gradio_demo_bg.py

import os
import math
import gradio as gr
import numpy as np
import torch
import safetensors.torch as sf
import db_examples

from PIL import Image
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
from diffusers import AutoencoderKL, UNet2DConditionModel, DDIMScheduler, EulerAncestralDiscreteScheduler, DPMSolverMultistepScheduler
from diffusers.models.attention_processor import AttnProcessor2_0
from transformers import CLIPTextModel, CLIPTokenizer
from briarmbg import BriaRMBG
from enum import Enum
from torch.hub import download_url_to_file


# 'stablediffusionapi/realistic-vision-v51'
# 'runwayml/stable-diffusion-v1-5'
sd15_name = 'stablediffusionapi/realistic-vision-v51'
tokenizer = CLIPTokenizer.from_pretrained(sd15_name, subfolder="tokenizer")
text_encoder = CLIPTextModel.from_pretrained(sd15_name, subfolder="text_encoder")
vae = AutoencoderKL.from_pretrained(sd15_name, subfolder="vae")
unet = UNet2DConditionModel.from_pretrained(sd15_name, subfolder="unet")
rmbg = BriaRMBG.from_pretrained("briaai/RMBG-1.4")

# Change UNet

with torch.no_grad():
    new_conv_in = torch.nn.Conv2d(12, unet.conv_in.out_channels, unet.conv_in.kernel_size, unet.conv_in.stride, unet.conv_in.padding)
    new_conv_in.weight.zero_()
    new_conv_in.weight[:, :4, :, :].copy_(unet.conv_in.weight)
    new_conv_in.bias = unet.conv_in.bias
    unet.conv_in = new_conv_in

unet_original_forward = unet.forward


def hooked_unet_forward(sample, timestep, encoder_hidden_states, **kwargs):
    c_concat = kwargs['cross_attention_kwargs']['concat_conds'].to(sample)
    c_concat = torch.cat([c_concat] * (sample.shape[0] // c_concat.shape[0]), dim=0)
    new_sample = torch.cat([sample, c_concat], dim=1)
    kwargs['cross_attention_kwargs'] = {}
    return unet_original_forward(new_sample, timestep, encoder_hidden_states, **kwargs)


unet.forward = hooked_unet_forward

# Load

model_path = './models/iclight_sd15_fbc.safetensors'

if not os.path.exists(model_path):
    download_url_to_file(url='https://huggingface.co/lllyasviel/ic-light/resolve/main/iclight_sd15_fbc.safetensors', dst=model_path)

sd_offset = sf.load_file(model_path)
sd_origin = unet.state_dict()
keys = sd_origin.keys()
sd_merged = {k: sd_origin[k] + sd_offset[k] for k in sd_origin.keys()}
unet.load_state_dict(sd_merged, strict=True)
del sd_offset, sd_origin, sd_merged, keys

# Device

device = torch.device('cuda')
text_encoder = text_encoder.to(device=device, dtype=torch.float16)
vae = vae.to(device=device, dtype=torch.bfloat16)
unet = unet.to(device=device, dtype=torch.float16)
rmbg = rmbg.to(device=device, dtype=torch.float32)

# SDP

unet.set_attn_processor(AttnProcessor2_0())
vae.set_attn_processor(AttnProcessor2_0())

# Samplers

ddim_scheduler = DDIMScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    clip_sample=False,
    set_alpha_to_one=False,
    steps_offset=1,
)

euler_a_scheduler = EulerAncestralDiscreteScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    steps_offset=1
)

dpmpp_2m_sde_karras_scheduler = DPMSolverMultistepScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    algorithm_type="sde-dpmsolver++",
    use_karras_sigmas=True,
    steps_offset=1
)

# Pipelines

t2i_pipe = StableDiffusionPipeline(
    vae=vae,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    unet=unet,
    scheduler=dpmpp_2m_sde_karras_scheduler,
    safety_checker=None,
    requires_safety_checker=False,
    feature_extractor=None,
    image_encoder=None
)

i2i_pipe = StableDiffusionImg2ImgPipeline(
    vae=vae,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    unet=unet,
    scheduler=dpmpp_2m_sde_karras_scheduler,
    safety_checker=None,
    requires_safety_checker=False,
    feature_extractor=None,
    image_encoder=None
)


