Spaces:
Running
on
A10G
Running
on
A10G
Commit
•
c1ee41f
1
Parent(s):
86a6a3c
Update app.py
Browse files
app.py
CHANGED
@@ -6,12 +6,9 @@ import random
|
|
6 |
from io import BytesIO
|
7 |
from utils import *
|
8 |
from constants import *
|
9 |
-
# from inversion_utils import *
|
10 |
-
# from inversion_utils_dpmplusplus import *
|
11 |
-
#from modified_pipeline_semantic_stable_diffusion import SemanticStableDiffusionPipeline
|
12 |
from pipeline_semantic_stable_diffusion_img2img_solver import SemanticStableDiffusionImg2ImgPipeline_DPMSolver
|
13 |
from torch import autocast, inference_mode
|
14 |
-
from diffusers import StableDiffusionPipeline
|
15 |
from diffusers.schedulers import DDIMScheduler
|
16 |
from scheduling_dpmsolver_multistep_inject import DPMSolverMultistepSchedulerInject
|
17 |
from transformers import AutoProcessor, BlipForConditionalGeneration
|
@@ -20,17 +17,14 @@ from share_btn import community_icon_html, loading_icon_html, share_js
|
|
20 |
# load pipelines
|
21 |
sd_model_id = "runwayml/stable-diffusion-v1-5"
|
22 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
23 |
-
|
24 |
-
pipe = SemanticStableDiffusionImg2ImgPipeline_DPMSolver.from_pretrained(sd_model_id,torch_dtype=torch.float16).to(device)
|
25 |
-
# pipe.scheduler = DDIMScheduler.from_config(sd_model_id, subfolder = "scheduler")
|
26 |
pipe.scheduler = DPMSolverMultistepSchedulerInject.from_pretrained(sd_model_id, subfolder="scheduler"
|
27 |
, algorithm_type="sde-dpmsolver++", solver_order=2)
|
28 |
|
29 |
blip_processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
30 |
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base",torch_dtype=torch.float16).to(device)
|
31 |
|
32 |
-
|
33 |
-
|
34 |
## IMAGE CPATIONING ##
|
35 |
def caption_image(input_image):
|
36 |
inputs = blip_processor(images=input_image, return_tensors="pt").to(device, torch.float16)
|
@@ -40,65 +34,6 @@ def caption_image(input_image):
|
|
40 |
generated_caption = blip_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
41 |
return generated_caption, generated_caption
|
42 |
|
43 |
-
|
44 |
-
|
45 |
-
## DDPM INVERSION AND SAMPLING ##
|
46 |
-
# def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1):
|
47 |
-
|
48 |
-
# # inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf,
|
49 |
-
# # based on the code in https://github.com/inbarhub/DDPM_inversion
|
50 |
-
|
51 |
-
# # returns wt, zs, wts:
|
52 |
-
# # wt - inverted latent
|
53 |
-
# # wts - intermediate inverted latents
|
54 |
-
# # zs - noise maps
|
55 |
-
|
56 |
-
# sd_pipe.scheduler.set_timesteps(num_diffusion_steps)
|
57 |
-
|
58 |
-
# # vae encode image
|
59 |
-
# with inference_mode():
|
60 |
-
# w0 = (sd_pipe.vae.encode(x0).latent_dist.mode() * 0.18215)
|
61 |
-
|
62 |
-
# # find Zs and wts - forward process
|
63 |
-
# wt, zs, wts = inversion_forward_process(sd_pipe, w0, etas=eta, prompt=prompt_src, cfg_scale=cfg_scale_src, prog_bar=True, num_inference_steps=num_diffusion_steps)
|
64 |
-
# return zs, wts
|
65 |
-
|
66 |
-
|
67 |
-
# def sample(zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):
|
68 |
-
|
69 |
-
# # reverse process (via Zs and wT)
|
70 |
-
# w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=True, zs=zs[skip:])
|
71 |
-
|
72 |
-
# # vae decode image
|
73 |
-
# with inference_mode():
|
74 |
-
# x0_dec = sd_pipe.vae.decode(1 / 0.18215 * w0).sample
|
75 |
-
# if x0_dec.dim()<4:
|
76 |
-
# x0_dec = x0_dec[None,:,:,:]
|
77 |
-
# img = image_grid(x0_dec)
|
78 |
-
# return img
|
79 |
-
|
