app.py

import os
import random
import numpy as np
import gradio as gr
import base64
from io import BytesIO
import PIL.Image
from typing import Tuple
from novita_client import NovitaClient, V3TaskResponseStatus
from time import time

from style_template import styles



# global variable
MAX_SEED = np.iinfo(np.int32).max
STYLE_NAMES = list(styles.keys())
DEFAULT_STYLE_NAME = 'Watercolor'
DEFAULT_MODEL_NAME = 'sdxlUnstableDiffusers_v8HEAVENSWRATH_133813'
enable_lcm_arg = False

# Path to InstantID models
face_adapter = f'./checkpoints/ip-adapter.bin'
controlnet_path = f'./checkpoints/ControlNetModel'

# controlnet-pose/canny/depth
controlnet_pose_model = 'thibaud/controlnet-openpose-sdxl-1.0'
controlnet_canny_model = 'diffusers/controlnet-canny-sdxl-1.0'
controlnet_depth_model = 'diffusers/controlnet-depth-sdxl-1.0-small'

SDXL_MODELS = [
	"albedobaseXL_v04_130099",
	"altxl_v60_146691",
	"animeArtDiffusionXL_alpha2_91872",
	"animeArtDiffusionXL_alpha3_93120",
	"animeIllustDiffusion_v04_117809",
	"animagineXLV3_v30_231047",
	"breakdomainxl_V05g_124265",
	"brixlAMustInYour_v40Dagobah_145992",
	"cinemaxAlphaSDXLCinema_alpha1_107473",
	"cineroXLPhotomatic_v12aPHENO_137703",
	"clearhungAnimeXL_v10_117716",
	"copaxTimelessxlSDXL1_colorfulV2_100729",
	"counterfeitxl__98184",
	"counterfeitxl_v10_108721",
	"crystalClearXL_ccxl_97637",
	"dreamshaperXL09Alpha_alpha2Xl10_91562",
	"dynavisionXLAllInOneStylized_alpha036FP16Bakedvae_99980",
	"dynavisionXLAllInOneStylized_beta0411Bakedvae_109970",
	"dynavisionXLAllInOneStylized_release0534bakedvae_129001",
	"fenrisxl_145_134980",
	"foddaxlPhotorealism_v45_122788",
	"formulaxl_v10_104889",
	"juggernautXL_version2_113240",
	"juggernautXL_version5_126522",
	"juggernautXL_v8Rundiffusion_227002",
	"kohakuXL_alpha7_111843",
	"LahMysteriousSDXL_v40_122478",
	"leosamsHelloworldSDXLModel_helloworldSDXL10_112178",
	"leosamsHelloworldSDXL_helloworldSDXL50_268813",
	"mbbxlUltimate_v10RC_94686",
	"moefusionSDXL_v10_114018",
	"nightvisionXLPhotorealisticPortrait_beta0681Bakedvae_108833",
	"nightvisionXLPhotorealisticPortrait_beta0702Bakedvae_113098",
	"nightvisionXLPhotorealisticPortrait_release0770Bakedvae_154525",
	"novaPrimeXL_v10_107899",
	"pixelwave_v10_117722",
	"protovisionXLHighFidelity3D_beta0520Bakedvae_106612",
	"protovisionXLHighFidelity3D_release0620Bakedvae_131308",
	"protovisionXLHighFidelity3D_release0630Bakedvae_154359",
	"protovisionXLHighFidelity3D_releaseV660Bakedvae_207131",
	"realismEngineSDXL_v05b_131513",
	"realismEngineSDXL_v10_136287",
	"realisticStockPhoto_v10_115618",
	"RealitiesEdgeXL_4_122673",
	"realvisxlV20_v20Bakedvae_129156",
	"riotDiffusionXL_v20_139293",
	"roxl_v10_109354",
	"sd_xl_base_0.9",
	"sd_xl_base_1.0",
	"sdxlNijiSpecial_sdxlNijiSE_115638",
	"sdxlNijiV3_sdxlNijiV3_104571",
	"sdxlNijiV51_sdxlNijiV51_112807",
	"sd_xl_refiner_1.0",
	"sdxlUnstableDiffusers_v8HEAVENSWRATH_133813",
	"sdXL_v10Refiner_91495",
	"sdxlYamersAnimeUltra_yamersAnimeV3_121537",
	"shikianimexl_v10_93788",
	#"stable-diffusion-xl-base-1.0",
	"theTalosProject_v10_117893",
	"thinkdiffusionxl_v10_145931",
	"voidnoisecorexl_r1486_150780",
	"wlopArienwlopstylexl_v10_101973",
	"wlopSTYLEXL_v2_126171",
	"xl13AsmodeusSFWNSFW_v22BakedVAE_111954",
	"xxmix9realisticsdxl_v10_123235",
	"zavychromaxl_b2_103298",
	"zavychromaxl_v21_129006",
]

