Spaces:
Runtime error
Runtime error
Main update
Browse files- README.md +1 -1
- app.py +81 -67
- generation_sdxl.py +474 -0
- requirements.txt +1 -0
README.md
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
emoji: πΌ
|
| 4 |
colorFrom: purple
|
| 5 |
colorTo: red
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Demo App
|
| 3 |
emoji: πΌ
|
| 4 |
colorFrom: purple
|
| 5 |
colorTo: red
|
app.py
CHANGED
|
@@ -1,46 +1,76 @@
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import numpy as np
|
| 3 |
import random
|
| 4 |
-
|
|
|
|
|
|
|
| 5 |
import torch
|
| 6 |
|
| 7 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 8 |
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
MAX_SEED = np.iinfo(np.int32).max
|
| 19 |
MAX_IMAGE_SIZE = 1024
|
| 20 |
|
| 21 |
-
|
|
|
|
|
|
|
| 22 |
|
| 23 |
if randomize_seed:
|
| 24 |
seed = random.randint(0, MAX_SEED)
|
| 25 |
|
| 26 |
generator = torch.Generator().manual_seed(seed)
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
)
|
| 37 |
-
|
| 38 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
|
| 40 |
examples = [
|
| 41 |
-
"Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
|
| 42 |
"An astronaut riding a green horse",
|
| 43 |
-
|
|
|
|
|
|
|
| 44 |
]
|
| 45 |
|
| 46 |
css="""
|
|
@@ -58,11 +88,20 @@ else:
|
|
| 58 |
with gr.Blocks(css=css) as demo:
|
| 59 |
|
| 60 |
with gr.Column(elem_id="col-container"):
|
| 61 |
-
gr.Markdown(
|
| 62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
Currently running on {power_device}.
|
| 64 |
-
"""
|
| 65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
with gr.Row():
|
| 67 |
|
| 68 |
prompt = gr.Text(
|
|
@@ -79,13 +118,6 @@ with gr.Blocks(css=css) as demo:
|
|
| 79 |
|
| 80 |
with gr.Accordion("Advanced Settings", open=False):
|
| 81 |
|
| 82 |
-
negative_prompt = gr.Text(
|
| 83 |
-
label="Negative prompt",
|
| 84 |
-
max_lines=1,
|
| 85 |
-
placeholder="Enter a negative prompt",
|
| 86 |
-
visible=False,
|
| 87 |
-
)
|
| 88 |
-
|
| 89 |
seed = gr.Slider(
|
| 90 |
label="Seed",
|
| 91 |
minimum=0,
|
|
@@ -94,53 +126,35 @@ with gr.Blocks(css=css) as demo:
|
|
| 94 |
value=0,
|
| 95 |
)
|
| 96 |
|
| 97 |
-
randomize_seed = gr.Checkbox(label="Randomize seed", value=
|
| 98 |
-
|
| 99 |
-
with gr.Row():
|
| 100 |
-
|
| 101 |
-
width = gr.Slider(
|
| 102 |
-
label="Width",
|
| 103 |
-
minimum=256,
|
| 104 |
-
maximum=MAX_IMAGE_SIZE,
|
| 105 |
-
step=32,
|
| 106 |
-
value=512,
|
| 107 |
-
)
|
| 108 |
-
|
| 109 |
-
height = gr.Slider(
|
| 110 |
-
label="Height",
|
| 111 |
-
minimum=256,
|
| 112 |
-
maximum=MAX_IMAGE_SIZE,
|
| 113 |
-
step=32,
|
| 114 |
-
value=512,
|
| 115 |
-
)
|
| 116 |
|
| 117 |
with gr.Row():
|
| 118 |
|
| 119 |
guidance_scale = gr.Slider(
|
| 120 |
label="Guidance scale",
|
| 121 |
minimum=0.0,
|
| 122 |
-
maximum=
|
| 123 |
-
step=0
|
| 124 |
-
value=
|
| 125 |
)
|
| 126 |
|
| 127 |
-
|
| 128 |
-
label="
|
| 129 |
-
minimum=
|
| 130 |
-
maximum=
|
| 131 |
-
step=1,
|
| 132 |
-
value=
|
| 133 |
)
|
| 134 |
|
| 135 |
gr.Examples(
|
| 136 |
examples = examples,
|
| 137 |
-
inputs = [prompt]
|
|
|
|
| 138 |
)
|
| 139 |
-
|
| 140 |
run_button.click(
|
| 141 |
fn = infer,
|
| 142 |
-
inputs = [prompt,
|
| 143 |
outputs = [result]
|
| 144 |
)
|
| 145 |
|
| 146 |
-
demo.queue().launch()
|
|
|
|
| 1 |
+
import spaces
|
| 2 |
import gradio as gr
|
| 3 |
import numpy as np
|
| 4 |
import random
|
