Update app.py
Browse files
app.py
CHANGED
@@ -12,7 +12,7 @@ from preprocess.openpose.run_openpose import OpenPose
|
|
12 |
|
13 |
import gradio as gr
|
14 |
|
15 |
-
# Download checkpoints
|
16 |
snapshot_download(repo_id="franciszzj/Leffa", local_dir="./ckpts")
|
17 |
|
18 |
|
@@ -34,64 +34,66 @@ class LeffaPredictor:
|
|
34 |
body_model_path="./ckpts/openpose/body_pose_model.pth",
|
35 |
)
|
36 |
|
37 |
-
|
|
|
38 |
pretrained_model_name_or_path="./ckpts/stable-diffusion-inpainting",
|
39 |
pretrained_model="./ckpts/virtual_tryon.pth",
|
40 |
dtype="float16",
|
41 |
)
|
42 |
-
self.
|
43 |
|
44 |
-
|
|
|
45 |
pretrained_model_name_or_path="./ckpts/stable-diffusion-inpainting",
|
46 |
pretrained_model="./ckpts/virtual_tryon_dc.pth",
|
47 |
dtype="float16",
|
48 |
)
|
49 |
-
self.
|
50 |
|
51 |
-
|
|
|
52 |
pretrained_model_name_or_path="./ckpts/stable-diffusion-xl-1.0-inpainting-0.1",
|
53 |
pretrained_model="./ckpts/pose_transfer.pth",
|
54 |
dtype="float16",
|
55 |
)
|
56 |
-
self.
|
57 |
-
|
58 |
-
def
|
59 |
-
|
60 |
-
|
61 |
-
assert control_type in ["virtual_tryon", "pose_transfer"]
|
62 |
-
src = Image.open(src_image_path)
|
63 |
-
ref = Image.open(ref_image_path)
|
64 |
src = resize_and_center(src, 768, 1024)
|
65 |
ref = resize_and_center(ref, 768, 1024)
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
|
|
|
|
|
|
78 |
else:
|
79 |
-
mask =
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
iuv = self.densepose_predictor.predict_iuv(arr)
|
88 |
-
seg = np.repeat(iuv[:, :, 0:1], 3, axis=-1)
|
89 |
-
densepose = Image.fromarray(seg)
|
90 |
else:
|
91 |
-
iuv = self.densepose_predictor.predict_iuv(
|
92 |
-
|
|
|
|
|
93 |
|
94 |
-
#
|
95 |
data = {
|
96 |
"src_image": [src],
|
97 |
"ref_image": [ref],
|
@@ -99,100 +101,123 @@ class LeffaPredictor:
|
|
99 |
"densepose": [densepose],
|
100 |
}
|
101 |
data = LeffaTransform()(data)
|
102 |
-
|
103 |
-
if control_type == "virtual_tryon":
|
104 |
-
inf = self.vt_inference_hd if vt_model_type == "viton_hd" else self.vt_inference_dc
|
105 |
-
else:
|
106 |
-
inf = self.pt_inference
|
107 |
-
|
108 |
out = inf(
|
109 |
data,
|
110 |
-
ref_acceleration=
|
111 |
-
num_inference_steps=
|
112 |
-
guidance_scale=scale,
|
113 |
-
seed=seed,
|
114 |
-
repaint=
|
115 |
)
|
116 |
-
|
117 |
-
return np.array(
|
118 |
-
|
119 |
-
def
|
120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
|
122 |
-
|
123 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
|
125 |
|
126 |
if __name__ == "__main__":
|
127 |
lp = LeffaPredictor()
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
gr.
|
138 |
-
gr.