@torch.inference_mode()
def encode_prompt_inner(txt: str):
    max_length = tokenizer.model_max_length
    chunk_length = tokenizer.model_max_length - 2
    id_start = tokenizer.bos_token_id
    id_end = tokenizer.eos_token_id
    id_pad = id_end

    def pad(x, p, i):
        return x[:i] if len(x) >= i else x + [p] * (i - len(x))

    tokens = tokenizer(txt, truncation=False, add_special_tokens=False)["input_ids"]
    chunks = [[id_start] + tokens[i: i + chunk_length] + [id_end] for i in range(0, len(tokens), chunk_length)]
    chunks = [pad(ck, id_pad, max_length) for ck in chunks]

    token_ids = torch.tensor(chunks).to(device=device, dtype=torch.int64)
    conds = text_encoder(token_ids).last_hidden_state

    return conds


@torch.inference_mode()
def encode_prompt_pair(positive_prompt, negative_prompt):
    c = encode_prompt_inner(positive_prompt)
    uc = encode_prompt_inner(negative_prompt)

    c_len = float(len(c))
    uc_len = float(len(uc))
    max_count = max(c_len, uc_len)
    c_repeat = int(math.ceil(max_count / c_len))
    uc_repeat = int(math.ceil(max_count / uc_len))
    max_chunk = max(len(c), len(uc))

    c = torch.cat([c] * c_repeat, dim=0)[:max_chunk]
    uc = torch.cat([uc] * uc_repeat, dim=0)[:max_chunk]

    c = torch.cat([p[None, ...] for p in c], dim=1)
    uc = torch.cat([p[None, ...] for p in uc], dim=1)

    return c, uc


@torch.inference_mode()
def pytorch2numpy(imgs, quant=True):
    results = []
    for x in imgs:
        y = x.movedim(0, -1)

        if quant:
            y = y * 127.5 + 127.5
            y = y.detach().float().cpu().numpy().clip(0, 255).astype(np.uint8)
        else:
            y = y * 0.5 + 0.5
            y = y.detach().float().cpu().numpy().clip(0, 1).astype(np.float32)

        results.append(y)
    return results


@torch.inference_mode()
def numpy2pytorch(imgs):
    h = torch.from_numpy(np.stack(imgs, axis=0)).float() / 127.0 - 1.0  # so that 127 must be strictly 0.0
    h = h.movedim(-1, 1)
    return h


def resize_and_center_crop(image, target_width, target_height):
    pil_image = Image.fromarray(image)
    original_width, original_height = pil_image.size
    scale_factor = max(target_width / original_width, target_height / original_height)
    resized_width = int(round(original_width * scale_factor))
    resized_height = int(round(original_height * scale_factor))
    resized_image = pil_image.resize((resized_width, resized_height), Image.LANCZOS)
    left = (resized_width - target_width) / 2
    top = (resized_height - target_height) / 2
    right = (resized_width + target_width) / 2
    bottom = (resized_height + target_height) / 2
    cropped_image = resized_image.crop((left, top, right, bottom))
    return np.array(cropped_image)


def resize_without_crop(image, target_width, target_height):
    pil_image = Image.fromarray(image)
    resized_image = pil_image.resize((target_width, target_height), Image.LANCZOS)
    return np.array(resized_image)


@torch.inference_mode()
def run_rmbg(img, sigma=0.0):
    H, W, C = img.shape
    assert C == 3
    k = (256.0 / float(H * W)) ** 0.5
    feed = resize_without_crop(img, int(64 * round(W * k)), int(64 * round(H * k)))
    feed = numpy2pytorch([feed]).to(device=device, dtype=torch.float32)
    alpha = rmbg(feed)[0][0]
    alpha = torch.nn.functional.interpolate(alpha, size=(H, W), mode="bilinear")
    alpha = alpha.movedim(1, -1)[0]
    alpha = alpha.detach().float().cpu().numpy().clip(0, 1)
    result = 127 + (img.astype(np.float32) - 127 + sigma) * alpha
    return result.clip(0, 255).astype(np.uint8), alpha


@torch.inference_mode()
def process(input_fg, input_bg, prompt, image_width, image_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source):
    bg_source = BGSource(bg_source)

    if bg_source == BGSource.UPLOAD:
        pass
    elif bg_source == BGSource.UPLOAD_FLIP:
        input_bg = np.fliplr(input_bg)
    elif bg_source == BGSource.GREY:
        input_bg = np.zeros(shape=(image_height, image_width, 3), dtype=np.uint8) + 64
    elif bg_source == BGSource.LEFT:
        gradient = np.linspace(224, 32, image_width)
        image = np.tile(gradient, (image_height, 1))
        input_bg = np.stack((image,) * 3, axis=-1).astype(np.uint8)
    elif bg_source == BGSource.RIGHT:
        gradient = np.linspace(32, 224, image_width)
        image = np.tile(gradient, (image_height, 1))
        input_bg = np.stack((image,) * 3, axis=-1).astype(np.uint8)
    elif bg_source == BGSource.TOP:
        gradient = np.linspace(224, 32, image_height)[:, None]
        image = np.tile(gradient, (1, image_width))
        input_bg = np.stack((image,) * 3, axis=-1).astype(np.uint8)
    elif bg_source == BGSource.BOTTOM:
        gradient = np.linspace(32, 224, image_height)[:, None]
        image = np.tile(gradient, (1, image_width))
        input_bg = np.stack((image,) * 3, axis=-1).astype(np.uint8)
    else:
        raise 'Wrong background source!'