| 5 |
+
import generation_sdxl
|
| 6 |
+
import functools
|
| 7 |
+
from diffusers import DiffusionPipeline, UNet2DConditionModel, StableDiffusionXLPipeline, DDIMScheduler
|
| 8 |
import torch
|
| 9 |
|
| 10 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 11 |
|
| 12 |
+
torch.cuda.max_memory_allocated(device=device)
|
| 13 |
+
model_id = 'stabilityai/stable-diffusion-xl-base-1.0'
|
| 14 |
+
pipe = StableDiffusionXLPipeline.from_pretrained(model_id,
|
| 15 |
+
torch_dtype=torch.float16,
|
| 16 |
+
scheduler=DDIMScheduler.from_pretrained(model_id, subfolder="scheduler"),
|
| 17 |
+
variant="fp16").to(device)
|
| 18 |
+
pipe = pipe.to(device)
|
| 19 |
+
unet = UNet2DConditionModel.from_pretrained("dbaranchuk/sdxl-cfg-distill-unet").to(device)
|
| 20 |
+
pipe.unet = unet
|
| 21 |
+
pipe.load_lora_weights("dbaranchuk/icd-lora-sdxl",
|
| 22 |
+
weight_name='reverse-249-499-699-999.safetensors')
|
| 23 |
+
pipe.fuse_lora()
|
| 24 |
+
pipe.to(dtype=torch.float16, device=device)
|
| 25 |
|
| 26 |
MAX_SEED = np.iinfo(np.int32).max
|
| 27 |
MAX_IMAGE_SIZE = 1024
|
| 28 |
|
| 29 |
+
@spaces.GPU(duration=30)
|
| 30 |
+
def infer(prompt, seed, randomize_seed, tau,
|
| 31 |
+
guidance_scale):
|
| 32 |
|
| 33 |
if randomize_seed:
|
| 34 |
seed = random.randint(0, MAX_SEED)
|
| 35 |
|
| 36 |
generator = torch.Generator().manual_seed(seed)
|
| 37 |
+
prompt = [prompt]
|
| 38 |
+
text_encoders = [pipe.text_encoder, pipe.text_encoder_2]
|
| 39 |
+
tokenizers = [pipe.tokenizer, pipe.tokenizer_2]
|
| 40 |
+
|
| 41 |
+
compute_embeddings_fn = functools.partial(
|
| 42 |
+
generation_sdxl.compute_embeddings,
|
| 43 |
+
proportion_empty_prompts=0,
|
| 44 |
+
text_encoders=text_encoders,
|
| 45 |
+
tokenizers=tokenizers,
|
| 46 |
+
)
|
| 47 |
+
|
| 48 |
+
if tau < 1.0:
|
| 49 |
+
use_dynamic_guidance=True
|
| 50 |
+
else:
|
| 51 |
+
use_dynamic_guidance=False
|
| 52 |
+
|
| 53 |
+
images = generation_sdxl.sample_deterministic(
|
| 54 |
+
pipe,
|
| 55 |
+
prompt,
|
| 56 |
+
num_inference_steps=4,
|
| 57 |
+
generator=generator,
|
| 58 |
+
guidance_scale=guidance_scale,
|
| 59 |
+
is_sdxl=True,
|
| 60 |
+
timesteps=[249, 499, 699, 999],
|
| 61 |
+
use_dynamic_guidance=use_dynamic_guidance,
|
| 62 |
+
tau1=tau,
|
| 63 |
+
tau2=tau,
|
| 64 |
+
compute_embeddings_fn=compute_embeddings_fn
|
| 65 |
+
)[0]
|
| 66 |
+
|
| 67 |
+
return images
|
| 68 |
|
| 69 |
examples = [
|
|
|
|
| 70 |
"An astronaut riding a green horse",
|
| 71 |
+
'Long-exposure night photography of a starry sky over a mountain range, with light trails.',
|
| 72 |
+
"Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
|
| 73 |
+
"A portrait of a girl with blonde, tousled hair, blue eyes",
|
| 74 |
]
|
| 75 |
|
| 76 |
css="""
|
|
|
|
| 88 |
with gr.Blocks(css=css) as demo:
|
| 89 |
|
| 90 |
with gr.Column(elem_id="col-container"):
|
| 91 |
+
gr.Markdown(
|
| 92 |
+
f"""
|
| 93 |
+
# β‘ Invertible Consistency Distillation β‘
|
| 94 |
+
# β‘ Image Generation with 4-step iCD-XL β‘
|
| 95 |
+
This is a demo of [Invertible Consistency Distillation](https://yandex-research.github.io/invertible-cd/),
|
| 96 |
+
a diffusion distillation method proposed in [Invertible Consistency Distillation for Text-Guided Image Editing in Around 7 Steps](https://arxiv.org/abs/2406.14539)
|
| 97 |
+
by [Yandex Research](https://github.com/yandex-research).
|
| 98 |
Currently running on {power_device}.