|
139 |
-
|
140 |
-
"[🤖 Code](https://github.com/franciszzj/Leffa) • "
|
141 |
-
"[🤗 Model](https://huggingface.co/franciszzj/Leffa)"
|
142 |
-
)
|
143 |
|
144 |
-
|
|
|
|
|
145 |
with gr.Row():
|
146 |
with gr.Column():
|
147 |
-
vt_src = gr.Image(type="filepath", label="Person
|
148 |
-
gr.Examples(
|
|
|
149 |
with gr.Column():
|
150 |
-
vt_ref = gr.Image(type="filepath", label="Garment
|
151 |
-
gr.Examples(
|
|
|
152 |
with gr.Column():
|
153 |
-
vt_out
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
outputs=[vt_out],
|
173 |
-
)
|
174 |
|
175 |
with gr.Tab("Pose Transfer"):
|
176 |
with gr.Row():
|
177 |
with gr.Column():
|
178 |
-
|
179 |
-
gr.Examples(
|
|
|
180 |
with gr.Column():
|
181 |
-
|
182 |
-
gr.Examples(
|
|
|
183 |
with gr.Column():
|
184 |
-
pt_out
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
import gradio as gr
|
14 |
|
15 |
+
# Download checkpoints once at startup
|
16 |
snapshot_download(repo_id="franciszzj/Leffa", local_dir="./ckpts")
|
17 |
|
18 |
|
|
|
34 |
body_model_path="./ckpts/openpose/body_pose_model.pth",
|
35 |
)
|
36 |
|
37 |
+
# Virtual try‑on HD
|
38 |
+
vt_hd = LeffaModel(
|
39 |
pretrained_model_name_or_path="./ckpts/stable-diffusion-inpainting",
|
40 |
pretrained_model="./ckpts/virtual_tryon.pth",
|
41 |
dtype="float16",
|
42 |
)
|
43 |
+
self.vt_hd_inf = LeffaInference(model=vt_hd)
|
44 |
|
45 |
+
# Virtual try‑on DressCode
|
46 |
+
vt_dc = LeffaModel(
|
47 |
pretrained_model_name_or_path="./ckpts/stable-diffusion-inpainting",
|
48 |
pretrained_model="./ckpts/virtual_tryon_dc.pth",
|
49 |
dtype="float16",
|
50 |
)
|
51 |
+
self.vt_dc_inf = LeffaInference(model=vt_dc)
|
52 |
|
53 |
+
# Pose transfer
|
54 |
+
pt = LeffaModel(
|
55 |
pretrained_model_name_or_path="./ckpts/stable-diffusion-xl-1.0-inpainting-0.1",
|
56 |
pretrained_model="./ckpts/pose_transfer.pth",
|
57 |
dtype="float16",
|
58 |
)
|
59 |
+
self.pt_inf = LeffaInference(model=pt)
|
60 |
+
|
61 |
+
def _prepare(self, src_path, ref_path):
|
62 |
+
src = Image.open(src_path)
|
63 |
+
ref = Image.open(ref_path)
|
|
|
|
|
|
|
64 |
src = resize_and_center(src, 768, 1024)
|
65 |
ref = resize_and_center(ref, 768, 1024)
|
66 |
+
return src, ref
|
67 |
+
|
68 |
+
def predict_virtual_tryon(
|
69 |
+
self, src_path, ref_path,
|
70 |
+
accelerate_ref, steps, scale, seed,
|
71 |
+
model_type, garment_type, repaint
|
72 |
+
):
|
73 |
+
src, ref = self._prepare(src_path, ref_path)
|
74 |
+
src_arr = np.array(src.convert("RGB"))
|
75 |
+
|
76 |
+
# 1) parsing + keypoints → agnostic mask
|
77 |
+
parse, _ = self.parsing(src.resize((384, 512)))
|
78 |
+
