Spaces:
Runtime error
Runtime error
add app.py
Browse files- app.py +804 -0
- requirements.txt +9 -0
app.py
ADDED
@@ -0,0 +1,804 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Graphit
|
3 |
+
Copyright (c) 2023-present NAVER Corp.
|
4 |
+
Apache-2.0
|
5 |
+
"""
|
6 |
+
import os
|
7 |
+
import numpy as np
|
8 |
+
import base64
|
9 |
+
import requests
|
10 |
+
from io import BytesIO
|
11 |
+
import json
|
12 |
+
import time
|
13 |
+
import math
|
14 |
+
import argparse
|
15 |
+
|
16 |
+
import torch
|
17 |
+
import torch.nn.functional as F
|
18 |
+
import gradio as gr
|
19 |
+
|
20 |
+
import types
|
21 |
+
from typing import Union, List, Optional, Callable
|
22 |
+
import diffusers
|
23 |
+
import torch
|
24 |
+
from diffusers import AutoencoderKL, UNet2DConditionModel
|
25 |
+
from diffusers import StableDiffusionInstructPix2PixPipeline, EulerAncestralDiscreteScheduler
|
26 |
+
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
|
27 |
+
from diffusers.models import AutoencoderKL
|
28 |
+
from transformers import CLIPTextModel
|
29 |
+
|
30 |
+
import datasets
|
31 |
+
|
32 |
+
from torchvision import transforms
|
33 |
+
from torchvision.transforms.functional import to_pil_image, pil_to_tensor
|
34 |
+
|
35 |
+
import PIL
|
36 |
+
from PIL import Image, ImageOps
|
37 |
+
|
38 |
+
import compodiff
|
39 |
+
from transformers import DPTFeatureExtractor, DPTForDepthEstimation
|
40 |
+
from transparent_background import Remover
|
41 |
+
from huggingface_hub import hf_hub_url, cached_download
|
42 |
+
from RealESRGAN import RealESRGAN
|
43 |
+
import einops
|
44 |
+
import cv2
|
45 |
+
from skimage import segmentation, color, graph
|
46 |
+
import random
|
47 |
+
|
48 |
+
|
49 |
+
def preprocess(image, mode):
|
50 |
+
image = np.array(image)[None, :].astype(np.float32) / 255.0
|
51 |
+
image = image
|
52 |
+
image = image.transpose(0, 3, 1, 2)
|
53 |
+
image = 2.0 * image - 1.0
|
54 |
+
if mode == 'scr2i':
|
55 |
+
image[image > 0.0] = 0.0
|
56 |
+
image = torch.from_numpy(image)
|
57 |
+
return image
|
58 |
+
|
59 |
+
|
60 |
+
class GraphitPipeline(StableDiffusionInstructPix2PixPipeline):
|
61 |
+
'''
|
62 |
+
override:
|
63 |
+
/opt/conda/lib/python3.8/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py
|
64 |
+
'''
|
65 |
+
def prepare_image_latents(
|
66 |
+
self, image, mask, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None
|
67 |
+
):
|
68 |
+
if not isinstance(image, (torch.Tensor, Image.Image, list)):
|
69 |
+
raise ValueError(
|
70 |
+
f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
|
71 |
+
)
|
72 |
+
|
73 |
+
image = image.to(device=device, dtype=dtype)
|
74 |
+
mask = mask.to(device=device, dtype=dtype)
|
75 |
+
|
76 |
+
batch_size = batch_size * num_images_per_prompt
|
77 |
+
if isinstance(generator, list) and len(generator) != batch_size:
|
78 |
+
raise ValueError(
|
79 |
+
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
80 |
+
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
81 |
+
)
|
82 |
+
|
83 |
+
if isinstance(generator, list):
|
84 |
+
image_latents = [self.vae.encode(image[i : i + 1]).latent_dist.mode() for i in range(batch_size)]
|
85 |
+
image_latents = torch.cat(image_latents, dim=0)
|
86 |
+
else:
|
87 |
+
image_latents = self.vae.encode(image).latent_dist.mode()
|
88 |
+
|
89 |
+
mask = torch.nn.functional.interpolate(
|
90 |
+
mask, #.unsqueeze(0).unsqueeze(0),
|
91 |
+
size=(image_latents.shape[-2], image_latents.shape[-1]),
|
92 |
+
mode='bicubic',
|
93 |
+
align_corners=False,
|
94 |
+
)
|
95 |
+
|
96 |
+
if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
|
97 |
+
# expand image_latents for batch_size
|
98 |
+
deprecation_message = (
|
99 |
+
f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial"
|
100 |
+
" images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
|
101 |
+
" that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
|
102 |
+
" your script to pass as many initial images as text prompts to suppress this warning."
|
103 |
+
)
|
104 |
+
#deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
|
105 |
+
additional_image_per_prompt = batch_size // image_latents.shape[0]
|
106 |
+
image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
|
107 |
+
mask = torch.cat([mask] * additional_image_per_prompt, dim=0)
|
108 |
+
elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
|
109 |
+
raise ValueError(
|
110 |
+
f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
|
111 |
+
)
|
112 |
+
else:
|
113 |
+
image_latents = torch.cat([image_latents], dim=0)
|
114 |
+
image_latents *= 0.18215
|
115 |
+
if do_classifier_free_guidance:
|
116 |
+
uncond_image_latents = torch.zeros_like(image_latents)
|
117 |
+
image_latents = torch.cat([image_latents, image_latents], dim=0)
|
118 |
+
mask = torch.cat([mask, mask], dim=0)
|
119 |
+
image_latents = torch.cat([image_latents, mask], dim=1)
|
120 |
+
|
121 |
+
return image_latents
|
122 |
+
|
123 |
+
@torch.no_grad()
|
124 |
+
def __call__(
|
125 |
+
self,
|
126 |
+
prompt: Union[str, List[str]] = None,
|
127 |
+
image: Union[torch.FloatTensor, PIL.Image.Image] = None,
|
128 |
+
mask: Union[torch.FloatTensor, PIL.Image.Image] = None,
|
129 |
+
depth_map: Union[torch.FloatTensor, PIL.Image.Image] = None,
|
130 |
+
num_inference_steps: int = 100,
|
131 |
+
guidance_scale: float = 3.5,
|
132 |
+
use_depth_map_as_input: bool = False,
|
133 |
+
apply_mask_to_input: bool = True,
|
134 |
+
mode: str = None,
|
135 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
136 |
+
num_images_per_prompt: Optional[int] = 1,
|
137 |
+
eta: float = 0.0,
|
138 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
139 |
+
latents: Optional[torch.FloatTensor] = None,
|
140 |
+
image_cond_embeds: Optional[torch.FloatTensor] = None,
|
141 |
+
negative_image_cond_embeds: Optional[torch.FloatTensor] = None,
|
142 |
+
output_type: Optional[str] = "pil",
|
143 |
+
return_dict: bool = True,
|
144 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
145 |
+
callback_steps: Optional[int] = 1,
|
146 |
+
):
|
147 |
+
# 0. Check inputs
|
148 |
+
self.check_inputs(prompt, callback_steps)
|
149 |
+
|
150 |
+
if image is None:
|
151 |
+
raise ValueError("`image` input cannot be undefined.")