|
| 99 |
+
"""
|
| 100 |
+
)
|
| 101 |
+
gr.Markdown(
|
| 102 |
+
"If you enjoy the space, feel free to give a β to the <a href='https://github.com/yandex-research/invertible-cd' target='_blank'>Github Repo</a>. [](https://github.com/yandex-research/invertible-cd)"
|
| 103 |
+
)
|
| 104 |
+
|
| 105 |
with gr.Row():
|
| 106 |
|
| 107 |
prompt = gr.Text(
|
|
|
|
| 118 |
|
| 119 |
with gr.Accordion("Advanced Settings", open=False):
|
| 120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
seed = gr.Slider(
|
| 122 |
label="Seed",
|
| 123 |
minimum=0,
|
|
|
|
| 126 |
value=0,
|
| 127 |
)
|
| 128 |
|
| 129 |
+
randomize_seed = gr.Checkbox(label="Randomize seed", value=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
|
| 131 |
with gr.Row():
|
| 132 |
|
| 133 |
guidance_scale = gr.Slider(
|
| 134 |
label="Guidance scale",
|
| 135 |
minimum=0.0,
|
| 136 |
+
maximum=19.0,
|
| 137 |
+
step=1.0,
|
| 138 |
+
value=7.0,
|
| 139 |
)
|
| 140 |
|
| 141 |
+
dynamic_guidance_tau = gr.Slider(
|
| 142 |
+
label="Dynamic guidance tau",
|
| 143 |
+
minimum=0,
|
| 144 |
+
maximum=1,
|
| 145 |
+
step=0.1,
|
| 146 |
+
value=1.0,
|
| 147 |
)
|
| 148 |
|
| 149 |
gr.Examples(
|
| 150 |
examples = examples,
|
| 151 |
+
inputs = [prompt],
|
| 152 |
+
cache_examples=False
|
| 153 |
)
|
|
|
|
| 154 |
run_button.click(
|
| 155 |
fn = infer,
|
| 156 |
+
inputs = [prompt, seed, randomize_seed, dynamic_guidance_tau, guidance_scale],
|
| 157 |
outputs = [result]
|
| 158 |
)
|
| 159 |
|
| 160 |
+
demo.queue().launch(share=False)
|
generation_sdxl.py
ADDED
|
@@ -0,0 +1,474 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import copy
|
| 3 |
+
import random
|
| 4 |
+
import numpy as np
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
# Diffusion util
|
| 8 |
+
# ------------------------------------------------------------------------
|
| 9 |
+
def encode_prompt(prompt_batch, text_encoders, tokenizers, proportion_empty_prompts, is_train=True):
|
| 10 |
+
prompt_embeds_list = []
|
| 11 |
+
|
| 12 |
+
captions = []
|
| 13 |
+
for caption in prompt_batch:
|
| 14 |
+
if random.random() < proportion_empty_prompts:
|
| 15 |
+
captions.append("")
|
| 16 |
+
elif isinstance(caption, str):
|
| 17 |
+
captions.append(caption)
|
| 18 |
+
elif isinstance(caption, (list, np.ndarray)):
|
| 19 |
+
# take a random caption if there are multiple
|
| 20 |
+
captions.append(random.choice(caption) if is_train else caption[0])
|
| 21 |
+
|
| 22 |
+
with torch.no_grad():
|
| 23 |
+
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
|
| 24 |
+
text_inputs = tokenizer(
|
| 25 |
+
captions,
|
| 26 |
+
padding="max_length",
|
| 27 |
+
max_length=tokenizer.model_max_length,
|
| 28 |
+
truncation=True,
|
| 29 |
+
return_tensors="pt",
|
| 30 |
+
)
|
| 31 |
+
text_input_ids = text_inputs.input_ids
|
| 32 |
+
prompt_embeds = text_encoder(
|
| 33 |
+
text_input_ids.to(text_encoder.device),
|
| 34 |
+
output_hidden_states=True,
|
| 35 |
+
)
|
| 36 |
+
|
| 37 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
| 38 |
+
pooled_prompt_embeds = prompt_embeds[0]
|
| 39 |
+
prompt_embeds = prompt_embeds.hidden_states[-2]
|
| 40 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
| 41 |
+
prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
|
| 42 |
+
prompt_embeds_list.append(prompt_embeds)
|
| 43 |
+
|
| 44 |
+
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
| 45 |
+
pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
|
| 46 |
+
return prompt_embeds, pooled_prompt_embeds
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def compute_embeddings(
|
| 50 |
+
prompt_batch, original_sizes, crop_coords, proportion_empty_prompts, text_encoders, tokenizers, is_train=True,
|
| 51 |
+
device='cuda'
|
| 52 |
+
):
|
| 53 |
+
target_size = (1024, 1024)
|
| 54 |
+
original_sizes = original_sizes #list(map(list, zip(*original_sizes)))
|
| 55 |
+
crops_coords_top_left = crop_coords #list(map(list, zip(*crop_coords)))