kpts = self.openpose(src.resize((384, 512)))
|
79 |
+
if model_type == "viton_hd":
|
80 |
+
mask = get_agnostic_mask_hd(parse, kpts, garment_type)
|
81 |
else:
|
82 |
+
mask = get_agnostic_mask_dc(parse, kpts, garment_type)
|
83 |
+
mask = mask.resize((768, 1024))
|
84 |
+
|
85 |
+
# 2) DensePose → seg or IUV
|
86 |
+
if model_type == "viton_hd":
|
87 |
+
seg = self.densepose_predictor.predict_seg(src_arr)[:, :, ::-1]
|
88 |
+
densepose = Image.fromarray(seg)
|
89 |
+
inf = self.vt_hd_inf
|
|
|
|
|
|
|
90 |
else:
|
91 |
+
iuv = self.densepose_predictor.predict_iuv(src_arr)
|
92 |
+
seg = np.concatenate([iuv[:, :, :1]] * 3, axis=-1)
|
93 |
+
densepose = Image.fromarray(seg)
|
94 |
+
inf = self.vt_dc_inf
|
95 |
|
96 |
+
# 3) run Leffa
|
97 |
data = {
|
98 |
"src_image": [src],
|
99 |
"ref_image": [ref],
|
|
|
101 |
"densepose": [densepose],
|
102 |
}
|
103 |
data = LeffaTransform()(data)
|
|
|
|
|
|
|
|
|
|
|
|
|
104 |
out = inf(
|
105 |
data,
|
106 |
+
ref_acceleration=accelerate_ref,
|
107 |
+
num_inference_steps=int(steps),
|
108 |
+
guidance_scale=float(scale),
|
109 |
+
seed=int(seed),
|
110 |
+
repaint=repaint,
|
111 |
)
|
112 |
+
gen = out["generated_image"][0]
|
113 |
+
return np.array(gen), np.array(mask), np.array(densepose)
|
114 |
+
|
115 |
+
def predict_pose_transfer(
|
116 |
+
self, src_path, ref_path,
|
117 |
+
accelerate_ref, steps, scale, seed
|
118 |
+
):
|
119 |
+
src, ref = self._prepare(src_path, ref_path)
|
120 |
+
src_arr = np.array(src)
|
121 |
+
mask = Image.fromarray(np.ones_like(src_arr) * 255)
|
122 |
+
iuv = self.densepose_predictor.predict_iuv(src_arr)[:, :, ::-1]
|
123 |
+
densepose = Image.fromarray(iuv)
|
124 |
|
125 |
+
data = {
|
126 |
+
"src_image": [src],
|
127 |
+
"ref_image": [ref],
|
128 |
+
"mask": [mask],
|
129 |
+
"densepose": [densepose],
|
130 |
+
}
|
131 |
+
data = LeffaTransform()(data)
|
132 |
+
out = self.pt_inf(
|
133 |
+
data,
|
134 |
+
ref_acceleration=accelerate_ref,
|
135 |
+
num_inference_steps=int(steps),
|
136 |
+
guidance_scale=float(scale),
|
137 |
+
seed=int(seed),
|
138 |
+
)
|
139 |
+
gen = out["generated_image"][0]
|
140 |
+
return np.array(gen), np.array(mask), np.array(densepose)
|
141 |
|
142 |
|
143 |
if __name__ == "__main__":
|
144 |
lp = LeffaPredictor()
|
145 |
+
examples = "./ckpts/examples"
|
146 |
+
person1 = list_dir(f"{examples}/person1")
|
147 |
+
person2 = list_dir(f"{examples}/person2")
|
148 |
+
garments = list_dir(f"{examples}/garment")
|
149 |
+
|
150 |
+
title = "## Leffa: Controllable Person Image Generation"
|
151 |
+
note = "Note: Virtual Try‑On uses VITON‑HD/DressCode; Pose Transfer uses DeepFashion."