|
152 |
+
|
153 |
+
# 1. Define call parameters
|
154 |
+
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
155 |
+
device = self._execution_device
|
156 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
157 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
158 |
+
# corresponds to doing no classifier free guidance.
|
159 |
+
do_classifier_free_guidance = True#guidance_scale >= 1.0 and image_guidance_scale >= 1.0
|
160 |
+
# check if scheduler is in sigmas space
|
161 |
+
scheduler_is_in_sigma_space = hasattr(self.scheduler, "sigmas")
|
162 |
+
|
163 |
+
# 2. Encode input prompt
|
164 |
+
cond_embeds = torch.cat([image_cond_embeds, negative_image_cond_embeds])
|
165 |
+
cond_embeds = einops.repeat(cond_embeds, 'b n d -> (b num) n d', num=num_images_per_prompt).to(torch.float16)
|
166 |
+
prompt_embeds = cond_embeds
|
167 |
+
|
168 |
+
# 3. Preprocess image
|
169 |
+
image = preprocess(image, mode)
|
170 |
+
|
171 |
+
if len(mask.shape) > 2:
|
172 |
+
edge_map = mask[:,:,1:]
|
173 |
+
edge_map = preprocess(edge_map, mode)
|
174 |
+
mask = mask[:,:,0]
|
175 |
+
else:
|
176 |
+
edge_map = None
|
177 |
+
mask = mask.unsqueeze(0).unsqueeze(0)
|
178 |
+
if torch.sum(mask).item() == 0.0 and use_depth_map_as_input:
|
179 |
+
image = depth_map
|
180 |
+
if edge_map is None:
|
181 |
+
if apply_mask_to_input:
|
182 |
+
image = image * (1 - mask)
|
183 |
+
else:
|
184 |
+
image = image * (1 - mask) + edge_map * mask
|
185 |
+
height, width = image.shape[-2:]
|
186 |
+
|
187 |
+
# 4. set timesteps
|
188 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
189 |
+
timesteps = self.scheduler.timesteps
|
190 |
+
|
191 |
+
# 5. Prepare Image latents
|
192 |
+
image_latents = self.prepare_image_latents(
|
193 |
+
image,
|
194 |
+
mask,
|
195 |
+
batch_size,
|
196 |
+
num_images_per_prompt,
|
197 |
+
prompt_embeds.dtype,
|
198 |
+
device,
|
199 |
+
do_classifier_free_guidance,
|
200 |
+
generator,
|
201 |
+
)
|
202 |
+
|
203 |
+
if mode == 't2i':
|
204 |
+
image_latents = torch.zeros_like(image_latents)
|
205 |
+
|
206 |
+
# 6. Prepare latent variables
|
207 |
+
num_channels_latents = self.vae.config.latent_channels
|
208 |
+
latents = self.prepare_latents(
|
209 |
+
batch_size * num_images_per_prompt,
|
210 |
+
num_channels_latents,
|
211 |
+
height,
|
212 |
+
width,
|
213 |
+
prompt_embeds.dtype,
|
214 |
+
device,
|
215 |
+
generator,
|
216 |
+
latents,
|
217 |
+
)
|
218 |
+
|
219 |
+
# 7. Check that shapes of latents and image match the UNet channels
|
220 |
+
num_channels_image = image_latents.shape[1]
|
221 |
+
if num_channels_latents + num_channels_image != self.unet.config.in_channels:
|
222 |
+
raise ValueError(
|
223 |
+
f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
|
224 |
+
f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
|
225 |
+
f" `num_channels_image`: {num_channels_image} "
|
226 |
+
f" = {num_channels_latents+num_channels_image}. Please verify the config of"
|
227 |
+
" `pipeline.unet` or your `image` input."
|
228 |
+
)
|
229 |
+
|
230 |
+
# 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
231 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
232 |
+
|
233 |
+
# 9. Denoising loop
|
234 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
235 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
236 |
+
for i, t in enumerate(timesteps):
|
237 |
+
# Expand the latents if we are doing classifier free guidance.
|
238 |
+
# The latents are expanded 3 times because for pix2pix the guidance\
|
239 |
+
# is applied for both the text and the input image.
|
240 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
241 |
+
|
242 |
+
# concat latents, image_latents in the channel dimension
|
243 |
+
scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
244 |
+
scaled_latent_model_input = torch.cat([scaled_latent_model_input, image_latents], dim=1)
|
245 |
+
|
246 |
+
# predict the noise residual
|
247 |
+
noise_pred = self.unet(scaled_latent_model_input, t, encoder_hidden_states=prompt_embeds).sample
|
248 |
+
|
249 |
+
# Hack:
|
250 |
+
# For karras style schedulers the model does classifer free guidance using the
|
251 |
+
# predicted_original_sample instead of the noise_pred. So we need to compute the
|
252 |
+
# predicted_original_sample here if we are using a karras style scheduler.
|
253 |
+
if scheduler_is_in_sigma_space:
|
254 |
+
step_index = (self.scheduler.timesteps == t).nonzero().item()
|
255 |
+
sigma = self.scheduler.sigmas[step_index]
|
256 |
+
noise_pred = latent_model_input - sigma * noise_pred
|
257 |
+
|
258 |
+
# perform guidance
|
259 |
+
if do_classifier_free_guidance:
|
260 |
+
noise_pred_full, noise_pred_uncond = noise_pred.chunk(2)
|
261 |
+
noise_pred = (
|
262 |
+
noise_pred_uncond
|
263 |
+
+ guidance_scale * (noise_pred_full - noise_pred_uncond)
|
264 |
+
)