|
| 56 |
+
|
| 57 |
+
original_sizes = torch.tensor(original_sizes, dtype=torch.long)
|
| 58 |
+
crops_coords_top_left = torch.tensor(crops_coords_top_left, dtype=torch.long)
|
| 59 |
+
|
| 60 |
+
prompt_embeds, pooled_prompt_embeds = encode_prompt(
|
| 61 |
+
prompt_batch, text_encoders, tokenizers, proportion_empty_prompts, is_train
|
| 62 |
+
)
|
| 63 |
+
add_text_embeds = pooled_prompt_embeds
|
| 64 |
+
|
| 65 |
+
# Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
|
| 66 |
+
add_time_ids = list(target_size)
|
| 67 |
+
add_time_ids = torch.tensor([add_time_ids])
|
| 68 |
+
add_time_ids = add_time_ids.repeat(len(prompt_batch), 1)
|
| 69 |
+
add_time_ids = torch.cat([original_sizes, crops_coords_top_left, add_time_ids], dim=-1)
|
| 70 |
+
add_time_ids = add_time_ids.to(device, dtype=prompt_embeds.dtype)
|
| 71 |
+
|
| 72 |
+
prompt_embeds = prompt_embeds.to(device)
|
| 73 |
+
add_text_embeds = add_text_embeds.to(device)
|
| 74 |
+
unet_added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
| 75 |
+
|
| 76 |
+
return {"prompt_embeds": prompt_embeds, **unet_added_cond_kwargs}
|
| 77 |
+
|
| 78 |
+
def extract_into_tensor(a, t, x_shape):
|
| 79 |
+
b, *_ = t.shape
|
| 80 |
+
out = a.gather(-1, t)
|
| 81 |
+
return out.reshape(b, *((1,) * (len(x_shape) - 1)))
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
def guidance_scale_embedding(w, embedding_dim=512, dtype=torch.float32):
|
| 85 |
+
"""
|
| 86 |
+
See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
|
| 87 |
+
|
| 88 |
+
Args:
|
| 89 |
+
timesteps (`torch.Tensor`):
|
| 90 |
+
generate embedding vectors at these timesteps
|
| 91 |
+
embedding_dim (`int`, *optional*, defaults to 512):
|
| 92 |
+
dimension of the embeddings to generate
|
| 93 |
+
dtype:
|
| 94 |
+
data type of the generated embeddings
|
| 95 |
+
|
| 96 |
+
Returns:
|
| 97 |
+
`torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
|
| 98 |
+
"""
|
| 99 |
+
assert len(w.shape) == 1
|
| 100 |
+
w = w * 1000.0
|
| 101 |
+
|
| 102 |
+
half_dim = embedding_dim // 2
|
| 103 |
+
emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
|
| 104 |
+
emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
|
| 105 |
+
emb = w.to(dtype)[:, None] * emb[None, :]
|
| 106 |
+
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
|
| 107 |
+
if embedding_dim % 2 == 1: # zero pad
|
| 108 |
+
emb = torch.nn.functional.pad(emb, (0, 1))
|
| 109 |
+
assert emb.shape == (w.shape[0], embedding_dim)
|
| 110 |
+
return emb
|
| 111 |
+
|
| 112 |
+
def predicted_origin(model_output, timesteps, boundary_timesteps, sample, prediction_type, alphas, sigmas):
|
| 113 |
+
sigmas_s = extract_into_tensor(sigmas, boundary_timesteps, sample.shape)
|
| 114 |
+
alphas_s = extract_into_tensor(alphas, boundary_timesteps, sample.shape)
|
| 115 |
+
|
| 116 |
+
sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
|
| 117 |
+
alphas = extract_into_tensor(alphas, timesteps, sample.shape)
|
| 118 |
+
|
| 119 |
+
# Set hard boundaries to ensure equivalence with forward (direct) CD
|
| 120 |
+
alphas_s[boundary_timesteps == 0] = 1.0
|
| 121 |
+
sigmas_s[boundary_timesteps == 0] = 0.0
|
| 122 |
+
|
| 123 |
+
if prediction_type == "epsilon":
|
| 124 |
+
pred_x_0 = (sample - sigmas * model_output) / alphas # x0 prediction
|
| 125 |
+
pred_x_0 = alphas_s * pred_x_0 + sigmas_s * model_output # Euler step to the boundary step
|
| 126 |
+
elif prediction_type == "v_prediction":
|
| 127 |
+
assert boundary_timesteps == 0, "v_prediction does not support multiple endpoints at the moment"
|
| 128 |
+
pred_x_0 = alphas * sample - sigmas * model_output
|
| 129 |
+
else:
|
| 130 |
+
raise ValueError(f"Prediction type {prediction_type} currently not supported.")
|
| 131 |
+
|
| 132 |
+
return pred_x_0
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
class DDIMSolver:
|
| 136 |
+
def __init__(
|
| 137 |
+
self, alpha_cumprods, timesteps=1000, ddim_timesteps=50,