    rng = torch.Generator(device=device).manual_seed(seed)

    fg = resize_and_center_crop(input_fg, image_width, image_height)
    bg = resize_and_center_crop(input_bg, image_width, image_height)
    concat_conds = numpy2pytorch([fg, bg]).to(device=vae.device, dtype=vae.dtype)
    concat_conds = vae.encode(concat_conds).latent_dist.mode() * vae.config.scaling_factor
    concat_conds = torch.cat([c[None, ...] for c in concat_conds], dim=1)

    conds, unconds = encode_prompt_pair(positive_prompt=prompt + ', ' + a_prompt, negative_prompt=n_prompt)

    latents = t2i_pipe(
        prompt_embeds=conds,
        negative_prompt_embeds=unconds,
        width=image_width,
        height=image_height,
        num_inference_steps=steps,
        num_images_per_prompt=num_samples,
        generator=rng,
        output_type='latent',
        guidance_scale=cfg,
        cross_attention_kwargs={'concat_conds': concat_conds},
    ).images.to(vae.dtype) / vae.config.scaling_factor

    pixels = vae.decode(latents).sample
    pixels = pytorch2numpy(pixels)
    pixels = [resize_without_crop(
        image=p,
        target_width=int(round(image_width * highres_scale / 64.0) * 64),
        target_height=int(round(image_height * highres_scale / 64.0) * 64))
    for p in pixels]

    pixels = numpy2pytorch(pixels).to(device=vae.device, dtype=vae.dtype)
    latents = vae.encode(pixels).latent_dist.mode() * vae.config.scaling_factor
    latents = latents.to(device=unet.device, dtype=unet.dtype)

    image_height, image_width = latents.shape[2] * 8, latents.shape[3] * 8
    fg = resize_and_center_crop(input_fg, image_width, image_height)
    bg = resize_and_center_crop(input_bg, image_width, image_height)
    concat_conds = numpy2pytorch([fg, bg]).to(device=vae.device, dtype=vae.dtype)
    concat_conds = vae.encode(concat_conds).latent_dist.mode() * vae.config.scaling_factor
    concat_conds = torch.cat([c[None, ...] for c in concat_conds], dim=1)

    latents = i2i_pipe(
        image=latents,
        strength=highres_denoise,
        prompt_embeds=conds,
        negative_prompt_embeds=unconds,
        width=image_width,
        height=image_height,
        num_inference_steps=int(round(steps / highres_denoise)),
        num_images_per_prompt=num_samples,
        generator=rng,
        output_type='latent',
        guidance_scale=cfg,
        cross_attention_kwargs={'concat_conds': concat_conds},
    ).images.to(vae.dtype) / vae.config.scaling_factor

    pixels = vae.decode(latents).sample
    pixels = pytorch2numpy(pixels, quant=False)

    return pixels, [fg, bg]


@torch.inference_mode()
def process_relight(input_fg, input_bg, prompt, image_width, image_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source):
    input_fg, matting = run_rmbg(input_fg)
    results, extra_images = process(input_fg, input_bg, prompt, image_width, image_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source)
    results = [(x * 255.0).clip(0, 255).astype(np.uint8) for x in results]
    return results + extra_images


@torch.inference_mode()
def process_normal(input_fg, input_bg, prompt, image_width, image_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source):
    input_fg, matting = run_rmbg(input_fg, sigma=16)

    print('left ...')
    left = process(input_fg, input_bg, prompt, image_width, image_height, 1, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, BGSource.LEFT.value)[0][0]

    print('right ...')
    right = process(input_fg, input_bg, prompt, image_width, image_height, 1, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, BGSource.RIGHT.value)[0][0]

    print('bottom ...')
    bottom = process(input_fg, input_bg, prompt, image_width, image_height, 1, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, BGSource.BOTTOM.value)[0][0]

    print('top ...')
    top = process(input_fg, input_bg, prompt, image_width, image_height, 1, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, BGSource.TOP.value)[0][0]

    inner_results = [left * 2.0 - 1.0, right * 2.0 - 1.0, bottom * 2.0 - 1.0, top * 2.0 - 1.0]

    ambient = (left + right + bottom + top) / 4.0
    h, w, _ = ambient.shape
    matting = resize_and_center_crop((matting[..., 0] * 255.0).clip(0, 255).astype(np.uint8), w, h).astype(np.float32)[..., None] / 255.0

    def safa_divide(a, b):
        e = 1e-5
        return ((a + e) / (b + e)) - 1.0

    left = safa_divide(left, ambient)
    right = safa_divide(right, ambient)
    bottom = safa_divide(bottom, ambient)
    top = safa_divide(top, ambient)