|
265 |
+
|
266 |
+
# Hack:
|
267 |
+
# For karras style schedulers the model does classifer free guidance using the
|
268 |
+
# predicted_original_sample instead of the noise_pred. But the scheduler.step function
|
269 |
+
# expects the noise_pred and computes the predicted_original_sample internally. So we
|
270 |
+
# need to overwrite the noise_pred here such that the value of the computed
|
271 |
+
# predicted_original_sample is correct.
|
272 |
+
if scheduler_is_in_sigma_space:
|
273 |
+
noise_pred = (noise_pred - latents) / (-sigma)
|
274 |
+
|
275 |
+
# compute the previous noisy sample x_t -> x_t-1
|
276 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
277 |
+
|
278 |
+
# call the callback, if provided
|
279 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
280 |
+
progress_bar.update()
|
281 |
+
if callback is not None and i % callback_steps == 0:
|
282 |
+
callback(i, t, latents)
|
283 |
+
|
284 |
+
# 10. Post-processing
|
285 |
+
image = self.decode_latents(latents)
|
286 |
+
|
287 |
+
# 11. Run safety checker
|
288 |
+
image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
|
289 |
+
|
290 |
+
# 12. Convert to PIL
|
291 |
+
if output_type == "pil":
|
292 |
+
image = self.numpy_to_pil(image)
|
293 |
+
|
294 |
+
if not return_dict:
|
295 |
+
return (image, has_nsfw_concept)
|
296 |
+
|
297 |
+
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|
298 |
+
|
299 |
+
|
300 |
+
class CustomRealESRGAN(RealESRGAN):
|
301 |
+
@torch.no_grad()
|
302 |
+
@torch.cuda.amp.autocast()
|
303 |
+
def predict(self, pil_lr_image_list):
|
304 |
+
device = self.device
|
305 |
+
# batchfy
|
306 |
+
batch_lr_images = (torch.stack([pil_to_tensor(pil_lr_image) for pil_lr_image in pil_lr_image_list]).float() / 255).to(device)
|
307 |
+
batch_outputs = self.model(batch_lr_images).clamp_(0, 1)
|
308 |
+
|
309 |
+
# to pil images
|
310 |
+
return [to_pil_image(output) for output in batch_outputs]
|
311 |
+
|
312 |
+
|
313 |
+
def build_models(args):
|
314 |
+
# Load scheduler, tokenizer and models.
|
315 |
+
|
316 |
+
model_path = 'navervision/Graphit-SD'
|
317 |
+
unet = UNet2DConditionModel.from_pretrained(
|
318 |
+
model_path, torch_dtype=torch.float16,
|
319 |
+
)
|
320 |
+
|
321 |
+
vae_name = 'stabilityai/sd-vae-ft-ema'
|
322 |
+
vae = AutoencoderKL.from_pretrained(vae_name, torch_dtype=torch.float16)
|
323 |
+
|
324 |
+
model_name = 'timbrooks/instruct-pix2pix'
|
325 |
+
pipe = GraphitPipeline.from_pretrained(model_name, torch_dtype=torch.float16, safety_checker=None,
|
326 |
+
unet = unet,
|
327 |
+
vae = vae,
|
328 |
+
)
|
329 |
+
pipe = pipe.to('cuda:0')
|
330 |
+
|
331 |
+
## load CompoDiff
|
332 |
+
compodiff_model, clip_model, clip_preprocess, clip_tokenizer = compodiff.build_model()
|
333 |
+
compodiff_model, clip_model = compodiff_model.to('cuda:0'), clip_model.to('cuda:0')
|
334 |
+
|
335 |
+
## load third-party models
|
336 |
+
model_name = 'Intel/dpt-large'
|
337 |
+
depth_preprocess = DPTFeatureExtractor.from_pretrained(model_name)
|
338 |
+
depth_predictor = DPTForDepthEstimation.from_pretrained(model_name, torch_dtype=torch.float16)
|
339 |
+
depth_predictor = depth_predictor.to('cuda:0')
|
340 |
+
|
341 |
+
if not os.path.exists('./third_party/remover_fast.pth'):
|
342 |
+
model_file_url = hf_hub_url(repo_id='Geonmo/remover_fast', filename='remover_fast.pth')
|
343 |
+
cached_download(model_file_url, cache_dir='./third_party', force_filename='remover_fast.pth')
|
344 |
+
remover = Remover(fast=True, jit=False, device='cuda:0', ckpt='./third_party/remover_fast.pth')
|
345 |
+
|
346 |
+
sr_model = CustomRealESRGAN('cuda:0', scale=2)
|
347 |
+
sr_model.load_weights('./third_party/RealESRGAN_x2.pth', download=True)
|
348 |
+
|
349 |
+
dataset = datasets.load_dataset("FredZhang7/stable-diffusion-prompts-2.47M")
|
350 |
+
|
351 |
+
train = dataset["train"]
|
352 |
+
prompts = train["text"]
|
353 |
+
|
354 |
+
model_dict = {'pipe': pipe,
|
355 |
+
'compodiff': compodiff_model,
|
356 |
+
'clip_preprocess': clip_preprocess,
|
357 |
+
'clip_tokenizer': clip_tokenizer,
|
358 |
+
'clip_model': clip_model,
|
359 |
+
'depth_preprocess': depth_preprocess,
|
360 |
+
'depth_predictor': depth_predictor,
|
361 |
+
'remover': remover,
|
362 |
+
'sr_model': sr_model,
|
363 |
+
'prompt_candidates': prompts,
|
364 |
+
}
|
365 |
+
return model_dict
|
366 |
+
|
367 |
+
|
368 |
+
def predict_compodiff(image, text_input, negative_text, cfg_image_scale, cfg_text_scale, mask, random_seed):
|
369 |
+
text_token_dict = model_dict['clip_tokenizer'](text=text_input, return_tensors='pt', padding='max_length', truncation=True)
|
370 |
+
text_tokens, text_attention_mask = text_token_dict['input_ids'].to('cuda:0'), text_token_dict['attention_mask'].to('cuda:0')
|
371 |
+
|
372 |
+
negative_text_token_dict = model_dict['clip_tokenizer'](text=negative_text, return_tensors='pt', padding='max_length', truncation=True)
|
373 |
+
negative_text_tokens, negative_text_attention_mask = negative_text_token_dict['input_ids'].to('cuda:0'), text_token_dict['attention_mask'].to('cuda:0')
|
374 |
+
|
375 |
+
with torch.no_grad():
|
376 |
+
if image is None:
|
377 |
+
image_cond = torch.zeros([1,1,768]).to('cuda:0')
|
378 |
+
mask = torch.tensor(np.zeros([64, 64], dtype='float32')).to('cuda:0').unsqueeze(0)
|
379 |
+
else:
|
380 |
+
image_source = image.resize((512, 512))
|
381 |
+
image_source = model_dict['clip_preprocess'](image_source, return_tensors='pt')['pixel_values'].to('cuda:0')
|
382 |
+
mask = mask.resize((512, 512))
|
383 |
+
mask = model_dict['clip_preprocess'](mask, do_normalize=False, return_tensors='pt')['pixel_values']
|
384 |
+
mask = mask[:,:1,:,:]
|
385 |
+
mask = (mask > 0.5).float().to('cuda:0')
|
386 |
+
image_source = image_source * (1 - mask)
|
387 |
+
image_cond = model_dict['clip_model'].encode_images(image_source)
|
388 |
+
mask = transforms.Resize([64, 64])(mask)[:,0,:,:]
