Spaces:
Runtime error
Runtime error
saner defaults, more input sanitization, shorter queue
Browse files- README.md +0 -2
- app.py +62 -58
- example.webp +0 -3
- examples/example_04_furry_moster/params.json +1 -1
- examples/example_06_sophie/params.json +1 -1
- makeavid_sd/inference.py +5 -5
README.md
CHANGED
|
@@ -19,5 +19,3 @@ models:
|
|
| 19 |
tags:
|
| 20 |
- jax-diffusers-event
|
| 21 |
---
|
| 22 |
-
|
| 23 |
-
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
|
| 19 |
tags:
|
| 20 |
- jax-diffusers-event
|
| 21 |
---
|
|
|
|
|
|
app.py
CHANGED
|
@@ -33,7 +33,7 @@ if _model.failed != False:
|
|
| 33 |
|
| 34 |
_examples = []
|
| 35 |
_expath = 'examples'
|
| 36 |
-
for x in os.listdir(_expath):
|
| 37 |
with open(os.path.join(_expath, x, 'params.json'), 'r') as f:
|
| 38 |
ex = json.load(f)
|
| 39 |
ex['image_input'] = None
|
|
@@ -56,22 +56,23 @@ def generate(
|
|
| 56 |
cfg = 15.0,
|
| 57 |
cfg_image = 9.0,
|
| 58 |
seed = 0,
|
| 59 |
-
fps =
|
| 60 |
num_frames = 24,
|
| 61 |
height = 512,
|
| 62 |
width = 512,
|
| 63 |
-
scheduler_type = '
|
| 64 |
-
output_format = '
|
| 65 |
) -> str:
|
| 66 |
-
num_frames = int(num_frames)
|
| 67 |
-
inference_steps = int(inference_steps)
|
| 68 |
-
height = int(height)
|
| 69 |
-
width = int(width)
|
| 70 |
height = (height // 64) * 64
|
| 71 |
width = (width // 64) * 64
|
| 72 |
cfg = max(cfg, 1.0)
|
| 73 |
cfg_image = max(cfg_image, 1.0)
|
| 74 |
-
|
|
|
|
| 75 |
if seed < 0:
|
| 76 |
seed = -seed
|
| 77 |
if hint_image is not None:
|
|
@@ -79,11 +80,12 @@ def generate(
|
|
| 79 |
hint_image = hint_image.convert('RGB')
|
| 80 |
if hint_image.size != (width, height):
|
| 81 |
hint_image = ImageOps.fit(hint_image, (width, height), method = Image.Resampling.LANCZOS)
|
|
|
|
| 82 |
if scheduler_type not in SCHEDULERS:
|
| 83 |
-
scheduler_type = '
|
| 84 |
output_format = output_format.lower()
|
| 85 |
if output_format not in _output_formats:
|
| 86 |
-
output_format = '
|
| 87 |
mask_image = None
|
| 88 |
images = _model.generate(
|
| 89 |
prompt = [prompt] * _model.device_count,
|
|
@@ -100,26 +102,24 @@ def generate(
|
|
| 100 |
scheduler_type = scheduler_type
|
| 101 |
)
|
| 102 |
_seen_compilations.add((hint_image is None, inference_steps, height, width, num_frames))
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
first_data = f'data:image/png;base64,' + base64.b64encode(buffer.getvalue()).decode()
|
| 122 |
-
buffer.close()
|
| 123 |
return data, last_data, first_data
|
| 124 |
|
| 125 |
def check_if_compiled(hint_image, inference_steps, height, width, num_frames, scheduler_type, message):
|
|
@@ -140,11 +140,11 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 140 |
intro1 = gr.Markdown("""
|
| 141 |
# Make-A-Video Stable Diffusion JAX
|
| 142 |
|
| 143 |
-
We have extended a pretrained
|
| 144 |
-
|
| 145 |
In this demo the hint image can be given by the user, otherwise it is generated by an generative image model.
|
| 146 |
|
| 147 |
-
The temporal layers are a port of [Make-A-Video PyTorch](https://github.com/lucidrains/make-a-video-pytorch) to FLAX.
|
| 148 |
The convolution is pseudo 3D and seperately convolves accross the spatial dimension in 2D and over the temporal dimension in 1D.
|
| 149 |
Temporal attention is purely self attention and also separately attends to time.
|
| 150 |
|
|
@@ -160,7 +160,7 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 160 |
**Please be patient. The model might have to compile with current parameters.**
|
| 161 |
|
| 162 |
This can take up to 5 minutes on the first run, and 2-3 minutes on later runs.