|
| 138 |
+
num_endpoints=1, num_inverse_endpoints=1,
|
| 139 |
+
max_inverse_timestep_index=49,
|
| 140 |
+
endpoints=None, inverse_endpoints=None
|
| 141 |
+
):
|
| 142 |
+
# DDIM sampling parameters
|
| 143 |
+
step_ratio = timesteps // ddim_timesteps
|
| 144 |
+
self.ddim_timesteps = (np.arange(1, ddim_timesteps + 1) * step_ratio).round().astype(
|
| 145 |
+
np.int64) - 1 # [19, ..., 999]
|
| 146 |
+
self.ddim_alpha_cumprods = alpha_cumprods[self.ddim_timesteps]
|
| 147 |
+
self.ddim_alpha_cumprods_prev = np.asarray(
|
| 148 |
+
[alpha_cumprods[0]] + alpha_cumprods[self.ddim_timesteps[:-1]].tolist()
|
| 149 |
+
)
|
| 150 |
+
self.ddim_alpha_cumprods_next = np.asarray(
|
| 151 |
+
alpha_cumprods[self.ddim_timesteps[1:]].tolist() + [0.0]
|
| 152 |
+
)
|
| 153 |
+
# convert to torch tensors
|
| 154 |
+
self.ddim_timesteps = torch.from_numpy(self.ddim_timesteps).long()
|
| 155 |
+
self.ddim_alpha_cumprods = torch.from_numpy(self.ddim_alpha_cumprods)
|
| 156 |
+
self.ddim_alpha_cumprods_prev = torch.from_numpy(self.ddim_alpha_cumprods_prev)
|
| 157 |
+
self.ddim_alpha_cumprods_next = torch.from_numpy(self.ddim_alpha_cumprods_next)
|
| 158 |
+
|
| 159 |
+
# Set endpoints for direct CTM
|
| 160 |
+
if endpoints is None:
|
| 161 |
+
timestep_interval = ddim_timesteps // num_endpoints + int(ddim_timesteps % num_endpoints > 0)
|
| 162 |
+
endpoint_idxs = torch.arange(timestep_interval, ddim_timesteps, timestep_interval) - 1
|
| 163 |
+
self.endpoints = torch.tensor([0] + self.ddim_timesteps[endpoint_idxs].tolist())
|
| 164 |
+
else:
|
| 165 |
+
self.endpoints = torch.tensor([int(endpoint) for endpoint in endpoints.split(',')])
|
| 166 |
+
assert len(self.endpoints) == num_endpoints
|
| 167 |
+
|
| 168 |
+
# Set endpoints for inverse CTM
|
| 169 |
+
if inverse_endpoints is None:
|
| 170 |
+
timestep_interval = ddim_timesteps // num_inverse_endpoints + int(
|
| 171 |
+
ddim_timesteps % num_inverse_endpoints > 0)
|
| 172 |
+
inverse_endpoint_idxs = torch.arange(timestep_interval, ddim_timesteps, timestep_interval) - 1
|
| 173 |
+
inverse_endpoint_idxs = torch.tensor(inverse_endpoint_idxs.tolist() + [max_inverse_timestep_index])
|
| 174 |
+
self.inverse_endpoints = self.ddim_timesteps[inverse_endpoint_idxs]
|
| 175 |
+
else:
|
| 176 |
+
self.inverse_endpoints = torch.tensor([int(endpoint) for endpoint in inverse_endpoints.split(',')])
|
| 177 |
+
assert len(self.inverse_endpoints) == num_inverse_endpoints
|
| 178 |
+
|
| 179 |
+
def to(self, device):
|
| 180 |
+
self.endpoints = self.endpoints.to(device)
|
| 181 |
+
self.inverse_endpoints = self.inverse_endpoints.to(device)
|
| 182 |
+
|
| 183 |
+
self.ddim_timesteps = self.ddim_timesteps.to(device)
|
| 184 |
+
self.ddim_alpha_cumprods = self.ddim_alpha_cumprods.to(device)
|
| 185 |
+
self.ddim_alpha_cumprods_prev = self.ddim_alpha_cumprods_prev.to(device)
|
| 186 |
+
self.ddim_alpha_cumprods_next = self.ddim_alpha_cumprods_next.to(device)
|
| 187 |
+
return self
|
| 188 |
+
|
| 189 |
+
def ddim_step(self, pred_x0, pred_noise, timestep_index):
|
| 190 |
+
alpha_cumprod_prev = extract_into_tensor(self.ddim_alpha_cumprods_prev, timestep_index, pred_x0.shape)
|
| 191 |
+
dir_xt = (1.0 - alpha_cumprod_prev).sqrt() * pred_noise
|
| 192 |
+
x_prev = alpha_cumprod_prev.sqrt() * pred_x0 + dir_xt
|
| 193 |
+
return x_prev
|
| 194 |
+
|
| 195 |
+
def inverse_ddim_step(self, pred_x0, pred_noise, timestep_index):
|
| 196 |
+
alpha_cumprod_next = extract_into_tensor(self.ddim_alpha_cumprods_next, timestep_index, pred_x0.shape)
|
| 197 |
+
dir_xt = (1.0 - alpha_cumprod_next).sqrt() * pred_noise
|
| 198 |
+
x_next = alpha_cumprod_next.sqrt() * pred_x0 + dir_xt
|
| 199 |
+
return x_next
|
| 200 |
+
# ------------------------------------------------------------------------
|
| 201 |
+
|
| 202 |
+
# Distillation specific
|
| 203 |
+
# ------------------------------------------------------------------------
|