|
389 |
+
mask = (mask > 0.5).float()
|
390 |
+
|
391 |
+
text_cond = model_dict['clip_model'].encode_texts(text_tokens, text_attention_mask)
|
392 |
+
negative_text_cond = model_dict['clip_model'].encode_texts(negative_text_tokens, negative_text_attention_mask)
|
393 |
+
|
394 |
+
sampled_image_features = model_dict['compodiff'].sample(image_cond, text_cond, negative_text_cond, mask, timesteps=25, cond_scale=(1.0 if image is None else 1.3, cfg_text_scale), num_samples_per_batch=4, random_seed=random_seed).unsqueeze(1)
|
395 |
+
return sampled_image_features, image_cond
|
396 |
+
|
397 |
+
|
398 |
+
def generate_depth_map(image, height, width):
|
399 |
+
depth_inputs = {k: v.to('cuda:0', dtype=torch.float16) for k, v in model_dict['depth_preprocess'](images=image, return_tensors='pt').items()}
|
400 |
+
depth_map = model_dict['depth_predictor'](**depth_inputs).predicted_depth.unsqueeze(1)
|
401 |
+
depth_min = torch.amin(depth_map, dim=[1,2,3], keepdim=True)
|
402 |
+
depth_max = torch.amax(depth_map, dim=[1,2,3], keepdim=True)
|
403 |
+
depth_map = 2.0 * ((depth_map - depth_min) / (depth_max - depth_min)) - 1.0
|
404 |
+
depth_map = torch.nn.functional.interpolate(
|
405 |
+
depth_map,
|
406 |
+
size=(height, width),
|
407 |
+
mode='bicubic',
|
408 |
+
align_corners=False,
|
409 |
+
)
|
410 |
+
return depth_map
|
411 |
+
|
412 |
+
|
413 |
+
def generate_color(image, compactness=30, n_segments=100, thresh=35, blur_kernel=3, blur_std=0):
|
414 |
+
img = image # 0 ~ 255 uint8
|
415 |
+
labels = segmentation.slic(img, compactness=compactness, n_segments=n_segments)#, start_label=1)
|
416 |
+
g = graph.rag_mean_color(img, labels)
|
417 |
+
labels2 = graph.cut_threshold(labels, g, thresh=thresh)
|
418 |
+
out = color.label2rgb(labels2, img, kind='avg', bg_label=-1)
|
419 |
+
return out
|
420 |
+
|
421 |
+
|
422 |
+
@torch.no_grad()
|
423 |
+
def generate(image_source, image_reference, text_input, negative_prompt, steps, random_seed, cfg_image_scale, cfg_text_scale, cfg_image_space_scale, cfg_image_reference_mix_weight, cfg_image_source_mix_weight, mask_scale, use_edge, t2i_height, t2i_width, do_sr, mode):
|
424 |
+
text_input = text_input.lower()
|
425 |
+
if negative_prompt == '':
|
426 |
+
print('running without a negative prompt')
|
427 |
+
# prepare an input image
|
428 |
+
use_mask = False
|
429 |
+
mask = None
|
430 |
+
is_null_image_source = False
|
431 |
+
if type(image_source) == dict:
|
432 |
+
image_source, mask = image_source['image'], image_source['mask']
|
433 |
+
elif image_source is None:
|
434 |
+
image_source = Image.fromarray(np.zeros([t2i_height, t2i_width, 3]).astype('uint8'))
|
435 |
+
is_null_image_source = True
|
436 |
+
|
437 |
+
try:
|
438 |
+
image_source = ImageOps.exif_transpose(image_source)
|
439 |
+
except:
|
440 |
+
pass
|
441 |
+
|
442 |
+
width, height = image_source.size
|
443 |
+
factor = 512 / max(width, height)
|
444 |
+
factor = math.ceil(min(width, height) * factor / 64) * 64 / min(width, height)
|
445 |
+
width = int((width * factor) // 64) * 64
|
446 |
+
height = int((height * factor) // 64) * 64
|
447 |
+
|
448 |
+
image_source = org_image_source = ImageOps.fit(image_source, (width, height), method=Image.Resampling.LANCZOS)
|
449 |
+
|
450 |
+
if mask is not None:
|
451 |
+
mask_pil = mask = ImageOps.fit(mask, (width, height), method=Image.Resampling.LANCZOS)
|
452 |
+
mask = ((torch.tensor(np.array(mask.convert('L'))).float() / 255.0) > 0.5).float()
|
453 |
+
if torch.sum(mask).item() > 0.0:
|
454 |
+
print('now using mask')
|
455 |
+
use_mask = True
|
456 |
+
else:
|
457 |
+
mask = torch.zeros([height, width])
|
458 |
+
mask_pil = to_pil_image(mask)
|
459 |
+
|
460 |
+
use_depth_map_as_input = False
|
461 |
+
if mode == 's2i' or mode == 'scr2i': # sketch to image
|
462 |
+
image_source = mask
|
463 |
+
image_source = einops.repeat(image_source, 'h w -> r h w', r=3)
|
464 |
+
mask = image_source[0,:,:]
|
465 |
+
image_source = org_image_source = to_pil_image(image_source)
|
466 |
+
mask_pil = to_pil_image(mask)
|
467 |
+
mask *= mask_scale
|
468 |
+
use_mask = False
|
469 |
+
elif mode == 'cs2i':
|
470 |
+
mask = torch.tensor((np.array(image_source)[:,:,0] != 255)).float() * mask_scale
|
471 |
+
mask_pil = Image.fromarray(((np.array(image_source)[:,:,0] != 255) * 255).astype('uint8'))
|
472 |
+
use_mask = False #True
|
473 |
+
elif mode == 'd2i': # depth to image
|
474 |
+
use_depth_map_as_input = True
|
475 |
+
elif mode == 'e2i': # edge to image
|
476 |
+
image_source = einops.repeat(cv2.Canny(cv2.cvtColor(np.array(image_source)[:,:,::-1], cv2.COLOR_BGR2GRAY), threshold1=100, threshold2=200), 'h w -> h w r', r=3)
|
477 |
+
image_source = Image.fromarray(image_source) #to_pil_image(image_source)
|
478 |
+
org_image_source = image_source
|
479 |
+
elif mode == 'inped':
|
480 |
+
# mask = torch.Size([512, 512])
|
481 |
+
mask_np = (einops.repeat(mask.numpy(), 'h w -> h w r', r=1) * 255).astype('uint8')
|
482 |
+
gray = mask_np #cv2.cvtColor(mask_np, cv2.COLOR_BGR2GRAY)
|
483 |
+
_, thresh = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY)
|
484 |
+
contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
485 |
+
x, y, w, h = cv2.boundingRect(contours[0])
|
486 |
+
cv2.rectangle(mask_np, (x, y), (x+w, y+h), 255, -1)
|
487 |
+
mask_np = mask_np.astype('float32') / 255
|
488 |
+
if image_reference is not None:
|
489 |
+
edge_reference = image_reference.resize((w, h))
|
490 |
+
color_map = generate_color(np.array(edge_reference)).astype('float32')
|
491 |
+
reference_map = (model_dict['remover'].process(edge_reference, type='map') > 16).astype('float32')
|
492 |
+
edge_reference = einops.repeat(cv2.Canny(cv2.cvtColor(np.array(edge_reference)[:,:,::-1], cv2.COLOR_BGR2GRAY), threshold1=100, threshold2=200), 'h w -> h w r', r=3).astype('float32')
|
493 |
+
edge_np = np.zeros_like(np.array(image_source)).astype('float32')