|
152 |
+
|
153 |
+
with gr.Blocks(theme=gr.themes.Default(
|
154 |
+
primary_hue=gr.themes.colors.pink,
|
155 |
+
secondary_hue=gr.themes.colors.red
|
156 |
+
)).queue() as demo:
|
|
|
|
|
|
|
157 |
|
158 |
+
gr.Markdown(title)
|
159 |
+
|
160 |
+
with gr.Tab("Virtual Try‑On"):
|
161 |
with gr.Row():
|
162 |
with gr.Column():
|
163 |
+
vt_src = gr.Image(source="upload", type="filepath", label="Person")
|
164 |
+
gr.Examples(examples=person1, inputs=vt_src)
|
165 |
+
|
166 |
with gr.Column():
|
167 |
+
vt_ref = gr.Image(source="upload", type="filepath", label="Garment")
|
168 |
+
gr.Examples(examples=garments, inputs=vt_ref)
|
169 |
+
|
170 |
with gr.Column():
|
171 |
+
vt_out = gr.Image(label="Result")
|
172 |
+
vt_mask = gr.Image(label="Mask")
|
173 |
+
vt_dp = gr.Image(label="DensePose")
|
174 |
+
vt_btn = gr.Button("Generate")
|
175 |
+
|
176 |
+
with gr.Accordion("Advanced Options", open=False):
|
177 |
+
vt_model = gr.Radio(["viton_hd","dress_code"], value="viton_hd", label="Model")
|
178 |
+
vt_garment = gr.Radio(["upper_body","lower_body","dresses"], value="upper_body", label="Garment Type")
|
179 |
+
vt_accel_ref = gr.Checkbox(label="Accelerate Reference UNet")
|
180 |
+
vt_repaint = gr.Checkbox(label="Repaint Mode")
|
181 |
+
vt_steps = gr.Slider(30,100,value=30,step=1,label="Steps")
|
182 |
+
vt_scale = gr.Slider(0.1,5.0,value=2.5,step=0.1,label="Guidance Scale")
|
183 |
+
vt_seed = gr.Number(value=42, label="Seed")
|
184 |
+
|
185 |
+
vt_btn.click(
|
186 |
+
fn=lp.predict_virtual_tryon,
|
187 |
+
inputs=[vt_src, vt_ref, vt_accel_ref, vt_steps, vt_scale, vt_seed, vt_model, vt_garment, vt_repaint],
|
188 |
+
outputs=[vt_out, vt_mask, vt_dp],
|
189 |
+
)
|
|
|
|
|
190 |
|
191 |
with gr.Tab("Pose Transfer"):
|
192 |
with gr.Row():
|
193 |
with gr.Column():
|
194 |
+
pt_src = gr.Image(source="upload", type="filepath", label="Source Pose")
|
195 |
+
gr.Examples(examples=person2, inputs=pt_src)
|
196 |
+
|
197 |
with gr.Column():
|
198 |
+
pt_ref = gr.Image(source="upload", type="filepath", label="Target Person")
|
199 |
+
gr.Examples(examples=person1, inputs=pt_ref)
|
200 |
+
|
201 |
with gr.Column():
|
202 |
+
pt_out = gr.Image(label="Result")
|
203 |
+
pt_mask = gr.Image(label="Mask")
|
204 |
+
pt_dp = gr.Image(label="DensePose")
|
205 |
+
pt_btn = gr.Button("Generate")
|
206 |
+
|
207 |
+
with gr.Accordion("Advanced Options", open=False):
|
208 |
+
pt_accel_ref = gr.Checkbox(label="Accelerate Reference UNet")
|
209 |
+
pt_steps = gr.Slider(30,100,value=30,step=1,label="Steps")
|
210 |
+
pt_scale = gr.Slider(0.1,5.0,value=2.5,step=0.1,label="Guidance Scale")
|
211 |
+
pt_seed = gr.Number(value=42, label="Seed")
|
212 |
+
|
213 |
+
pt_btn.click(
|
214 |
+
fn=lp.predict_pose_transfer,
|
215 |
+
inputs=[pt_src, pt_ref, pt_accel_ref, pt_steps, pt_scale, pt_seed],
|
216 |
+
outputs=[pt_out, pt_mask, pt_dp],
|
217 |
+
)
|
218 |
+
|
219 |
+
gr.Markdown(note)
|
220 |
+
|
221 |
+
# expose publicly
|
222 |
+
demo.launch(share=True, server_port=7860,
|
223 |
+
allowed_paths=["./ckpts/examples"])
|