| 204 |
+
def inverse_sample_deterministic(
|
| 205 |
+
pipe,
|
| 206 |
+
images,
|
| 207 |
+
prompt,
|
| 208 |
+
generator=None,
|
| 209 |
+
num_scales=50,
|
| 210 |
+
num_inference_steps=1,
|
| 211 |
+
timesteps=None,
|
| 212 |
+
start_timestep=19,
|
| 213 |
+
max_inverse_timestep_index=49,
|
| 214 |
+
return_start_latent=False,
|
| 215 |
+
guidance_scale=None, # Used only if the student has w_embedding
|
| 216 |
+
compute_embeddings_fn=None,
|
| 217 |
+
is_sdxl=False,
|
| 218 |
+
inverse_endpoints=None,
|
| 219 |
+
seed=0,
|
| 220 |
+
):
|
| 221 |
+
# assert isinstance(pipe, StableDiffusionImg2ImgPipeline), f"Does not support the pipeline {type(pipe)}"
|
| 222 |
+
|
| 223 |
+
if prompt is not None and isinstance(prompt, str):
|
| 224 |
+
batch_size = 1
|
| 225 |
+
elif prompt is not None and isinstance(prompt, list):
|
| 226 |
+
batch_size = len(prompt)
|
| 227 |
+
|
| 228 |
+
device = pipe._execution_device
|
| 229 |
+
|
| 230 |
+
# Prepare text embeddings
|
| 231 |
+
if compute_embeddings_fn is not None:
|
| 232 |
+
if is_sdxl:
|
| 233 |
+
orig_size = [(1024, 1024)] * len(prompt)
|
| 234 |
+
crop_coords = [(0, 0)] * len(prompt)
|
| 235 |
+
encoded_text = compute_embeddings_fn(prompt, orig_size, crop_coords)
|
| 236 |
+
prompt_embeds = encoded_text.pop("prompt_embeds")
|
| 237 |
+
else:
|
| 238 |
+
prompt_embeds = compute_embeddings_fn(prompt)["prompt_embeds"]
|
| 239 |
+
encoded_text = {}
|
| 240 |
+
prompt_embeds = prompt_embeds.to(pipe.unet.dtype)
|
| 241 |
+
else:
|
| 242 |
+
prompt_embeds = pipe.encode_prompt(prompt, device, 1, False)[0]
|
| 243 |
+
encoded_text = {}
|
| 244 |
+
assert prompt_embeds.dtype == pipe.unet.dtype
|
| 245 |
+
|
| 246 |
+
# Prepare the DDIM solver
|
| 247 |
+
endpoints = ','.join(['0'] + inverse_endpoints.split(',')[:-1]) if inverse_endpoints is not None else None
|
| 248 |
+
solver = DDIMSolver(
|
| 249 |
+
pipe.scheduler.alphas_cumprod.cpu().numpy(),
|
| 250 |
+
timesteps=pipe.scheduler.num_train_timesteps,
|
| 251 |
+
ddim_timesteps=num_scales,
|
| 252 |
+
num_endpoints=num_inference_steps,
|
| 253 |
+
num_inverse_endpoints=num_inference_steps,
|
| 254 |
+
max_inverse_timestep_index=max_inverse_timestep_index,
|
| 255 |
+
endpoints=endpoints,
|
| 256 |
+
inverse_endpoints=inverse_endpoints
|
| 257 |
+
).to(device)
|
| 258 |
+
|
| 259 |
+
if timesteps is None:
|
| 260 |
+
timesteps = solver.inverse_endpoints.flip(0)
|
| 261 |
+
boundary_timesteps = solver.endpoints.flip(0)
|
| 262 |
+
else:
|
| 263 |
+
timesteps, boundary_timesteps = timesteps, timesteps
|
| 264 |
+
boundary_timesteps = boundary_timesteps[1:] + [boundary_timesteps[0]]
|
| 265 |
+
boundary_timesteps[-1] = 999
|
| 266 |
+
timesteps, boundary_timesteps = torch.tensor(timesteps), torch.tensor(boundary_timesteps)
|
| 267 |
+
|
| 268 |
+
alpha_schedule = torch.sqrt(pipe.scheduler.alphas_cumprod).to(device)
|
| 269 |
+
sigma_schedule = torch.sqrt(1 - pipe.scheduler.alphas_cumprod).to(device)
|
| 270 |
+
|
| 271 |
+
# 5. Prepare latent variables
|
| 272 |
+
num_channels_latents = pipe.unet.config.in_channels
|
| 273 |
+
start_latents = pipe.prepare_latents(
|
| 274 |
+
images, timesteps[0], batch_size, 1, prompt_embeds.dtype, device,
|
| 275 |
+
generator=torch.Generator().manual_seed(seed),
|
| 276 |
+
)
|
| 277 |
+
latents = start_latents.clone()
|
| 278 |
+
|
| 279 |
+
if guidance_scale is not None:
|
| 280 |
+
w = torch.ones(batch_size) * guidance_scale
|
| 281 |
+
w_embedding = guidance_scale_embedding(w, embedding_dim=512)
|
| 282 |
+
w_embedding = w_embedding.to(device=latents.device, dtype=latents.dtype)
|
| 283 |
+
else:
|
| 284 |
+
w_embedding = None
|
| 285 |
+
|
| 286 |
+
for i, (t, s) in enumerate(zip(timesteps, boundary_timesteps)):
|
| 287 |
+
# predict the noise residual
|
| 288 |
+
noise_pred = pipe.unet(
|
| 289 |
+
latents.to(prompt_embeds.dtype),
|
| 290 |
+
t,
|
| 291 |
+
encoder_hidden_states=prompt_embeds,