|
494 |
+
if text_input != '':
|
495 |
+
edge_np[y:y+h,x:x+w] = edge_reference * reference_map
|
496 |
+
elif use_edge and mask_scale > 0.0:
|
497 |
+
print('mode: color inped with with_edge')
|
498 |
+
edge_np[y:y+h,x:x+w] = (255 - edge_reference) / 255 * color_map * reference_map + (1 - mask_scale) * edge_reference / 255 * reference_map
|
499 |
+
else:
|
500 |
+
print('mode: color inped with no_edge')
|
501 |
+
edge_np[y:y+h,x:x+w] = color_map * reference_map
|
502 |
+
mask_np = np.zeros_like(np.array(image_source)).astype('float32')
|
503 |
+
mask_np[y:y+h,x:x+w] = reference_map #edge_reference
|
504 |
+
mask_np = mask_np[:,:,:1]
|
505 |
+
else:
|
506 |
+
edge_np = einops.repeat(cv2.Canny(cv2.cvtColor(np.array(image_source)[:,:,::-1], cv2.COLOR_BGR2GRAY), threshold1=100, threshold2=200), 'h w -> h w r', r=3).astype('float32')
|
507 |
+
# concat edge to mask_np
|
508 |
+
mask = torch.tensor(np.concatenate([mask_np, edge_np], axis=-1))
|
509 |
+
mask_pil = to_pil_image(mask_np[:,:,0].astype('uint8') * 255)
|
510 |
+
#mask_pil = to_pil_image((mask_np[:,:,0] * 255).astype('uint8'))
|
511 |
+
|
512 |
+
with torch.no_grad():
|
513 |
+
# do reference first
|
514 |
+
if image_reference is not None:
|
515 |
+
image_cond_reference = ImageOps.exif_transpose(image_reference)
|
516 |
+
image_cond_reference = model_dict['clip_preprocess'](image_cond_reference, return_tensors='pt')['pixel_values'].to('cuda:0')
|
517 |
+
image_cond_reference = model_dict['clip_model'].encode_images(image_cond_reference)
|
518 |
+
else:
|
519 |
+
image_cond_reference = torch.zeros([1, 1, 768]).to(torch.float16).to('cuda:0')
|
520 |
+
|
521 |
+
# do source or knn
|
522 |
+
image_cond_source = None
|
523 |
+
if text_input != '':
|
524 |
+
if mode in ['t2i', 'd2i', 'e2i', 's2i', 'scr2i', 'cs2i']:
|
525 |
+
if mode == 'cs2i':
|
526 |
+
image_cond, image_cond_source = predict_compodiff(None, text_input, negative_prompt, cfg_image_scale, cfg_text_scale, mask=mask_pil, random_seed=random_seed)
|
527 |
+
image_cond_color_compensation, _ = predict_compodiff(image_source, text_input, negative_prompt, cfg_image_scale, cfg_text_scale, mask=mask_pil, random_seed=random_seed)
|
528 |
+
image_cond = 0.9 * image_cond + 0.1 * image_cond_color_compensation
|
529 |
+
else:
|
530 |
+
image_cond, image_cond_source = predict_compodiff(None, text_input, negative_prompt, cfg_image_scale, cfg_text_scale, mask=mask_pil, random_seed=random_seed)
|
531 |
+
else:
|
532 |
+
image_cond, image_cond_source = predict_compodiff(image_source, text_input, negative_prompt, cfg_image_scale, cfg_text_scale, mask=mask_pil, random_seed=random_seed)
|
533 |
+
image_cond = image_cond.to(torch.float16).to('cuda:0')
|
534 |
+
image_cond_source = image_cond_source.to(torch.float16).to('cuda:0')
|
535 |
+
else:
|
536 |
+
image_cond = torch.zeros([1, 1, 768]).to(torch.float16).to('cuda:0')
|
537 |
+
|
538 |
+
if image_cond_source is None and mode != 't2i':
|
539 |
+
image_cond_source = image_source.resize((512, 512))
|
540 |
+
image_cond_source = model_dict['clip_preprocess'](image_cond_source, return_tensors='pt')['pixel_values'].to('cuda:0')
|
541 |
+
image_cond_source = model_dict['clip_model'].encode_images(image_cond_source)
|
542 |
+
|
543 |
+
if cfg_image_reference_mix_weight > 0.0 and torch.sum(image_cond_reference).item() != 0.0:
|
544 |
+
if torch.sum(image_cond).item() == 0.0:
|
545 |
+
image_cond = image_cond_reference
|
546 |
+
else:
|
547 |
+
image_cond = (1.0 - cfg_image_reference_mix_weight) * image_cond + cfg_image_reference_mix_weight * image_cond_reference
|
548 |
+
|
549 |
+
if cfg_image_source_mix_weight > 0.0:
|
550 |
+
image_cond = (1.0 - cfg_image_source_mix_weight) * image_cond + cfg_image_source_mix_weight * image_cond_source
|
551 |
+
|
552 |
+
if negative_prompt != '':
|
553 |
+
negative_image_cond, _ = predict_compodiff(None, negative_prompt, '', cfg_image_scale, cfg_text_scale, mask=mask_pil, random_seed=random_seed)
|
554 |
+
negative_image_cond = negative_image_cond.to(torch.float16).to('cuda:0')
|
555 |
+
else:
|
556 |
+
negative_image_cond = torch.zeros_like(image_cond)
|
557 |
+
|
558 |
+
# negative_prompt_embeds
|
559 |
+
image_source = torch.tensor(np.array(image_source))
|
560 |
+
depth_map = einops.repeat(generate_depth_map(image_source, height, width), 'n c h w -> n (c r) h w', r=3).float().cpu()
|
561 |
+
|
562 |
+
images = model_dict['pipe'](text_input,
|
563 |
+
image=image_source,
|
564 |
+
mask=mask,
|
565 |
+
depth_map=depth_map,
|
566 |
+
num_inference_steps=int(steps),
|
567 |
+
image_cond_embeds=image_cond,
|
568 |
+
negative_image_cond_embeds=negative_image_cond,
|
569 |
+
guidance_scale=cfg_image_space_scale,
|
570 |
+
use_depth_map_as_input=use_depth_map_as_input,
|
571 |
+
apply_mask_to_input=use_mask,
|
572 |
+
mode=mode,
|
573 |
+
generator=torch.manual_seed(random_seed),
|
574 |
+
num_images_per_prompt=2).images
|
575 |
+
if do_sr:
|
576 |
+
images = model_dict['sr_model'].predict(images)
|
577 |
+
|
578 |
+
return images, [org_image_source, mask_pil, to_pil_image(0.5 * (depth_map[0] + 1.0))]
|
579 |
+
|
580 |
+
|
581 |
+
def generate_canvas(image):
|
582 |
+
return Image.fromarray((np.ones([512, 512, 3]) * 255).astype('uint8'))
|
583 |
+
|
584 |
+
|
585 |
+
def surprise_me():
|
586 |
+
return random.sample(model_dict['prompt_candidates'], k=1)[0]
|
587 |
+
|
588 |
+
|
589 |
+
if __name__ == "__main__":
|
590 |
+
parser = argparse.ArgumentParser('Demo')
|
591 |
+
parser.add_argument('--model_folder', default=None, type=str, help='path to model_folder')
|
592 |
+
|
593 |
+
args = parser.parse_args()
|
594 |
+
|
595 |
+
|
596 |
+
global model_dict
|
597 |
+
|
598 |
+
model_dict = build_models(args)
|
599 |
+
|
600 |
+
### define gradio demo
|
601 |
+
title = 'Graphit demo'
|
602 |
+
|
603 |
+
md_title = f'''# {title}
|
604 |
+
Diffusion on GPU.