CONTROLNET_DICT = dict(
	pose={
		'model_name': 'controlnet-openpose-sdxl-1.0',
		'strength': 1,
		'preprocessor': 'openpose',
	},
	depth={
		'model_name': 'controlnet-depth-sdxl-1.0',
		'strength': 1,
		'preprocessor': 'depth',
	},
	canny={
		'model_name': 'controlnet-canny-sdxl-1.0',
		'strength': 1,
		'preprocessor': 'canny',
	},
)

last_check = 0



def get_novita_client (novita_key):
	client = NovitaClient(novita_key, os.getenv('NOVITA_API_URI', None))
	return client


get_local_storage = '''
	function () {
		globalThis.setStorage = (key, value)=>{
			localStorage.setItem(key, JSON.stringify(value))
		}
		globalThis.getStorage = (key, value)=>{
			return JSON.parse(localStorage.getItem(key))
		}

		const novita_key = getStorage("novita_key")
		return [novita_key];
	}
'''


def toggle_lcm_ui (value):
	if value:
		return (
			gr.update(minimum=0, maximum=100, step=1, value=5),
			gr.update(minimum=0.1, maximum=20.0, step=0.1, value=1.5),
		)
	else:
		return (
			gr.update(minimum=5, maximum=100, step=1, value=30),
			gr.update(minimum=0.1, maximum=20.0, step=0.1, value=5),
		)


def randomize_seed_fn (seed: int, randomize_seed: bool) -> int:
	if randomize_seed:
		seed = random.randint(0, MAX_SEED)
	return seed


def remove_tips ():
	return gr.update(visible=False)


def apply_style (style_name: str, positive: str, negative: str = "") -> Tuple[str, str]:
	p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
	return p.replace("{prompt}", positive), n + " " + negative


def get_example ():
	case = [
		[
			'./examples/yann-lecun_resize.jpg',
			None,
			'a man',
			'Spring Festival',
			'(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green',
		],
		[
			'./examples/musk_resize.jpeg',
			'./examples/poses/pose2.jpg',
			'a man flying in the sky in Mars',
			'Mars',
			'(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green',
		],
		[
			'./examples/sam_resize.png',
			'./examples/poses/pose4.jpg',
			'a man doing a silly pose wearing a suite',
			'Jungle',
			'(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, gree',
		],
		[
			'./examples/schmidhuber_resize.png',
			'./examples/poses/pose3.jpg',
			'a man sit on a chair',
			'Neon',
			'(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green',
		],
		[
			'./examples/kaifu_resize.png',
			'./examples/poses/pose.jpg',
			'a man',
			'Vibrant Color',
			'(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green',
		],
	]
	return case


def load_example (face_file, pose_file, prompt, style, negative_prompt):
	name = os.path.basename(face_file).split('_')[0]
	image = PIL.Image.open(open(f'./examples/generated/{name}.jpg', 'rb'))

	return image, gr.update(visible=True)