|
| 163 |
-
The compilation will be cached and
|
| 164 |
will be much faster.
|
| 165 |
|
| 166 |
Changes to the following parameters require the model to compile
|
|
@@ -170,7 +170,9 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 170 |
- Input image vs. no input image
|
| 171 |
- Noise scheduler type
|
| 172 |
|
| 173 |
-
If you encounter any issues, please report them here: [Space discussions](https://huggingface.co/spaces/TempoFunk/makeavid-sd-jax/discussions)
|
|
|
|
|
|
|
| 174 |
""")
|
| 175 |
|
| 176 |
with gr.Row(variant = variant):
|
|
@@ -221,14 +223,14 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 221 |
inference_steps_input = gr.Slider(
|
| 222 |
label = 'Steps',
|
| 223 |
minimum = 2,
|
| 224 |
-
maximum =
|
| 225 |
value = 20,
|
| 226 |
step = 1,
|
| 227 |
interactive = True
|
| 228 |
)
|
| 229 |
num_frames_input = gr.Slider(
|
| 230 |
label = 'Number of frames to generate',
|
| 231 |
-
minimum =
|
| 232 |
maximum = 24,
|
| 233 |
step = 1,
|
| 234 |
value = 24,
|
|
@@ -236,7 +238,7 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 236 |
)
|
| 237 |
width_input = gr.Slider(
|
| 238 |
label = 'Width',
|
| 239 |
-
minimum =
|
| 240 |
maximum = 576,
|
| 241 |
step = 64,
|
| 242 |
value = 512,
|
|
@@ -244,7 +246,7 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 244 |
)
|
| 245 |
height_input = gr.Slider(
|
| 246 |
label = 'Height',
|
| 247 |
-
minimum =
|
| 248 |
maximum = 576,
|
| 249 |
step = 64,
|
| 250 |
value = 512,
|
|
@@ -253,7 +255,7 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 253 |
scheduler_input = gr.Dropdown(
|
| 254 |
label = 'Noise scheduler',
|
| 255 |
choices = list(SCHEDULERS.keys()),
|
| 256 |
-
value = '
|
| 257 |
interactive = True
|
| 258 |
)
|
| 259 |
with gr.Row():
|
|
@@ -279,32 +281,33 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 279 |
value = 'example.gif',
|
| 280 |
interactive = False
|
| 281 |
)
|
| 282 |
-
tips = gr.Markdown('🤫 *Secret tip*:
|
| 283 |
with gr.Row():
|
| 284 |
last_frame_output = gr.Image(
|
| 285 |
label = 'Last frame',
|
| 286 |
interactive = False
|
| 287 |
)
|
| 288 |
first_frame_output = gr.Image(
|
| 289 |
-
label = '
|
| 290 |
interactive = False
|
| 291 |
)
|
| 292 |
examples_lst = []
|
| 293 |
for x in _examples:
|
| 294 |
examples_lst.append([
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
|
|
|
| 308 |
])
|
| 309 |
examples = gr.Examples(
|
| 310 |
examples = examples_lst,
|
|
@@ -317,10 +320,11 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 317 |
cfg_image_input,
|
| 318 |
seed_input,
|
| 319 |
fps_input,
|
|
|
|
|
|
|
| 320 |
num_frames_input,
|
| 321 |
height_input,
|
| 322 |
width_input,
|
| 323 |
-
scheduler_input,
|
| 324 |
output_format
|
| 325 |
],
|
| 326 |
postprocess = False
|
|
@@ -355,6 +359,6 @@ with gr.Blocks(title = 'Make-A-Video Stable Diffusion JAX', analytics_enabled =
|
|
| 355 |
)
|
| 356 |
#cancel_button.click(fn = lambda: None, cancels = ev)
|
| 357 |
|
| 358 |
-
demo.queue(concurrency_count = 1, max_size =
|
| 359 |
demo.launch()
|
| 360 |
|
|
|
|
| 33 |
|
| 34 |
_examples = []
|
| 35 |
_expath = 'examples'
|
| 36 |
+
for x in sorted(os.listdir(_expath)):
|
| 37 |
with open(os.path.join(_expath, x, 'params.json'), 'r') as f:
|
| 38 |
ex = json.load(f)
|
| 39 |
ex['image_input'] = None
|
|
|
|
| 56 |
cfg = 15.0,
|
| 57 |
cfg_image = 9.0,
|
| 58 |
seed = 0,
|
| 59 |
+
fps = 12,
|
| 60 |
num_frames = 24,
|
| 61 |
height = 512,
|
| 62 |
width = 512,
|
| 63 |
+
scheduler_type = 'dpm',
|
| 64 |
+
output_format = 'gif'
|
| 65 |
) -> str:
|
