Spaces:
Running
on
Zero
Running
on
Zero
<feat> complete app.py.
Browse files- .gitignore +1 -0
- app.py +261 -0
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
__pycache__/
|
app.py
CHANGED
@@ -18,10 +18,13 @@ from peft import LoraConfig
|
|
18 |
from omegaconf import OmegaConf
|
19 |
from safetensors.torch import safe_open
|
20 |
from PIL import Image, ImageDraw, ImageFilter
|
|
|
|
|
21 |
|
22 |
from models import HunyuanVideoTransformer3DModel
|
23 |
from pipelines import HunyuanVideoImageToVideoPipeline
|
24 |
|
|
|
25 |
header = """
|
26 |
# DRA-Ctrl Gradio App
|
27 |
|
@@ -33,9 +36,267 @@ header = """
|
|
33 |
</div>
|
34 |
"""
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
def create_app():
|
37 |
with gr.Blocks() as app:
|
38 |
gr.Markdown(header, elem_id="header")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
|
40 |
return app
|
41 |
|
|
|
18 |
from omegaconf import OmegaConf
|
19 |
from safetensors.torch import safe_open
|
20 |
from PIL import Image, ImageDraw, ImageFilter
|
21 |
+
from huggingface_hub import hf_hub_download
|
22 |
+
from transformers import pipeline
|
23 |
|
24 |
from models import HunyuanVideoTransformer3DModel
|
25 |
from pipelines import HunyuanVideoImageToVideoPipeline
|
26 |
|
27 |
+
|
28 |
header = """
|
29 |
# DRA-Ctrl Gradio App
|
30 |
|
|
|
36 |
</div>
|
37 |
"""
|
38 |
|
39 |
+
notice = """
|
40 |
+
For easier testing, in spatially-aligned image generation tasks, when passing the condition image to `gradio_app`,
|
41 |
+
there's no need to manually input edge maps, depth maps, or other condition images - only the original image is required.
|
42 |
+
The corresponding condition images will be automatically extracted.
|
43 |
+
"""
|
44 |
+
|
45 |
+
|
46 |
+
|
47 |
+
@spaces.GPU
|
48 |
+
def process_image_and_text(condition_image, target_prompt, condition_image_prompt, task):
|
49 |
+
# init models
|
50 |
+
transformer = HunyuanVideoTransformer3DModel.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V',
|
51 |
+
subfolder="transformer",
|
52 |
+
inference_subject_driven=task in ['subject_driven'])
|
53 |
+
scheduler = diffusers.FlowMatchEulerDiscreteScheduler()
|
54 |
+
vae = diffusers.AutoencoderKLHunyuanVideo.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V',
|
55 |
+
subfolder="vae")
|
56 |
+
text_encoder = transformers.LlavaForConditionalGeneration.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V',
|
57 |
+
subfolder="text_encoder")
|
58 |
+
text_encoder_2 = transformers.CLIPTextModel.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V',
|
59 |
+
subfolder="text_encoder_2")
|
60 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V',
|
61 |
+
subfolder="tokenizer")
|
62 |
+
tokenizer_2 = transformers.CLIPTokenizer.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V',
|
63 |
+
subfolder="tokenizer_2")
|
64 |
+
image_processor = transformers.CLIPImageProcessor.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V',