|
605 |
+
'''
|
606 |
+
neg_default = 'watermark, longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality'
|
607 |
+
with gr.Blocks(title=title) as demo:
|
608 |
+
gr.Markdown(md_title)
|
609 |
+
mode_t2i = gr.Textbox(value='t2i', label='mode selection', visible=False)
|
610 |
+
mode_i2i = gr.Textbox(value='i2i', label='mode selection', visible=False)
|
611 |
+
mode_inpaint = gr.Textbox(value='inpaint', label='mode selection', visible=False)
|
612 |
+
mode_s2i = gr.Textbox(value='s2i', label='mode selection', visible=False)
|
613 |
+
mode_scr2i = gr.Textbox(value='scr2i', label='mode selection', visible=False)
|
614 |
+
mode_d2i = gr.Textbox(value='d2i', label='mode selection', visible=False)
|
615 |
+
mode_e2i = gr.Textbox(value='e2i', label='mode selection', visible=False)
|
616 |
+
mode_inped = gr.Textbox(value='inped', label='mode selection', visible=False)
|
617 |
+
mode_cs2i = gr.Textbox(value='cs2i', label='mode selection', visible=False)
|
618 |
+
mask_scale_default = gr.Number(value=1.0, label='mask scale', visible=False)
|
619 |
+
use_edge_default = gr.Checkbox(value=True, label='use color map with edge map', visible=False)
|
620 |
+
height_default = gr.Number(value=512, precision=0, label='height', visible=False)
|
621 |
+
width_default = gr.Number(value=512, precision=0, label='width', visible=False)
|
622 |
+
with gr.Row():
|
623 |
+
with gr.Column():
|
624 |
+
with gr.Tabs():
|
625 |
+
'''
|
626 |
+
image to image
|
627 |
+
inpainting
|
628 |
+
depth to image
|
629 |
+
saliency map to image
|
630 |
+
'''
|
631 |
+
with gr.TabItem("Text to Image"):
|
632 |
+
image_source_t2i = gr.Image(type='pil', label='Source image', visible=False)
|
633 |
+
with gr.Row():
|
634 |
+
steps_input_t2i = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
635 |
+
random_seed_t2i = gr.Number(value=12345, precision=0, label='Seed')
|
636 |
+
with gr.Accordion('Advanced options', open=False):
|
637 |
+
with gr.Row():
|
638 |
+
cfg_image_scale_t2i = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
639 |
+
cfg_image_space_scale_t2i = gr.Number(value=7.5, label='attn image space scale')
|
640 |
+
cfg_text_scale_t2i = gr.Number(value=7.5, label='attn text scale')
|
641 |
+
negative_text_input_t2i = gr.Textbox(value=neg_default, label='Negative text')
|
642 |
+
with gr.Row():
|
643 |
+
cfg_image_source_mix_weight_t2i = gr.Number(value=0.0, label='weight for mixing source image (0.0~1.0)', visible=False)
|
644 |
+
cfg_image_reference_mix_weight_t2i = gr.Number(value=0.65, label='weight for mixing reference image (0.0~1.0)')
|
645 |
+
with gr.Row():
|
646 |
+
height_t2i = gr.Number(value=512, precision=0, label='height (~512)')
|
647 |
+
width_t2i = gr.Number(value=512, precision=0, label='width (~512)')
|
648 |
+
submit_button_t2i = gr.Button('Generate images')
|
649 |
+
with gr.TabItem("Image to Image"):
|
650 |
+
image_source_i2i = gr.Image(type='pil', label='Source image')
|
651 |
+
with gr.Row():
|
652 |
+
steps_input_i2i = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
653 |
+
random_seed_i2i = gr.Number(value=12345, precision=0, label='Seed')
|
654 |
+
with gr.Accordion('Advanced options', open=False):
|
655 |
+
with gr.Row():
|
656 |
+
cfg_image_scale_i2i = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
657 |
+
cfg_image_space_scale_i2i = gr.Number(value=7.5, label='attn image space scale')
|
658 |
+
cfg_text_scale_i2i = gr.Number(value=7.5, label='attn text scale')
|
659 |
+
negative_text_input_i2i = gr.Textbox(value=neg_default, label='Negative text')
|
660 |
+
with gr.Row():
|
661 |
+
cfg_image_source_mix_weight_i2i = gr.Number(value=0.05, label='weight for mixing source image (0.0~1.0)')
|
662 |
+
cfg_image_reference_mix_weight_i2i = gr.Number(value=0.65, label='weight for mixing reference image (0.0~1.0)')
|
663 |
+
submit_button_i2i = gr.Button('Generate images')
|
664 |
+
with gr.TabItem("Depth to Image"):
|
665 |
+
image_source_d2i = gr.Image(type='pil', label='Source image')
|
666 |
+
with gr.Row():
|
667 |
+
steps_input_d2i = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
668 |
+
random_seed_d2i = gr.Number(value=12345, precision=0, label='Seed')
|
669 |
+
with gr.Accordion('Advanced options', open=False):
|
670 |
+
with gr.Row():
|
671 |
+
cfg_image_scale_d2i = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
672 |
+
cfg_image_space_scale_d2i = gr.Number(value=7.5, label='attn image space scale')
|
673 |
+
cfg_text_scale_d2i = gr.Number(value=7.5, label='attn text scale')
|
674 |
+
negative_text_input_d2i = gr.Textbox(value=neg_default, label='Negative text')
|
675 |
+
with gr.Row():
|
676 |
+
cfg_image_source_mix_weight_d2i = gr.Number(value=0.0, label='weight for mixing source image (0.0~1.0)', visible=False)
|
677 |
+
cfg_image_reference_mix_weight_d2i = gr.Number(value=1.0, label='weight for mixing reference image (0.0~1.0)')
|
678 |
+
submit_button_d2i = gr.Button('Generate images')
|
679 |
+
with gr.TabItem("Edge to Image"):
|
680 |
+
image_source_e2i = gr.Image(type='pil', label='Source image')
|
681 |
+
with gr.Row():
|
682 |
+
steps_input_e2i = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
683 |
+
random_seed_e2i = gr.Number(value=12345, precision=0, label='Seed')
|
684 |
+
with gr.Accordion('Advanced options', open=False):
|
685 |
+
with gr.Row():
|
686 |
+
cfg_image_scale_e2i = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
687 |
+
cfg_image_space_scale_e2i = gr.Number(value=7.5, label='attn image space scale')
|
688 |
+
cfg_text_scale_e2i = gr.Number(value=7.5, label='attn text scale')
|
689 |
+
negative_text_input_e2i = gr.Textbox(value=neg_default, label='Negative text')
|
690 |
+
with gr.Row():
|
691 |
+
cfg_image_source_mix_weight_e2i = gr.Number(value=0.0, label='weight for mixing source image (0.0~1.0)', visible=False)
|
692 |