|
| 292 |
+
return_dict=False,
|
| 293 |
+
timestep_cond=w_embedding,
|
| 294 |
+
added_cond_kwargs=encoded_text,
|
| 295 |
+
)[0]
|
| 296 |
+
|
| 297 |
+
latents = predicted_origin(
|
| 298 |
+
noise_pred,
|
| 299 |
+
torch.tensor([t] * len(latents), device=device),
|
| 300 |
+
torch.tensor([s] * len(latents), device=device),
|
| 301 |
+
latents,
|
| 302 |
+
pipe.scheduler.config.prediction_type,
|
| 303 |
+
alpha_schedule,
|
| 304 |
+
sigma_schedule,
|
| 305 |
+
).to(prompt_embeds.dtype)
|
| 306 |
+
|
| 307 |
+
if return_start_latent:
|
| 308 |
+
return latents, start_latents
|
| 309 |
+
else:
|
| 310 |
+
return latents
|
| 311 |
+
|
| 312 |
+
|
| 313 |
+
def linear_schedule_old(t, guidance_scale, tau1, tau2):
|
| 314 |
+
t = t / 1000
|
| 315 |
+
if t <= tau1:
|
| 316 |
+
gamma = 1.0
|
| 317 |
+
elif t >= tau2:
|
| 318 |
+
gamma = 0.0
|
| 319 |
+
else:
|
| 320 |
+
gamma = (tau2 - t) / (tau2 - tau1)
|
| 321 |
+
return gamma * guidance_scale
|
| 322 |
+
|
| 323 |
+
|
| 324 |
+
@torch.no_grad()
|
| 325 |
+
def sample_deterministic(
|
| 326 |
+
pipe,
|
| 327 |
+
prompt,
|
| 328 |
+
latents=None,
|
| 329 |
+
generator=None,
|
| 330 |
+
num_scales=50,
|
| 331 |
+
num_inference_steps=1,
|
| 332 |
+
timesteps=None,
|
| 333 |
+
start_timestep=19,
|
| 334 |
+
max_inverse_timestep_index=49,
|
| 335 |
+
return_latent=False,
|
| 336 |
+
guidance_scale=None, # Used only if the student has w_embedding
|
| 337 |
+
compute_embeddings_fn=None,
|
| 338 |
+
is_sdxl=False,
|
| 339 |
+
endpoints=None,
|
| 340 |
+
use_dynamic_guidance=False,
|
| 341 |
+
tau1=0.7,
|
| 342 |
+
tau2=0.7,
|
| 343 |
+
amplify_prompt=None,
|
| 344 |
+
):
|
| 345 |
+
# assert isinstance(pipe, StableDiffusionPipeline), f"Does not support the pipeline {type(pipe)}"
|
| 346 |
+
height = pipe.unet.config.sample_size * pipe.vae_scale_factor
|
| 347 |
+
width = pipe.unet.config.sample_size * pipe.vae_scale_factor
|
| 348 |
+
|
| 349 |
+
# 1. Define call parameters
|
| 350 |
+
if prompt is not None and isinstance(prompt, str):
|
| 351 |
+
batch_size = 1
|
| 352 |
+
elif prompt is not None and isinstance(prompt, list):
|
| 353 |
+
batch_size = len(prompt)
|
| 354 |
+
|
| 355 |
+
device = pipe._execution_device
|
| 356 |
+
|
| 357 |
+
# Prepare text embeddings
|
| 358 |
+
if compute_embeddings_fn is not None:
|
| 359 |
+
if is_sdxl:
|
| 360 |
+
orig_size = [(1024, 1024)] * len(prompt)
|
| 361 |
+
crop_coords = [(0, 0)] * len(prompt)
|
| 362 |
+
encoded_text = compute_embeddings_fn(prompt, orig_size, crop_coords)
|
| 363 |
+
prompt_embeds = encoded_text.pop("prompt_embeds")
|
| 364 |
+
if amplify_prompt is not None:
|
| 365 |
+
orig_size = [(1024, 1024)] * len(amplify_prompt)
|
| 366 |
+
crop_coords = [(0, 0)] * len(amplify_prompt)
|
| 367 |
+
encoded_text_old = compute_embeddings_fn(amplify_prompt, orig_size, crop_coords)
|
| 368 |
+
amplify_prompt_embeds = encoded_text_old.pop("prompt_embeds")
|
| 369 |
+
else:
|
| 370 |
+
prompt_embeds = compute_embeddings_fn(prompt)["prompt_embeds"]
|
| 371 |
+
encoded_text = {}
|
| 372 |
+
prompt_embeds = prompt_embeds.to(pipe.unet.dtype)
|
| 373 |
+
else:
|
| 374 |
+
prompt_embeds = pipe.encode_prompt(prompt, device, 1, False)[0]
|
| 375 |
+
encoded_text = {}
|
| 376 |
+
assert prompt_embeds.dtype == pipe.unet.dtype
|
| 377 |
+
|
| 378 |
+
# Prepare the DDIM solver
|
| 379 |
+
inverse_endpoints = ','.join(endpoints.split(',')[1:] + ['999']) if endpoints is not None else None
|
| 380 |
+
solver = DDIMSolver(
|
| 381 |
+
pipe.scheduler.alphas_cumprod.numpy(),
|
| 382 |
+
timesteps=pipe.scheduler.num_train_timesteps,
|
| 383 |
+
ddim_timesteps=num_scales,
|
| 384 |
+
num_endpoints=num_inference_steps,
|
| 385 |
+
num_inverse_endpoints=num_inference_steps,
|
| 386 |
+
max_inverse_timestep_index=max_inverse_timestep_index,
|
| 387 |
+
endpoints=endpoints,
|
| 388 |
+
inverse_endpoints=inverse_endpoints
|
| 389 |
+
).to(device)
|
| 390 |
+
|
| 391 |
+