| 66 |
+
num_frames = min(24, max(2, int(num_frames)))
|
| 67 |
+
inference_steps = min(60, max(2, int(inference_steps)))
|
| 68 |
+
height = min(576, max(256, int(height)))
|
| 69 |
+
width = min(576, max(256, int(width)))
|
| 70 |
height = (height // 64) * 64
|
| 71 |
width = (width // 64) * 64
|
| 72 |
cfg = max(cfg, 1.0)
|
| 73 |
cfg_image = max(cfg_image, 1.0)
|
| 74 |
+
fps = min(1000, max(1, int(fps)))
|
| 75 |
+
seed = min(2**32-2, int(seed))
|
| 76 |
if seed < 0:
|
| 77 |
seed = -seed
|
| 78 |
if hint_image is not None:
|
|
|
|
| 80 |
hint_image = hint_image.convert('RGB')
|
| 81 |
if hint_image.size != (width, height):
|
| 82 |
hint_image = ImageOps.fit(hint_image, (width, height), method = Image.Resampling.LANCZOS)
|
| 83 |
+
scheduler_type = scheduler_type.lower()
|
| 84 |
if scheduler_type not in SCHEDULERS:
|
| 85 |
+
scheduler_type = 'dpm'
|
| 86 |
output_format = output_format.lower()
|
| 87 |
if output_format not in _output_formats:
|
| 88 |
+
output_format = 'gif'
|
| 89 |
mask_image = None
|
| 90 |
images = _model.generate(
|
| 91 |
prompt = [prompt] * _model.device_count,
|
|
|
|
| 102 |
scheduler_type = scheduler_type
|
| 103 |
)
|
| 104 |
_seen_compilations.add((hint_image is None, inference_steps, height, width, num_frames))
|
| 105 |
+
with BytesIO() as buffer:
|
| 106 |
+
images[1].save(
|
| 107 |
+
buffer,
|
| 108 |
+
format = output_format,
|
| 109 |
+
save_all = True,
|
| 110 |
+
append_images = images[2:],
|
| 111 |
+
loop = 0,
|
| 112 |
+
duration = round(1000 / fps),
|
| 113 |
+
allow_mixed = True,
|
| 114 |
+
optimize = True
|
| 115 |
+
)
|
| 116 |
+
data = f'data:image/{output_format};base64,' + base64.b64encode(buffer.getvalue()).decode()
|
| 117 |
+
with BytesIO() as buffer:
|
| 118 |
+
images[-1].save(buffer, format = 'png', optimize = True)
|
| 119 |
+
last_data = f'data:image/png;base64,' + base64.b64encode(buffer.getvalue()).decode()
|
| 120 |
+
with BytesIO() as buffer:
|
| 121 |
+
images[0].save(buffer, format ='png', optimize = True)
|
| 122 |
+
first_data = f'data:image/png;base64,' + base64.b64encode(buffer.getvalue()).decode()
|
|
|
|
|
|
|
| 123 |
return data, last_data, first_data
|
| 124 |
|
| 125 |
def check_if_compiled(hint_image, inference_steps, height, width, num_frames, scheduler_type, message):
|
|
|
|
| 140 |
intro1 = gr.Markdown("""
|
| 141 |
# Make-A-Video Stable Diffusion JAX
|
| 142 |
|
| 143 |
+
We have extended a pretrained latent-diffusion inpainting image generation model with **temporal convolutions and attention**.
|
| 144 |
+
We guide the video generation with a hint image by taking advantage of the extra 5 input channels of the inpainting model.
|
| 145 |
In this demo the hint image can be given by the user, otherwise it is generated by an generative image model.
|
| 146 |
|
| 147 |
+
The temporal layers are a port of [Make-A-Video PyTorch](https://github.com/lucidrains/make-a-video-pytorch) to [JAX](https://github.com/google/jax) utilizing [FLAX](https://github.com/google/flax).
|
| 148 |
The convolution is pseudo 3D and seperately convolves accross the spatial dimension in 2D and over the temporal dimension in 1D.
|
| 149 |
Temporal attention is purely self attention and also separately attends to time.
|
| 150 |
|
|
|
|
| 160 |
**Please be patient. The model might have to compile with current parameters.**
|
| 161 |
|
| 162 |
This can take up to 5 minutes on the first run, and 2-3 minutes on later runs.
|
| 163 |
+
The compilation will be cached and later runs with the same parameters
|
| 164 |
will be much faster.