+
cfg_image_reference_mix_weight_e2i = gr.Number(value=1.0, label='weight for mixing reference image (0.0~1.0)')
|
693 |
+
submit_button_e2i = gr.Button('Generate images')
|
694 |
+
with gr.TabItem("Inpaint"):
|
695 |
+
image_source_inp = gr.Image(type='pil', label='Source image', tool='sketch')
|
696 |
+
with gr.Row():
|
697 |
+
steps_input_inp = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
698 |
+
random_seed_inp = gr.Number(value=12345, precision=0, label='Seed')
|
699 |
+
with gr.Accordion('Advanced options', open=False):
|
700 |
+
with gr.Row():
|
701 |
+
cfg_image_scale_inp = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
702 |
+
cfg_image_space_scale_inp = gr.Number(value=7.5, label='attn image space scale')
|
703 |
+
cfg_text_scale_inp = gr.Number(value=7.5, label='attn text scale')
|
704 |
+
negative_text_input_inp = gr.Textbox(value='', label='Negative text')
|
705 |
+
with gr.Row():
|
706 |
+
cfg_image_source_mix_weight_inp = gr.Number(value=0.0, label='weight for mixing source image (0.0~1.0)', visible=False)
|
707 |
+
cfg_image_reference_mix_weight_inp = gr.Number(value=0.65, label='weight for mixing reference image (0.0~1.0)')
|
708 |
+
submit_button_inp = gr.Button('Generate images')
|
709 |
+
with gr.TabItem("Blending"):
|
710 |
+
image_source_inped = gr.Image(type='pil', label='Source image', tool='sketch')
|
711 |
+
with gr.Row():
|
712 |
+
steps_input_inped = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
713 |
+
random_seed_inped = gr.Number(value=12345, precision=0, label='Seed')
|
714 |
+
with gr.Accordion('Advanced options', open=False):
|
715 |
+
with gr.Row():
|
716 |
+
cfg_image_scale_inped = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
717 |
+
cfg_image_space_scale_inped = gr.Number(value=7.5, label='attn image space scale')
|
718 |
+
cfg_text_scale_inped = gr.Number(value=7.5, label='attn text scale')
|
719 |
+
negative_text_input_inped = gr.Textbox(value=neg_default, label='Negative text')
|
720 |
+
with gr.Row():
|
721 |
+
cfg_image_source_mix_weight_inped = gr.Number(value=0.0, label='weight for mixing source image (0.0~1.0)', visible=False)
|
722 |
+
cfg_image_reference_mix_weight_inped = gr.Number(value=0.35, label='weight for mixing reference image (0.0~1.0)')
|
723 |
+
with gr.Row():
|
724 |
+
mask_scale_inped = gr.Number(value=1.0, label='edge scale')
|
725 |
+
use_edge_inped = gr.Checkbox(value=False, label='use a color map with an edge map')
|
726 |
+
submit_button_inped = gr.Button('Generate images')
|
727 |
+
with gr.TabItem("Sketch (Rough) to Image"):
|
728 |
+
with gr.Column():
|
729 |
+
image_source_s2i = gr.Image(type='pil', label='Source image', tool='sketch', brush_radius=100).style(height=256, width=256)
|
730 |
+
build_canvas_s2i = gr.Button('Build canvas')
|
731 |
+
with gr.Row():
|
732 |
+
steps_input_s2i = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
733 |
+
random_seed_s2i = gr.Number(value=12345, precision=0, label='Seed')
|
734 |
+
with gr.Accordion('Advanced options', open=False):
|
735 |
+
with gr.Row():
|
736 |
+
cfg_image_scale_s2i = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
737 |
+
cfg_image_space_scale_s2i = gr.Number(value=7.5, label='attn image space scale')
|
738 |
+
cfg_text_scale_s2i = gr.Number(value=7.5, label='attn text scale')
|
739 |
+
negative_text_input_s2i = gr.Textbox(value=neg_default, label='Negative text')
|
740 |
+
with gr.Row():
|
741 |
+
cfg_image_source_mix_weight_s2i = gr.Number(value=0.0, label='weight for mixing source image (0.0~1.0)', visible=False)
|
742 |
+
cfg_image_reference_mix_weight_s2i = gr.Number(value=0.65, label='weight for mixing reference image (0.0~1.0)')
|
743 |
+
mask_scale_s2i = gr.Number(value=0.5, label='sketch weight (0.0~1.0)')
|
744 |
+
submit_button_s2i = gr.Button('Generate images')
|
745 |
+
with gr.TabItem("Sketch (Detail) to Image"):
|
746 |
+
with gr.Column():
|
747 |
+
image_source_scr2i = gr.Image(type='pil', label='Source image', tool='sketch', brush_radius=10).style(height=256, width=256)
|
748 |
+
build_canvas_scr2i = gr.Button('Build canvas')
|
749 |
+
with gr.Row():
|
750 |
+
steps_input_scr2i = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
751 |
+
random_seed_scr2i = gr.Number(value=12345, precision=0, label='Seed')
|
752 |
+
with gr.Accordion('Advanced options', open=False):
|
753 |
+
with gr.Row():
|
754 |
+
cfg_image_scale_scr2i = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
755 |
+
cfg_image_space_scale_scr2i = gr.Number(value=7.5, label='attn image space scale')
|
756 |
+
cfg_text_scale_scr2i = gr.Number(value=7.5, label='attn text scale')
|
757 |
+
negative_text_input_scr2i = gr.Textbox(value=neg_default, label='Negative text')
|
758 |
+
with gr.Row():
|
759 |
+
cfg_image_source_mix_weight_scr2i = gr.Number(value=0.0, label='weight for mixing source image (0.0~1.0)', visible=False)
|
760 |
+
cfg_image_reference_mix_weight_scr2i = gr.Number(value=0.65, label='weight for mixing reference image (0.0~1.0)')
|
761 |
+
mask_scale_scr2i = gr.Number(value=0.5, label='sketch weight (0.0~1.0)')
|
762 |
+
submit_button_scr2i = gr.Button('Generate images')
|
763 |
+
with gr.TabItem("Color Sketch to Image"):
|
764 |
+
with gr.Column():
|
765 |
+
image_source_cs2i = gr.Image(type='pil', source='canvas', label='Source image', tool='color-sketch').style(height=256, width=256)
|
766 |
+
#build_canvas_cs2i = gr.Button('Build canvas')
|
767 |
+
with gr.Row():
|
768 |
+
steps_input_cs2i = gr.Radio(['5', '10', '25', '50'], value='25', label='denoising steps')
|
769 |
+
random_seed_cs2i = gr.Number(value=12345, precision=0, label='Seed')
|
770 |
+
with gr.Accordion('Advanced options', open=False):
|
771 |
+
with gr.Row():
|
772 |
+
cfg_image_scale_cs2i = gr.Number(value=1.1, label='attn source image scale', visible=False)
|
773 |