|
65 |
+
subfolder="image_processor")
|
66 |
+
|
67 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
68 |
+
weight_dtype = torch.bfloat16
|
69 |
+
|
70 |
+
transformer.requires_grad_(False)
|
71 |
+
vae.requires_grad_(False).to(device, dtype=weight_dtype)
|
72 |
+
text_encoder.requires_grad_(False).to(device, dtype=weight_dtype)
|
73 |
+
text_encoder_2.requires_grad_(False).to(device, dtype=weight_dtype)
|
74 |
+
transformer.to(device, dtype=weight_dtype)
|
75 |
+
vae.enable_tiling()
|
76 |
+
vae.enable_slicing()
|
77 |
+
|
78 |
+
# insert LoRA
|
79 |
+
lora_config = LoraConfig(
|
80 |
+
r=16,
|
81 |
+
lora_alpha=16,
|
82 |
+
init_lora_weights="gaussian",
|
83 |
+
target_modules=[
|
84 |
+
'attn.to_k', 'attn.to_q', 'attn.to_v', 'attn.to_out.0',
|
85 |
+
'attn.add_k_proj', 'attn.add_q_proj', 'attn.add_v_proj', 'attn.to_add_out',
|
86 |
+
'ff.net.0.proj', 'ff.net.2',
|
87 |
+
'ff_context.net.0.proj', 'ff_context.net.2',
|
88 |
+
'norm1_context.linear', 'norm1.linear',
|
89 |
+
'norm.linear', 'proj_mlp', 'proj_out',
|
90 |
+
]
|
91 |
+
)
|
92 |
+
transformer.add_adapter(lora_config)
|
93 |
+
|
94 |
+
# hack LoRA forward
|
95 |
+
def create_hacked_forward(module):
|
96 |
+
lora_forward = module.forward
|
97 |
+
non_lora_forward = module.base_layer.forward
|
98 |
+
img_sequence_length = int((args.img_size / 8 / 2) ** 2)
|
99 |
+
encoder_sequence_length = 144 + 252 # encoder sequence: 144 img 252 txt
|
100 |
+
num_imgs = 4
|
101 |
+
num_generated_imgs = 3
|
102 |
+
num_encoder_sequences = 2 if args.task in ['subject_driven', 'style_transfer'] else 1
|
103 |
+
|
104 |
+
def hacked_lora_forward(self, x, *args, **kwargs):
|
105 |
+
if x.shape[1] == img_sequence_length * num_imgs and len(x.shape) > 2:
|
106 |
+
return torch.cat((
|
107 |
+
lora_forward(x[:, :-img_sequence_length*num_generated_imgs], *args, **kwargs),
|
108 |
+
non_lora_forward(x[:, -img_sequence_length*num_generated_imgs:], *args, **kwargs)
|
109 |
+
), dim=1)
|
110 |
+
elif x.shape[1] == encoder_sequence_length * num_encoder_sequences or x.shape[1] == encoder_sequence_length:
|
111 |
+
return lora_forward(x, *args, **kwargs)
|
112 |
+
elif x.shape[1] == img_sequence_length * num_imgs + encoder_sequence_length * num_encoder_sequences:
|
113 |
+
return torch.cat((
|
114 |
+
lora_forward(x[:, :(num_imgs - num_generated_imgs)*img_sequence_length], *args, **kwargs),
|
115 |
+
non_lora_forward(x[:, (num_imgs - num_generated_imgs)*img_sequence_length:-num_encoder_sequences*encoder_sequence_length], *args, **kwargs),
|
116 |
+
lora_forward(x[:, -num_encoder_sequences*encoder_sequence_length:], *args, **kwargs)
|
117 |
+
), dim=1)
|
118 |
+
elif x.shape[1] == 3072:
|
119 |
+
return non_lora_forward(x, *args, **kwargs)
|
120 |
+
else:
|
121 |
+
raise ValueError(
|
122 |
+
f"hacked_lora_forward receives unexpected sequence length: {x.shape[1]}, input shape: {x.shape}!"