upload_depot = {}

def upload_assets_with_cache (client, paths):
	global upload_depot

	pending_paths = [path for path in paths if not path in upload_depot]
	if pending_paths:
		print('uploading images:', pending_paths)
	for key, value in zip(pending_paths, client.upload_assets(pending_paths)):
		upload_depot[key] = value

	return [upload_depot[path] for path in paths]



def generate_image (
	novita_key1,
	model_name,
	face_image_path,
	pose_image_path,
	prompt,
	negative_prompt,
	style_name,
	num_steps,
	identitynet_strength_ratio,
	adapter_strength_ratio,
	pose_strength,
	canny_strength,
	depth_strength,
	controlnet_selection,
	guidance_scale,
	seed,
	scheduler,
	#enable_LCM,
	#enhance_face_region,
	progress=gr.Progress(track_tqdm=True),
):
	if face_image_path is None:
		raise gr.Error(f'Cannot find any input face image! Please refer to step 1️⃣')

	#print('novita_key:', novita_key1)
	#print('face_image_path:', face_image_path)
	if not novita_key1:
		raise gr.Error(f'Please input your Novita Key!')
	try:
		client = get_novita_client(novita_key1)
		prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt)
		prompt = prompt[:1024]
		#print('prompt:', prompt)
		#print('negative_prompt:', negative_prompt)
		#print('seed:', seed)
		#print('identitynet_strength_ratio:', identitynet_strength_ratio)
		#print('adapter_strength_ratio:', adapter_strength_ratio)
		#print('scheduler:', scheduler)
		#print('guidance_scale:', guidance_scale)
		#print('num_steps:', num_steps)

		ref_image_path = pose_image_path if pose_image_path else face_image_path
		ref_image = PIL.Image.open(ref_image_path)

		width, height = ref_image.size
		large_edge = max(width, height)
		if large_edge < 1024:
			scaling = 1024 / large_edge
			width = int(width * scaling)
			height = int(height * scaling)

		CONTROLNET_DICT['pose']['strength'] = pose_strength
		CONTROLNET_DICT['canny']['strength'] = canny_strength
		CONTROLNET_DICT['depth']['strength'] = depth_strength

		face_image_uploaded, ref_image_uploaded = upload_assets_with_cache(client, [face_image_path, ref_image_path])

		res = client._post('/v3/async/instant-id', {
			'extra': {
				'response_image_type': 'jpeg',
			},
			'model_name': f'{model_name}.safetensors',
			'face_image_assets_ids': [face_image_uploaded],
			'ref_image_assets_ids': [ref_image_uploaded],
			'prompt': prompt,
			'negative_prompt': negative_prompt,
			'controlnet': {
				'units': [CONTROLNET_DICT[name] for name in controlnet_selection if name in CONTROLNET_DICT],
			},
			'image_num': 1,
			'steps': num_steps,
			'seed': seed,
			'guidance_scale': guidance_scale,
			'sampler_name': scheduler,
			'id_strength': identitynet_strength_ratio,
			'adapter_strength': adapter_strength_ratio,
			'width': width,
			'height': height,
		})

		print('task_id:', res['task_id'])
		def progress (x):
			global last_check
			t = time()
			if t > last_check + 5:
				last_check = t
				print('progress:', t, x.task.status)
		final_res = client.wait_for_task_v3(res['task_id'], callback=progress)
		if final_res is None or final_res.task.status == V3TaskResponseStatus.TASK_STATUS_FAILED:
			raise RuntimeError(f'Novita task failed: {final_res and final_res.task.status}')
		print('status:', final_res.task.status)
		print('returned images:', final_res.images)

		final_res.download_images()
	except Exception as e:
		raise gr.Error(f'Error: {e}')

	#print('final_res:', final_res)
	#print('final_res.images_encoded:', final_res.images_encoded)

	image = PIL.Image.open(BytesIO(base64.b64decode(final_res.images_encoded[0])))

	return image, gr.update(visible=True)