|
| 165 |
|
| 166 |
Changes to the following parameters require the model to compile
|
|
|
|
| 170 |
- Input image vs. no input image
|
| 171 |
- Noise scheduler type
|
| 172 |
|
| 173 |
+
If you encounter any issues, please report them here: [Space discussions](https://huggingface.co/spaces/TempoFunk/makeavid-sd-jax/discussions) (or DM [@lopho](https://twitter.com/lopho))
|
| 174 |
+
|
| 175 |
+
<small>Leave a ❤️ like if you like. Consider it a dopamine donation at no cost.</small>
|
| 176 |
""")
|
| 177 |
|
| 178 |
with gr.Row(variant = variant):
|
|
|
|
| 223 |
inference_steps_input = gr.Slider(
|
| 224 |
label = 'Steps',
|
| 225 |
minimum = 2,
|
| 226 |
+
maximum = 60,
|
| 227 |
value = 20,
|
| 228 |
step = 1,
|
| 229 |
interactive = True
|
| 230 |
)
|
| 231 |
num_frames_input = gr.Slider(
|
| 232 |
label = 'Number of frames to generate',
|
| 233 |
+
minimum = 2,
|
| 234 |
maximum = 24,
|
| 235 |
step = 1,
|
| 236 |
value = 24,
|
|
|
|
| 238 |
)
|
| 239 |
width_input = gr.Slider(
|
| 240 |
label = 'Width',
|
| 241 |
+
minimum = 256,
|
| 242 |
maximum = 576,
|
| 243 |
step = 64,
|
| 244 |
value = 512,
|
|
|
|
| 246 |
)
|
| 247 |
height_input = gr.Slider(
|
| 248 |
label = 'Height',
|
| 249 |
+
minimum = 256,
|
| 250 |
maximum = 576,
|
| 251 |
step = 64,
|
| 252 |
value = 512,
|
|
|
|
| 255 |
scheduler_input = gr.Dropdown(
|
| 256 |
label = 'Noise scheduler',
|
| 257 |
choices = list(SCHEDULERS.keys()),
|
| 258 |
+
value = 'dpm',
|
| 259 |
interactive = True
|
| 260 |
)
|
| 261 |
with gr.Row():
|
|
|
|
| 281 |
value = 'example.gif',
|
| 282 |
interactive = False
|
| 283 |
)
|
| 284 |
+
tips = gr.Markdown('🤫 *Secret tip*: try using the last frame as input for the next generation.')
|
| 285 |
with gr.Row():
|
| 286 |
last_frame_output = gr.Image(
|
| 287 |
label = 'Last frame',
|
| 288 |
interactive = False
|
| 289 |
)
|
| 290 |
first_frame_output = gr.Image(
|
| 291 |
+
label = 'Initial frame',
|
| 292 |
interactive = False
|
| 293 |
)
|
| 294 |
examples_lst = []
|
| 295 |
for x in _examples:
|
| 296 |
examples_lst.append([
|
| 297 |
+
x['image_output'],
|
| 298 |
+
x['prompt'],
|
| 299 |
+
x['neg_prompt'],
|
| 300 |
+
x['image_input'],
|
| 301 |
+
x['cfg'],
|
| 302 |
+
x['cfg_image'],
|
| 303 |
+
x['seed'],
|
| 304 |
+
x['fps'],
|
| 305 |
+
x['steps'],
|
| 306 |
+
x['scheduler'],
|
| 307 |
+
x['num_frames'],
|
| 308 |
+
x['height'],
|
| 309 |
+
x['width'],
|
| 310 |
+
x['format']
|
| 311 |
])
|
| 312 |
examples = gr.Examples(
|
| 313 |
examples = examples_lst,
|
|
|
|
| 320 |
cfg_image_input,
|
| 321 |
seed_input,
|
| 322 |
fps_input,
|
| 323 |
+
inference_steps_input,
|
| 324 |
+
scheduler_input,
|
| 325 |
num_frames_input,
|
| 326 |
height_input,
|
| 327 |
width_input,
|
|
|
|
| 328 |
output_format
|
| 329 |
],
|
| 330 |
postprocess = False
|
|
|
|
| 359 |
)
|
| 360 |
#cancel_button.click(fn = lambda: None, cancels = ev)
|
| 361 |
|
| 362 |
+
demo.queue(concurrency_count = 1, max_size = 8)
|
| 363 |
demo.launch()
|
| 364 |
|
example.webp
DELETED
Git LFS Details
|
examples/example_04_furry_moster/params.json
CHANGED
|
@@ -8,7 +8,7 @@
|
