Commit
•
7ac68e0
1
Parent(s):
b1b84dc
Update app.py
Browse files
app.py
CHANGED
@@ -1,5 +1,6 @@
|
|
1 |
import gradio as gr
|
2 |
import torch
|
|
|
3 |
from diffusers.utils import load_image
|
4 |
from controlnet_flux import FluxControlNetModel
|
5 |
from transformer_flux import FluxTransformer2DModel
|
@@ -14,10 +15,12 @@ controlnet = FluxControlNetModel.from_pretrained("alimama-creative/FLUX.1-dev-Co
|
|
14 |
transformer = FluxTransformer2DModel.from_pretrained(
|
15 |
"black-forest-labs/FLUX.1-dev", subfolder='transformer', torch_dtype=torch.bfloat16
|
16 |
)
|
|
|
17 |
pipe = FluxControlNetInpaintingPipeline.from_pretrained(
|
18 |
"black-forest-labs/FLUX.1-dev",
|
19 |
controlnet=controlnet,
|
20 |
transformer=transformer,
|
|
|
21 |
torch_dtype=torch.bfloat16
|
22 |
)
|
23 |
repo_name = "ByteDance/Hyper-SD"
|
@@ -26,7 +29,7 @@ pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name))
|
|
26 |
pipe.fuse_lora(lora_scale=0.125)
|
27 |
pipe.transformer.to(torch.bfloat16)
|
28 |
pipe.controlnet.to(torch.bfloat16)
|
29 |
-
|
30 |
def can_expand(source_width, source_height, target_width, target_height, alignment):
|
31 |
if alignment in ("Left", "Right") and source_width >= target_width:
|
32 |
return False
|
@@ -133,7 +136,6 @@ def prepare_image_and_mask(image, width, height, overlap_percentage, resize_opti
|
|
133 |
|
134 |
@spaces.GPU
|
135 |
def inpaint(image, width, height, overlap_percentage, num_inference_steps, resize_option, custom_resize_percentage, prompt_input, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom, progress=gr.Progress(track_tqdm=True)):
|
136 |
-
pipe.enable_model_cpu_offload()
|
137 |
|
138 |
background, mask = prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom)
|
139 |
|
@@ -158,9 +160,12 @@ def inpaint(image, width, height, overlap_percentage, num_inference_steps, resiz
|
|
158 |
controlnet_conditioning_scale=0.9,
|
159 |
guidance_scale=3.5,
|
160 |
negative_prompt="",
|
161 |
-
true_guidance_scale=3.5
|
|
|
162 |
).images[0]
|
163 |
-
|
|
|
|
|
164 |
result = result.convert("RGBA")
|
165 |
cnet_image.paste(result, (0, 0), mask)
|
166 |
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
+
from diffusers import AutoencoderKL
|
4 |
from diffusers.utils import load_image
|
5 |
from controlnet_flux import FluxControlNetModel
|
6 |
from transformer_flux import FluxTransformer2DModel
|
|
|
15 |
transformer = FluxTransformer2DModel.from_pretrained(
|
16 |
"black-forest-labs/FLUX.1-dev", subfolder='transformer', torch_dtype=torch.bfloat16
|
17 |
)
|
18 |
+
vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae").to("cuda")
|
19 |
pipe = FluxControlNetInpaintingPipeline.from_pretrained(
|
20 |
"black-forest-labs/FLUX.1-dev",
|
21 |
controlnet=controlnet,
|
22 |
transformer=transformer,
|
23 |
+
vae=vae,
|
24 |
torch_dtype=torch.bfloat16
|
25 |
)
|
26 |
repo_name = "ByteDance/Hyper-SD"
|
|
|
29 |
pipe.fuse_lora(lora_scale=0.125)
|
30 |
pipe.transformer.to(torch.bfloat16)
|
31 |
pipe.controlnet.to(torch.bfloat16)
|
32 |
+
pipe.to("cuda")
|
33 |
def can_expand(source_width, source_height, target_width, target_height, alignment):
|
34 |
if alignment in ("Left", "Right") and source_width >= target_width:
|
35 |
return False
|
|
|
136 |
|
137 |
@spaces.GPU
|
138 |
def inpaint(image, width, height, overlap_percentage, num_inference_steps, resize_option, custom_resize_percentage, prompt_input, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom, progress=gr.Progress(track_tqdm=True)):
|
|
|
139 |
|
140 |
background, mask = prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom)
|
141 |
|
|
|
160 |
controlnet_conditioning_scale=0.9,
|
161 |
guidance_scale=3.5,
|
162 |
negative_prompt="",
|
163 |
+
true_guidance_scale=3.5,
|
164 |
+
output_type="latent"
|
165 |
).images[0]
|
166 |
+
pipe.to("cpu")
|
167 |
+
vae.to("cuda")
|
168 |
+
result = vae.decode(latent_image).sample
|
169 |
result = result.convert("RGBA")
|
170 |
cnet_image.paste(result, (0, 0), mask)
|
171 |
|