+
cfg_image_space_scale_cs2i = gr.Number(value=7.5, label='attn image space scale')
|
774 |
+
cfg_text_scale_cs2i = gr.Number(value=7.5, label='attn text scale')
|
775 |
+
negative_text_input_cs2i = gr.Textbox(value=neg_default, label='Negative text')
|
776 |
+
with gr.Row():
|
777 |
+
cfg_image_source_mix_weight_cs2i = gr.Number(value=0.0, label='weight for mixing source image (0.0~1.0)', visible=False)
|
778 |
+
cfg_image_reference_mix_weight_cs2i = gr.Number(value=0.65, label='weight for mixing reference image (0.0~1.0)')
|
779 |
+
mask_scale_cs2i = gr.Number(value=0.5, label='sketch weight (0.0~1.0)')
|
780 |
+
submit_button_cs2i = gr.Button('Generate images')
|
781 |
+
text_input = gr.Textbox(value='', label='Input text')
|
782 |
+
submit_surprise_me = gr.Button('Surprise me')
|
783 |
+
#swap_button = gr.Button('Swap source with reference', visible=False)
|
784 |
+
with gr.Column():
|
785 |
+
with gr.Row():
|
786 |
+
do_sr = gr.Checkbox(value=False, label='Super-resolution')
|
787 |
+
image_reference = gr.Image(type='pil', label='Reference image')
|
788 |
+
gallery_outputs = gr.Gallery(label='Generated outputs').style(grid=[2], height='auto')
|
789 |
+
gallery_inputs = gr.Gallery(label='Processed inputs').style(grid=[2], height='auto')
|
790 |
+
|
791 |
+
submit_button_t2i.click(generate, inputs=[image_source_t2i, image_reference, text_input, negative_text_input_t2i, steps_input_t2i, random_seed_t2i, cfg_image_scale_t2i, cfg_text_scale_t2i, cfg_image_space_scale_t2i, cfg_image_reference_mix_weight_t2i, cfg_image_source_mix_weight_t2i, mask_scale_default, use_edge_default, height_t2i, width_t2i, do_sr, mode_t2i], outputs=[gallery_outputs, gallery_inputs])
|
792 |
+
submit_button_i2i.click(generate, inputs=[image_source_i2i, image_reference, text_input, negative_text_input_i2i, steps_input_i2i, random_seed_i2i, cfg_image_scale_i2i, cfg_text_scale_i2i, cfg_image_space_scale_i2i, cfg_image_reference_mix_weight_i2i, cfg_image_source_mix_weight_i2i, mask_scale_default, use_edge_default, height_default, width_default, do_sr, mode_i2i], outputs=[gallery_outputs, gallery_inputs])
|
793 |
+
submit_button_d2i.click(generate, inputs=[image_source_d2i, image_reference, text_input, negative_text_input_d2i, steps_input_d2i, random_seed_d2i, cfg_image_scale_d2i, cfg_text_scale_d2i, cfg_image_space_scale_d2i, cfg_image_reference_mix_weight_d2i, cfg_image_source_mix_weight_d2i, mask_scale_default, use_edge_default, height_default, width_default, do_sr, mode_d2i], outputs=[gallery_outputs, gallery_inputs])
|
794 |
+
submit_button_e2i.click(generate, inputs=[image_source_e2i, image_reference, text_input, negative_text_input_e2i, steps_input_e2i, random_seed_e2i, cfg_image_scale_e2i, cfg_text_scale_e2i, cfg_image_space_scale_e2i, cfg_image_reference_mix_weight_e2i, cfg_image_source_mix_weight_e2i, mask_scale_default, use_edge_default, height_default, width_default, do_sr, mode_e2i], outputs=[gallery_outputs, gallery_inputs])
|
795 |
+
submit_button_inp.click(generate, inputs=[image_source_inp, image_reference, text_input, negative_text_input_inp, steps_input_inp, random_seed_inp, cfg_image_scale_inp, cfg_text_scale_inp, cfg_image_space_scale_inp, cfg_image_reference_mix_weight_inp, cfg_image_source_mix_weight_inp, mask_scale_default, use_edge_default, height_default, width_default, do_sr, mode_inpaint], outputs=[gallery_outputs, gallery_inputs])
|
796 |
+
submit_button_inped.click(generate, inputs=[image_source_inped, image_reference, text_input, negative_text_input_inped, steps_input_inped, random_seed_inped, cfg_image_scale_inped, cfg_text_scale_inped, cfg_image_space_scale_inped, cfg_image_reference_mix_weight_inped, cfg_image_source_mix_weight_inped, mask_scale_inped, use_edge_inped, height_default, width_default, do_sr, mode_inped], outputs=[gallery_outputs, gallery_inputs])
|
797 |
+
submit_button_s2i.click(generate, inputs=[image_source_s2i, image_reference, text_input, negative_text_input_s2i, steps_input_s2i, random_seed_s2i, cfg_image_scale_s2i, cfg_text_scale_s2i, cfg_image_space_scale_s2i, cfg_image_reference_mix_weight_s2i, cfg_image_source_mix_weight_s2i, mask_scale_s2i, use_edge_default, height_default, width_default, do_sr, mode_s2i], outputs=[gallery_outputs, gallery_inputs])
|
798 |
+
submit_button_scr2i.click(generate, inputs=[image_source_scr2i, image_reference, text_input, negative_text_input_scr2i, steps_input_scr2i, random_seed_scr2i, cfg_image_scale_scr2i, cfg_text_scale_scr2i, cfg_image_space_scale_scr2i, cfg_image_reference_mix_weight_scr2i, cfg_image_source_mix_weight_scr2i, mask_scale_scr2i, use_edge_default, height_default, width_default, do_sr, mode_scr2i], outputs=[gallery_outputs, gallery_inputs])
|
799 |
+
submit_button_cs2i.click(generate, inputs=[image_source_cs2i, image_reference, text_input, negative_text_input_cs2i, steps_input_cs2i, random_seed_cs2i, cfg_image_scale_cs2i, cfg_text_scale_cs2i, cfg_image_space_scale_cs2i, cfg_image_reference_mix_weight_cs2i, cfg_image_source_mix_weight_cs2i, mask_scale_cs2i, use_edge_default, height_default, width_default, do_sr, mode_cs2i], outputs=[gallery_outputs, gallery_inputs])
|
800 |
+
build_canvas_s2i.click(generate_canvas, inputs=[image_source_s2i], outputs=[image_source_s2i])
|
801 |
+
build_canvas_scr2i.click(generate_canvas, inputs=[image_source_scr2i], outputs=[image_source_scr2i])
|
802 |
+
submit_surprise_me.click(surprise_me, outputs=[text_input])
|
803 |
+
demo.queue()
|
804 |
+
demo.launch()
|
requirements.txt
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch>=1.13.0
|
2 |
+
torchvision>=0.9
|
3 |
+
transformers
|
4 |
+
diffusers
|
5 |
+
huggingface_hub
|
6 |
+
git+https://github.com/navervision/CompoDiff.git
|
7 |
+
transparent-background
|
8 |
+
git+https://github.com/sberbank-ai/Real-ESRGAN.git
|
9 |
+
gradio
|