|
123 |
+
)
|
124 |
+
|
125 |
+
return hacked_lora_forward.__get__(module, type(module))
|
126 |
+
|
127 |
+
for n, m in transformer.named_modules():
|
128 |
+
if isinstance(m, peft.tuners.lora.layer.Linear):
|
129 |
+
m.forward = create_hacked_forward(m)
|
130 |
+
|
131 |
+
# load LoRA weights
|
132 |
+
model_root = hf_hub_download(
|
133 |
+
repo_id="Kunbyte/DRA-Ctrl",
|
134 |
+
filename=f"{task}.safetensors",
|
135 |
+
resume_download=True)
|
136 |
+
|
137 |
+
try:
|
138 |
+
with safe_open(model_root, framework="pt") as f:
|
139 |
+
lora_weights = {}
|
140 |
+
for k in f.keys():
|
141 |
+
param = f.get_tensor(k)
|
142 |
+
if k.endswith(".weight"):
|
143 |
+
k = k.replace('.weight', '.default.weight')
|
144 |
+
lora_weights[k] = param
|
145 |
+
transformer.load_state_dict(lora_weights, strict=False)
|
146 |
+
except Exception as e:
|
147 |
+
raise ValueError(f'{e}')
|
148 |
+
|
149 |
+
transformer.requires_grad_(False)
|
150 |
+
|
151 |
+
pipe = HunyuanVideoImageToVideoPipeline(
|
152 |
+
text_encoder=text_encoder,
|
153 |
+
tokenizer=tokenizer,
|
154 |
+
transformer=transformer,
|
155 |
+
vae=vae,
|
156 |
+
scheduler=copy.deepcopy(scheduler),
|
157 |
+
text_encoder_2=text_encoder_2,
|
158 |
+
tokenizer_2=tokenizer_2,
|
159 |
+
image_processor=image_processor,
|
160 |
+
)
|
161 |
+
|
162 |
+
# start generation
|
163 |
+
c_txt = None if condition_image_prompt == "" else condition_image_prompt
|
164 |
+
c_img = condition_image.resize((512, 512))
|
165 |
+
t_txt = target_prompt
|
166 |
+
|
167 |
+
if args.task not in ['subject_driven', 'style_transfer']:
|
168 |
+
if args.task == "canny":
|
169 |
+
def get_canny_edge(img):
|
170 |
+
img_np = np.array(img)
|
171 |
+
img_gray = cv2.cvtColor(img_np, cv2.COLOR_BGR2GRAY)
|
172 |
+
edges = cv2.Canny(img_gray, 100, 200)
|
173 |
+
edges_tmp = Image.fromarray(edges).convert("RGB")
|
174 |
+
edges_tmp.save(os.path.join(save_dir, f"edges.png"))
|
175 |
+
edges[edges == 0] = 128
|
176 |
+
return Image.fromarray(edges).convert("RGB")
|
177 |
+
c_img = get_canny_edge(c_img)
|
178 |
+
elif args.task == "coloring":
|
179 |
+
c_img = (
|
180 |
+
c_img.resize((args.img_size, args.img_size))
|
181 |
+
.convert("L")
|
182 |
+
.convert("RGB")
|
183 |
+
)
|
184 |
+
elif args.task == "deblurring":
|
185 |
+
blur_radius = 10
|
186 |
+
c_img = (
|
187 |
+
c_img.convert("RGB")
|
188 |
+
.filter(ImageFilter.GaussianBlur(blur_radius))
|
189 |
+
.resize((args.img_size, args.img_size))
|
190 |
+
.convert("RGB")
|
191 |
+
)
|
192 |
+
elif args.task == "depth":
|
193 |
+
def get_depth_map(img):
|
194 |
+
from transformers import pipeline
|
195 |
+
|
196 |
+
depth_pipe = pipeline(
|
197 |
+
task="depth-estimation",
|
198 |
+
model="LiheYoung/depth-anything-small-hf",
|
199 |
+
device="cpu",
|
200 |
+
)
|
201 |
+
return depth_pipe(img)["depth"].convert("RGB").resize((args.img_size, args.img_size))
|
202 |
+
c_img = get_depth_map(c_img)
|
203 |
+
c_img.save(os.path.join(save_dir, f"depth.png"))
|
204 |
+
k = (255 - 128) / 255
|
205 |
+
b = 128
|
206 |
+
c_img = c_img.point(lambda x: k * x + b)
|
207 |
+
elif args.task == "depth_pred":
|
208 |
+
c_img = c_img
|
209 |
+
elif args.task == "fill":
|
210 |
+
c_img = c_img.resize((args.img_size, args.img_size)).convert("RGB")
|
211 |
+
x1, x2 = args.fill_x1, args.fill_x2
|
212 |
+
y1, y2 = args.fill_y1, args.fill_y2
|
213 |
+
mask = Image.new("L", (args.img_size, args.img_size), 0)
|
214 |