# Description
title = r'''
<h1 align="center">InstantID: Zero-shot Identity-Preserving Generation in Seconds (via Novita)</h1>
'''

description = r'''
<a href='https://github.com/InstantID/InstantID' target="_blank"><b>InstantID</b></a> demo via <a href="https://novita.ai/" target="_blank"><b>Novita API</b></a>.<br>

How to use:<br>
0. Input your <a href="https://novita.ai/dashboard/key" target="_blank"><b>Novita API Key</b></a>.
1. Upload an image with a face. For images with multiple faces, we will only detect the largest face. Ensure the face is not too small and is clearly visible without significant obstructions or blurring.
2. (Optional) You can upload another image as a reference for the face pose. If you don't, we will use the first detected face image to extract facial landmarks. If you use a cropped face at step 1, it is recommended to upload it to define a new face pose.
3. (Optional) You can select multiple ControlNet models to control the generation process. The default is to use the IdentityNet only. The ControlNet models include pose skeleton, canny, and depth. You can adjust the strength of each ControlNet model to control the generation process.
4. Enter a text prompt, as done in normal text-to-image models.
5. Click the <b>Submit</b> button to begin customization.
6. Share your customized photo with your friends and enjoy! 😊'''

article = r'''
---
'''

tips = r'''
### Usage tips of InstantID
1. If you're not satisfied with the similarity, try increasing the weight of "IdentityNet Strength" and "Adapter Strength."	
2. If you feel that the saturation is too high, first decrease the Adapter strength. If it remains too high, then decrease the IdentityNet strength.
3. If you find that text control is not as expected, decrease Adapter strength.
4. If you find that realistic style is not good enough, go for our Github repo and use a more realistic base model.
'''

css = '''
.gradio-container {width: 85% !important}
'''
with gr.Blocks(css=css) as demo:
	# description
	gr.Markdown(title)
	gr.Markdown(description)

	with gr.Row():
		with gr.Column(scale=1):
			novita_key = gr.Textbox(value='', label='Novita.AI API KEY', placeholder='novita.ai api key', type='password')
		with gr.Column(scale=1):
			user_balance = gr.Textbox(label='User Balance', value='0.0')

	with gr.Row():
		with gr.Column():
			with gr.Row(equal_height=True):
				# upload face image
				face_file = gr.Image(
					label='Upload a photo of your face', type='filepath'
				)
				# optional: upload a reference pose image
				pose_file = gr.Image(
					label='Upload a reference pose image (Optional)',
					type='filepath',
				)

			# prompt
			prompt = gr.Textbox(
				label='Prompt',
				info='Give simple prompt is enough to achieve good face fidelity',
				placeholder='A photo of a person',
				value='',
			)

			submit = gr.Button('Submit', variant='primary')
			#enable_LCM = gr.Checkbox(
			#	label='Enable Fast Inference with LCM', value=enable_lcm_arg,
			#	info='LCM speeds up the inference step, the trade-off is the quality of the generated image. It performs better with portrait face images rather than distant faces',
			#)

			model_name = gr.Dropdown(
				label='Base model',
				choices=SDXL_MODELS,
				value=DEFAULT_MODEL_NAME,
			)

			style = gr.Dropdown(
				label='Style template',
				choices=STYLE_NAMES,
				value=DEFAULT_STYLE_NAME,
			)

