, 0.5])

t1 = torch.load("xl_plus.ipadpt")
t2 = torch.load("xl_plusface.ipadpt")
t3 = torch.load("xl_faceid.ipadpt")

image_embeds = [t1[0], t2[0], t3[0]]

image = pipeline(
    prompt="a woman",
    ip_adapter_image_embeds=image_embeds,
    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
    num_inference_steps=50,
    num_images_per_prompt=1,
    guidance_scale = 7.5,
    width=1024,
    height=1024, 
    generator=torch.Generator(device="cpu").manual_seed(0)
).images[0]

image.save("plusface_faceid.png")