+
draw = ImageDraw.Draw(mask)
|
215 |
+
draw.rectangle((x1, y1, x2, y2), fill=255)
|
216 |
+
if args.inpainting:
|
217 |
+
mask = Image.eval(mask, lambda a: 255 - a)
|
218 |
+
c_img = Image.composite(
|
219 |
+
c_img,
|
220 |
+
Image.new("RGB", (args.img_size, args.img_size), (255, 255, 255)),
|
221 |
+
mask
|
222 |
+
)
|
223 |
+
c_img.save(os.path.join(save_dir, f"mask.png"))
|
224 |
+
c_img = Image.composite(
|
225 |
+
c_img,
|
226 |
+
Image.new("RGB", (args.img_size, args.img_size), (128, 128, 128)),
|
227 |
+
mask
|
228 |
+
)
|
229 |
+
elif args.task == "sr":
|
230 |
+
c_img = c_img.resize((int(args.img_size / 4), int(args.img_size / 4))).convert("RGB")
|
231 |
+
c_img.save(os.path.join(save_dir, f"low_resolution.png"))
|
232 |
+
c_img = c_img.resize((args.img_size, args.img_size))
|
233 |
+
c_img.save(os.path.join(save_dir, f"low_to_high.png"))
|
234 |
+
|
235 |
+
gen_img = pipe(
|
236 |
+
image=c_img,
|
237 |
+
prompt=[t_txt.strip()],
|
238 |
+
prompt_condition=[c_txt.strip()] if c_txt is not None else None,
|
239 |
+
prompt_2=[t_txt],
|
240 |
+
height=512,
|
241 |
+
width=512,
|
242 |
+
num_frames=5,
|
243 |
+
num_inference_steps=50,
|
244 |
+
guidance_scale=6.0,
|
245 |
+
num_videos_per_prompt=1,
|
246 |
+
generator=torch.Generator(device=pipe.transformer.device).manual_seed(0),
|
247 |
+
output_type='pt',
|
248 |
+
image_embed_interleave=4,
|
249 |
+
frame_gap=48,
|
250 |
+
mixup=True,
|
251 |
+
mixup_num_imgs=2,
|
252 |
+
).frames
|
253 |
+
|
254 |
+
gen_img = gen_img[:, 0:1, :, :, :]
|
255 |
+
gen_img = gen_img.squeeze(0).squeeze(0).cpu().to(torch.float32).numpy()
|
256 |
+
gen_img = np.transpose(gen_img, (1, 2, 0))
|
257 |
+
gen_img = (gen_img * 255).astype(np.uint8)
|
258 |
+
gen_img = Image.fromarray(gen_img)
|
259 |
+
|
260 |
+
return gen_img
|
261 |
+
|
262 |
def create_app():
|
263 |
with gr.Blocks() as app:
|
264 |
gr.Markdown(header, elem_id="header")
|
265 |
+
with gr.Row(equal_height=False):
|
266 |
+
with gr.Column(variant="panel", elem_classes="inputPanel"):
|
267 |
+
condition_image = gr.Image(
|
268 |
+
type="pil", label="Condition Image", width=300, elem_id="input"
|
269 |
+
)
|
270 |
+
task = gr.Radio(
|
271 |
+
[
|
272 |
+
("Subject-driven Image Generation", "subject_driven"),
|
273 |
+
("Canny-to-Image", "canny"),
|
274 |
+
("Colorization", "coloring"),
|
275 |
+
("Deblurring", "deblurring"),
|
276 |
+
("Depth-to-Image", "depth"),
|
277 |
+
("Depth Prediction", "depth_pred"),
|
278 |
+
("In/Out-Painting", "fill"),
|
279 |
+
("Super-Resolution", "sr"),
|
280 |
+
("Style Transfer", "style_transfer")
|
281 |
+
],
|
282 |
+
label="Task Selection",
|
283 |
+
value="subject_driven",
|
284 |
+
interactive=True,
|
285 |
+
elem_id="task_selection"
|
286 |
+
)
|
287 |
+
gr.Markdown(notice, elem_id="notice")
|
288 |
+
target_prompt = gr.Textbox(lines=2, label="Target Prompt", elem_id="text")
|
289 |
+
condition_image_prompt = gr.Textbox(lines=2, label="Condition Image Prompt", elem_id="text")
|
290 |
+
submit_btn = gr.Button("Run", elem_id="submit_btn")
|
291 |
+
|
292 |
+
with gr.Column(variant="panel", elem_classes="outputPanel"):
|
293 |
+
output_image = gr.Image(type="pil", elem_id="output")
|
294 |
+
|
295 |
+
submit_btn.click(
|
296 |
+
fn=process_image_and_text,
|
297 |
+
inputs=[condition_image, target_prompt, condition_image_prompt, task],
|
298 |
+
outputs=output_image,
|
299 |
+
)
|
300 |
|
301 |
return app
|
302 |
|