|
| 8 |
"width": 512,
|
| 9 |
"height": 512,
|
| 10 |
"scheduler": "dpm",
|
| 11 |
-
"fps":
|
| 12 |
"format": "gif",
|
| 13 |
"num_frames": 24
|
| 14 |
}
|
|
|
|
| 8 |
"width": 512,
|
| 9 |
"height": 512,
|
| 10 |
"scheduler": "dpm",
|
| 11 |
+
"fps": 12,
|
| 12 |
"format": "gif",
|
| 13 |
"num_frames": 24
|
| 14 |
}
|
examples/example_06_sophie/params.json
CHANGED
|
@@ -3,7 +3,7 @@
|
|
| 3 |
"neg_prompt": "",
|
| 4 |
"cfg": 15,
|
| 5 |
"cfg_image": 9,
|
| 6 |
-
"seed":
|
| 7 |
"steps": 20,
|
| 8 |
"width": 512,
|
| 9 |
"height": 512,
|
|
|
|
| 3 |
"neg_prompt": "",
|
| 4 |
"cfg": 15,
|
| 5 |
"cfg_image": 9,
|
| 6 |
+
"seed": 0,
|
| 7 |
"steps": 20,
|
| 8 |
"width": 512,
|
| 9 |
"height": 512,
|
makeavid_sd/inference.py
CHANGED
|
@@ -45,8 +45,8 @@ SchedulerStateType = Union[
|
|
| 45 |
]
|
| 46 |
|
| 47 |
SCHEDULERS: Dict[str, SchedulerType] = {
|
| 48 |
-
'
|
| 49 |
-
'
|
| 50 |
#'PLMS': FlaxPNDMScheduler, # its not correctly implemented in diffusers, output is bad, but at least it "works"
|
| 51 |
#'LMS': FlaxLMSDiscreteScheduler, # borked
|
| 52 |
# image_latents, image_scheduler_state = scheduler.step(
|
|
@@ -224,8 +224,8 @@ class InferenceUNetPseudo3D:
|
|
| 224 |
return tokens, neg_tokens, hint, mask
|
| 225 |
|
| 226 |
def generate(self,
|
| 227 |
-
prompt: Union[str, List[str]],
|
| 228 |
-
inference_steps: int,
|
| 229 |
hint_image: Union[Image.Image, List[Image.Image], None] = None,
|
| 230 |
mask_image: Union[Image.Image, List[Image.Image], None] = None,
|
| 231 |
neg_prompt: Union[str, List[str]] = '',
|
|
@@ -235,7 +235,7 @@ class InferenceUNetPseudo3D:
|
|
| 235 |
width: int = 512,
|
| 236 |
height: int = 512,
|
| 237 |
seed: int = 0,
|
| 238 |
-
scheduler_type: str = '
|
| 239 |
) -> List[List[Image.Image]]:
|
| 240 |
assert inference_steps > 0, f'number of inference steps must be > 0 but is {inference_steps}'
|
| 241 |
assert num_frames > 0, f'number of frames must be > 0 but is {num_frames}'
|
|
|
|
| 45 |
]
|
| 46 |
|
| 47 |
SCHEDULERS: Dict[str, SchedulerType] = {
|
| 48 |
+
'dpm': FlaxDPMSolverMultistepScheduler, # husbando
|
| 49 |
+
'ddim': FlaxDDIMScheduler,
|
| 50 |
#'PLMS': FlaxPNDMScheduler, # its not correctly implemented in diffusers, output is bad, but at least it "works"
|
| 51 |
#'LMS': FlaxLMSDiscreteScheduler, # borked
|
| 52 |
# image_latents, image_scheduler_state = scheduler.step(
|
|
|
|
| 224 |
return tokens, neg_tokens, hint, mask
|
| 225 |
|
| 226 |
def generate(self,
|
| 227 |
+
prompt: Union[str, List[str]] = '',
|
| 228 |
+
inference_steps: int = 20,
|
| 229 |
hint_image: Union[Image.Image, List[Image.Image], None] = None,
|
| 230 |
mask_image: Union[Image.Image, List[Image.Image], None] = None,
|
| 231 |
neg_prompt: Union[str, List[str]] = '',
|
|
|
|
| 235 |
width: int = 512,
|
| 236 |
height: int = 512,
|
| 237 |
seed: int = 0,
|
| 238 |
+
scheduler_type: str = 'dpm'
|
| 239 |
) -> List[List[Image.Image]]:
|
| 240 |
assert inference_steps > 0, f'number of inference steps must be > 0 but is {inference_steps}'
|
| 241 |
assert num_frames > 0, f'number of frames must be > 0 but is {num_frames}'
|