prompt_embeds_init = copy.deepcopy(prompt_embeds)
|
| 392 |
+
|
| 393 |
+
if timesteps is None:
|
| 394 |
+
timesteps = solver.inverse_endpoints.flip(0)
|
| 395 |
+
boundary_timesteps = solver.endpoints.flip(0)
|
| 396 |
+
else:
|
| 397 |
+
timesteps, boundary_timesteps = copy.deepcopy(timesteps), copy.deepcopy(timesteps)
|
| 398 |
+
timesteps.reverse()
|
| 399 |
+
boundary_timesteps.reverse()
|
| 400 |
+
boundary_timesteps = boundary_timesteps[1:] + [boundary_timesteps[0]]
|
| 401 |
+
boundary_timesteps[-1] = 0
|
| 402 |
+
timesteps, boundary_timesteps = torch.tensor(timesteps), torch.tensor(boundary_timesteps)
|
| 403 |
+
|
| 404 |
+
alpha_schedule = torch.sqrt(pipe.scheduler.alphas_cumprod).to(device)
|
| 405 |
+
sigma_schedule = torch.sqrt(1 - pipe.scheduler.alphas_cumprod).to(device)
|
| 406 |
+
|
| 407 |
+
# 5. Prepare latent variables
|
| 408 |
+
if latents is None:
|
| 409 |
+
num_channels_latents = pipe.unet.config.in_channels
|
| 410 |
+
latents = pipe.prepare_latents(
|
| 411 |
+
batch_size,
|
| 412 |
+
num_channels_latents,
|
| 413 |
+
height,
|
| 414 |
+
width,
|
| 415 |
+
prompt_embeds.dtype,
|
| 416 |
+
device,
|
| 417 |
+
generator,
|
| 418 |
+
None,
|
| 419 |
+
)
|
| 420 |
+
assert latents.dtype == pipe.unet.dtype
|
| 421 |
+
else:
|
| 422 |
+
latents = latents.to(prompt_embeds.dtype)
|
| 423 |
+
|
| 424 |
+
if guidance_scale is not None:
|
| 425 |
+
w = torch.ones(batch_size) * guidance_scale
|
| 426 |
+
w_embedding = guidance_scale_embedding(w, embedding_dim=512)
|
| 427 |
+
w_embedding = w_embedding.to(device=latents.device, dtype=latents.dtype)
|
| 428 |
+
else:
|
| 429 |
+
w_embedding = None
|
| 430 |
+
|
| 431 |
+
for i, (t, s) in enumerate(zip(timesteps, boundary_timesteps)):
|
| 432 |
+
if use_dynamic_guidance:
|
| 433 |
+
if not isinstance(t, int):
|
| 434 |
+
t_item = t.item()
|
| 435 |
+
if t_item > tau1 * 1000 and amplify_prompt is not None:
|
| 436 |
+
prompt_embeds = amplify_prompt_embeds
|
| 437 |
+
else:
|
| 438 |
+
prompt_embeds = prompt_embeds_init
|
| 439 |
+
guidance_scale = linear_schedule_old(t_item, w, tau1=tau1, tau2=tau2)
|
| 440 |
+
guidance_scale_tensor = torch.tensor([guidance_scale] * len(latents))
|
| 441 |
+
w_embedding = guidance_scale_embedding(guidance_scale_tensor, embedding_dim=512)
|
| 442 |
+
w_embedding = w_embedding.to(device=latents.device, dtype=latents.dtype)
|
| 443 |
+
|
| 444 |
+
# predict the noise residual
|
| 445 |
+
noise_pred = pipe.unet(
|
| 446 |
+
latents,
|
| 447 |
+
t,
|
| 448 |
+
encoder_hidden_states=prompt_embeds,
|
| 449 |
+
cross_attention_kwargs=None,
|
| 450 |
+
return_dict=False,
|
| 451 |
+
timestep_cond=w_embedding,
|
| 452 |
+
added_cond_kwargs=encoded_text,
|
| 453 |
+
)[0]
|
| 454 |
+
|
| 455 |
+
latents = predicted_origin(
|
| 456 |
+
noise_pred,
|
| 457 |
+
torch.tensor([t] * len(noise_pred)).to(device),
|
| 458 |
+
torch.tensor([s] * len(noise_pred)).to(device),
|
| 459 |
+
latents,
|
| 460 |
+
pipe.scheduler.config.prediction_type,
|
| 461 |
+
alpha_schedule,
|
| 462 |
+
sigma_schedule,
|
| 463 |
+
).to(pipe.unet.dtype)
|
| 464 |
+
|
| 465 |
+
pipe.vae.to(torch.float32)
|
| 466 |
+
image = pipe.vae.decode(latents.to(torch.float32) / pipe.vae.config.scaling_factor, return_dict=False)[0]
|
| 467 |
+
do_denormalize = [True] * image.shape[0]
|
| 468 |
+
image = pipe.image_processor.postprocess(image, output_type="pil", do_denormalize=do_denormalize)
|
| 469 |
+
|
| 470 |
+
if return_latent:
|
| 471 |
+
return image, latents
|
| 472 |
+
else:
|
| 473 |
+
return image
|
| 474 |
+
# ------------------------------------------------------------------------
|
requirements.txt
CHANGED
|
@@ -1,6 +1,7 @@
|
|
| 1 |
accelerate
|
| 2 |
diffusers
|
| 3 |
invisible_watermark
|
|
|
|
| 4 |
torch
|
| 5 |
transformers
|
| 6 |
xformers
|
|
|
|
| 1 |
accelerate
|
| 2 |
diffusers
|
| 3 |
invisible_watermark
|
| 4 |
+
peft
|
| 5 |
torch
|
| 6 |
transformers
|
| 7 |
xformers
|