			# strength
			identitynet_strength_ratio = gr.Slider(
				label='IdentityNet strength (for fidelity)',
				minimum=0,
				maximum=1.5,
				step=0.05,
				value=0.80,
			)
			adapter_strength_ratio = gr.Slider(
				label='Image adapter strength (for detail)',
				minimum=0,
				maximum=1.5,
				step=0.05,
				value=0.80,
			)
			with gr.Accordion('Controlnet'):
				controlnet_selection = gr.CheckboxGroup(
					['pose', 'canny', 'depth'], label='Controlnet', value=['pose'],
					info='Use pose for skeleton inference, canny for edge detection, and depth for depth map estimation. You can try all three to control the generation process'
				)
				pose_strength = gr.Slider(
					label='Pose strength',
					minimum=0,
					maximum=1.5,
					step=0.05,
					value=0.40,
				)
				canny_strength = gr.Slider(
					label='Canny strength',
					minimum=0,
					maximum=1.5,
					step=0.05,
					value=0.40,
				)
				depth_strength = gr.Slider(
					label='Depth strength',
					minimum=0,
					maximum=1.5,
					step=0.05,
					value=0.40,
				)
			with gr.Accordion(open=False, label='Advanced Options'):
				negative_prompt = gr.Textbox(
					label='Negative Prompt',
					placeholder='low quality',
					value='(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green',
				)
				num_steps = gr.Slider(
					label='Number of sample steps',
					minimum=1,
					maximum=100,
					step=1,
					value=5 if enable_lcm_arg else 30,
				)
				guidance_scale = gr.Slider(
					label='Guidance scale',
					minimum=1.,
					maximum=30.0,
					step=0.1,
					value=0.0 if enable_lcm_arg else 5.0,
				)
				seed = gr.Slider(
					label='Seed',
					minimum=0,
					maximum=MAX_SEED,
					step=1,
					value=42,
				)
				schedulers = [
					'Euler',
					'Euler a',
					'Heun',
					'DPM++ SDE',
					'DPM++ SDE Karras',
					'DPM2',
					'DPM2 Karras',
					'DPM2 a',
					'DPM2 a Karras',
				]
				scheduler = gr.Dropdown(
					label='Schedulers',
					choices=schedulers,
					value='Euler',
				)
				randomize_seed = gr.Checkbox(label='Randomize seed', value=True)
				#enhance_face_region = gr.Checkbox(label='Enhance non-face region', value=True)

		with gr.Column(scale=1):
			gallery = gr.Image(label='Generated Images')
			usage_tips = gr.Markdown(
				label='InstantID Usage Tips', value=tips, visible=False
			)

		submit.click(
			fn=remove_tips,
			outputs=usage_tips,
		).then(
			fn=randomize_seed_fn,
			inputs=[seed, randomize_seed],
			outputs=seed,
			queue=False,
			api_name=False,
		).then(
			fn=generate_image,
			inputs=[
				novita_key,
				model_name,
				face_file,
				pose_file,
				prompt,
				negative_prompt,
				style,
				num_steps,
				identitynet_strength_ratio,
				adapter_strength_ratio,
				pose_strength,
				canny_strength,
				depth_strength,
				controlnet_selection,
				guidance_scale,
				seed,
				scheduler,
				#enable_LCM,
				#enhance_face_region,
			],
			outputs=[gallery, usage_tips],
		)

		#enable_LCM.input(
		#	fn=toggle_lcm_ui,
		#	inputs=[enable_LCM],
		#	outputs=[num_steps, guidance_scale],
		#	queue=False,
		#)

	gr.Examples(
		examples=get_example(),
		inputs=[face_file, pose_file, prompt, style, negative_prompt],
		fn=load_example,
		outputs=[gallery, usage_tips],
		cache_examples=True,
	)

	gr.Markdown(article)

	def onload(novita_key):
		if novita_key is None or novita_key == '':
			return novita_key, f'$ UNKNOWN', gr.update(visible=False)
		try:
			user_info_json = get_novita_client(novita_key).user_info()
		except Exception as e:
			return novita_key, f'$ UNKNOWN'

		return novita_key, f'$ {user_info_json.credit_balance / 100 / 100:.2f}'

	novita_key.change(onload, inputs=novita_key, outputs=[novita_key, user_balance], js='v=>{ setStorage("novita_key", v); return [v]; }')

	demo.load(
		inputs=[novita_key],
		outputs=[novita_key, user_balance],
		fn=onload,
		js=get_local_storage,
	)


demo.queue(api_open=False)
demo.launch()