    u = (right - left) * 0.5
    v = (top - bottom) * 0.5

    sigma = 10.0
    u = np.mean(u, axis=2)
    v = np.mean(v, axis=2)
    h = (1.0 - u ** 2.0 - v ** 2.0).clip(0, 1e5) ** (0.5 * sigma)
    z = np.zeros_like(h)

    normal = np.stack([u, v, h], axis=2)
    normal /= np.sum(normal ** 2.0, axis=2, keepdims=True) ** 0.5
    normal = normal * matting + np.stack([z, z, 1 - z], axis=2) * (1 - matting)

    results = [normal, left, right, bottom, top] + inner_results
    results = [(x * 127.5 + 127.5).clip(0, 255).astype(np.uint8) for x in results]
    return results


quick_prompts = [
    'beautiful woman',
    'handsome man',
    'beautiful woman, cinematic lighting',
    'handsome man, cinematic lighting',
    'beautiful woman, natural lighting',
    'handsome man, natural lighting',
    'beautiful woman, neo punk lighting, cyberpunk',
    'handsome man, neo punk lighting, cyberpunk',
]
quick_prompts = [[x] for x in quick_prompts]


class BGSource(Enum):
    UPLOAD = "Use Background Image"
    UPLOAD_FLIP = "Use Flipped Background Image"
    LEFT = "Left Light"
    RIGHT = "Right Light"
    TOP = "Top Light"
    BOTTOM = "Bottom Light"
    GREY = "Ambient"


block = gr.Blocks().queue()
with block:
    with gr.Row():
        gr.Markdown("## IC-Light (Relighting with Foreground and Background Condition)")
    with gr.Row():
        with gr.Column():
            with gr.Row():
                input_fg = gr.Image(source='upload', type="numpy", label="Foreground", height=480)
                input_bg = gr.Image(source='upload', type="numpy", label="Background", height=480)
            prompt = gr.Textbox(label="Prompt")
            bg_source = gr.Radio(choices=[e.value for e in BGSource],
                                 value=BGSource.UPLOAD.value,
                                 label="Background Source", type='value')

            example_prompts = gr.Dataset(samples=quick_prompts, label='Prompt Quick List', components=[prompt])
            bg_gallery = gr.Gallery(height=450, object_fit='contain', label='Background Quick List', value=db_examples.bg_samples, columns=5, allow_preview=False)
            relight_button = gr.Button(value="Relight")

            with gr.Group():
                with gr.Row():
                    num_samples = gr.Slider(label="Images", minimum=1, maximum=12, value=1, step=1)
                    seed = gr.Number(label="Seed", value=12345, precision=0)
                with gr.Row():
                    image_width = gr.Slider(label="Image Width", minimum=256, maximum=1024, value=512, step=64)
                    image_height = gr.Slider(label="Image Height", minimum=256, maximum=1024, value=640, step=64)

            with gr.Accordion("Advanced options", open=False):
                steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=20, step=1)
                cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=7.0, step=0.01)
                highres_scale = gr.Slider(label="Highres Scale", minimum=1.0, maximum=3.0, value=1.5, step=0.01)
                highres_denoise = gr.Slider(label="Highres Denoise", minimum=0.1, maximum=0.9, value=0.5, step=0.01)
                a_prompt = gr.Textbox(label="Added Prompt", value='best quality')
                n_prompt = gr.Textbox(label="Negative Prompt",
                                      value='lowres, bad anatomy, bad hands, cropped, worst quality')
                normal_button = gr.Button(value="Compute Normal (4x Slower)")
        with gr.Column():
            result_gallery = gr.Gallery(height=832, object_fit='contain', label='Outputs')
    with gr.Row():
        dummy_image_for_outputs = gr.Image(visible=False, label='Result')
        gr.Examples(
            fn=lambda *args: [args[-1]],
            examples=db_examples.background_conditioned_examples,
            inputs=[
                input_fg, input_bg, prompt, bg_source, image_width, image_height, seed, dummy_image_for_outputs
            ],
            outputs=[result_gallery],
            run_on_click=True, examples_per_page=1024
        )
    ips = [input_fg, input_bg, prompt, image_width, image_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source]
    relight_button.click(fn=process_relight, inputs=ips, outputs=[result_gallery])
    normal_button.click(fn=process_normal, inputs=ips, outputs=[result_gallery])
    example_prompts.click(lambda x: x[0], inputs=example_prompts, outputs=prompt, show_progress=False, queue=False)

    def bg_gallery_selected(gal, evt: gr.SelectData):
        return gal[evt.index]['name']

    bg_gallery.select(bg_gallery_selected, inputs=bg_gallery, outputs=input_bg)


block.launch(server_name='0.0.0.0')