80 |
-
# def reconstruct(tar_prompt,
|
81 |
-
# image_caption,
|
82 |
-
# tar_cfg_scale,
|
83 |
-
# skip,
|
84 |
-
# wts, zs,
|
85 |
-
# do_reconstruction,
|
86 |
-
# reconstruction,
|
87 |
-
# reconstruct_button
|
88 |
-
# ):
|
89 |
-
|
90 |
-
# if reconstruct_button == "Hide Reconstruction":
|
91 |
-
# return reconstruction.value, reconstruction, ddpm_edited_image.update(visible=False), do_reconstruction, "Show Reconstruction"
|
92 |
-
|
93 |
-
# else:
|
94 |
-
# if do_reconstruction:
|
95 |
-
# if image_caption.lower() == tar_prompt.lower(): # if image caption was not changed, run actual reconstruction
|
96 |
-
# tar_prompt = ""
|
97 |
-
# reconstruction_img = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
|
98 |
-
# reconstruction = gr.State(value=reconstruction_img)
|
99 |
-
# do_reconstruction = False
|
100 |
-
# return reconstruction.value, reconstruction, ddpm_edited_image.update(visible=True), do_reconstruction, "Hide Reconstruction"
|
101 |
-
|
102 |
def sample(zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):
|
103 |
|
104 |
latnets = wts.value[-1].expand(1, -1, -1, -1)
|
@@ -112,8 +47,6 @@ def sample(zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):
|
|
112 |
zs=zs.value).images[0]
|
113 |
return img
|
114 |
|
115 |
-
|
116 |
-
|
117 |
def reconstruct(tar_prompt,
|
118 |
image_caption,
|
119 |
tar_cfg_scale,
|
@@ -903,5 +836,4 @@ with gr.Blocks(css="style.css") as demo:
|
|
903 |
|
904 |
|
905 |
demo.queue()
|
906 |
-
demo.launch()
|
907 |
-
# demo.launch(share=True)
|
|
|
6 |
from io import BytesIO
|
7 |
from utils import *
|
8 |
from constants import *
|
|
|
|
|
|
|
9 |
from pipeline_semantic_stable_diffusion_img2img_solver import SemanticStableDiffusionImg2ImgPipeline_DPMSolver
|
10 |
from torch import autocast, inference_mode
|
11 |
+
from diffusers import StableDiffusionPipeline, AutoencoderKL
|
12 |
from diffusers.schedulers import DDIMScheduler
|
13 |
from scheduling_dpmsolver_multistep_inject import DPMSolverMultistepSchedulerInject
|
14 |
from transformers import AutoProcessor, BlipForConditionalGeneration
|
|
|
17 |
# load pipelines
|
18 |
sd_model_id = "runwayml/stable-diffusion-v1-5"
|
19 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
20 |
+
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16)
|
21 |
+
pipe = SemanticStableDiffusionImg2ImgPipeline_DPMSolver.from_pretrained(sd_model_id,vae=vae,torch_dtype=torch.float16).to(device)
|
|
|
22 |
pipe.scheduler = DPMSolverMultistepSchedulerInject.from_pretrained(sd_model_id, subfolder="scheduler"
|
23 |
, algorithm_type="sde-dpmsolver++", solver_order=2)
|
24 |
|
25 |
blip_processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
26 |
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base",torch_dtype=torch.float16).to(device)
|
27 |
|
|
|
|
|
28 |
## IMAGE CPATIONING ##
|
29 |
def caption_image(input_image):
|
30 |
inputs = blip_processor(images=input_image, return_tensors="pt").to(device, torch.float16)
|
|
|
34 |
generated_caption = blip_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
35 |
return generated_caption, generated_caption
|
36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
def sample(zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):
|
38 |
|
39 |
latnets = wts.value[-1].expand(1, -1, -1, -1)
|
|
|
47 |
zs=zs.value).images[0]
|
48 |
return img
|
49 |
|
|
|
|
|
50 |
def reconstruct(tar_prompt,
|
51 |
image_caption,
|
52 |
tar_cfg_scale,
|
|
|
836 |
|
837 |
|
838 |
demo.queue()
|
839 |
+
demo.launch()
|
|