Spaces:
Runtime error
Runtime error
Jordan Legg
commited on
Commit
Β·
1c4aefd
1
Parent(s):
3be64a5
removed img2img :(
Browse files
app.py
CHANGED
|
@@ -1,145 +1,61 @@
|
|
| 1 |
import spaces
|
| 2 |
import gradio as gr
|
| 3 |
-
import numpy as np
|
| 4 |
-
import random
|
| 5 |
import torch
|
| 6 |
-
import torch.nn as nn
|
| 7 |
-
from PIL import Image
|
| 8 |
-
from torchvision import transforms
|
| 9 |
from diffusers import DiffusionPipeline
|
| 10 |
|
| 11 |
# Constants
|
| 12 |
-
|
| 13 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 14 |
-
MAX_SEED = np.iinfo(np.int32).max
|
| 15 |
MAX_IMAGE_SIZE = 2048
|
| 16 |
-
LATENT_CHANNELS = 16
|
| 17 |
-
TEXT_EMBED_DIM = 768
|
| 18 |
-
MAX_TEXT_EMBEDDINGS = 77
|
| 19 |
-
SCALING_FACTOR = 0.3611
|
| 20 |
|
| 21 |
# Load FLUX model
|
| 22 |
-
pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=
|
|
|
|
| 23 |
pipe.enable_model_cpu_offload()
|
| 24 |
pipe.vae.enable_slicing()
|
| 25 |
pipe.vae.enable_tiling()
|
| 26 |
|
| 27 |
-
# Add a projection layer to match text embedding dimension
|
| 28 |
-
projection = nn.Linear(LATENT_CHANNELS, TEXT_EMBED_DIM).to(device).to(dtype)
|
| 29 |
-
|
| 30 |
-
def preprocess_image(image, image_size):
|
| 31 |
-
preprocess = transforms.Compose([
|
| 32 |
-
transforms.Resize((image_size, image_size), interpolation=transforms.InterpolationMode.LANCZOS),
|
| 33 |
-
transforms.ToTensor(),
|
| 34 |
-
transforms.Normalize([0.5], [0.5])
|
| 35 |
-
])
|
| 36 |
-
image = preprocess(image).unsqueeze(0).to(device, dtype=dtype)
|
| 37 |
-
return image
|
| 38 |
-
|
| 39 |
-
def process_latents(latents, height, width):
|
| 40 |
-
print(f"Input latent shape: {latents.shape}")
|
| 41 |
-
|
| 42 |
-
# Ensure latents are the correct shape
|
| 43 |
-
if latents.shape[2:] != (height // 8, width // 8):
|
| 44 |
-
latents = torch.nn.functional.interpolate(latents, size=(height // 8, width // 8), mode='bilinear')
|
| 45 |
-
print(f"Latent shape after potential interpolation: {latents.shape}")
|
| 46 |
-
|
| 47 |
-
# Reshape latents to [batch_size, seq_len, channels]
|
| 48 |
-
latents = latents.permute(0, 2, 3, 1).reshape(1, -1, LATENT_CHANNELS)
|
| 49 |
-
print(f"Reshaped latent shape: {latents.shape}")
|
| 50 |
-
|
| 51 |
-
# Project latents to match text embedding dimension
|
| 52 |
-
latents = projection(latents)
|
| 53 |
-
print(f"Projected latent shape: {latents.shape}")
|
| 54 |
-
|
| 55 |
-
# Adjust sequence length to match text embeddings
|
| 56 |
-
seq_len = latents.shape[1]
|
| 57 |
-
if seq_len > MAX_TEXT_EMBEDDINGS:
|
| 58 |
-
latents = latents[:, :MAX_TEXT_EMBEDDINGS, :]
|
| 59 |
-
elif seq_len < MAX_TEXT_EMBEDDINGS:
|
| 60 |
-
pad_len = MAX_TEXT_EMBEDDINGS - seq_len
|
| 61 |
-
latents = torch.nn.functional.pad(latents, (0, 0, 0, pad_len, 0, 0))
|
| 62 |
-
print(f"Final latent shape: {latents.shape}")
|
| 63 |
-
|
| 64 |
-
return latents
|
| 65 |
-
|
| 66 |
@spaces.GPU()
|
| 67 |
-
def
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
generator = torch.Generator(device=device).manual_seed(seed)
|
| 71 |
-
|
| 72 |
try:
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
).images[0]
|
| 83 |
-
else:
|
| 84 |
-
# img2img case
|
| 85 |
-
init_image = init_image.convert("RGB")
|
| 86 |
-
init_image = preprocess_image(init_image, 1024) # Using 1024 as FLUX VAE sample size
|
| 87 |
-
|
| 88 |
-
# Encode the image using FLUX VAE
|
| 89 |
-
latents = pipe.vae.encode(init_image).latent_dist.sample() * SCALING_FACTOR
|
| 90 |
-
print(f"Initial latent shape from VAE: {latents.shape}")
|
| 91 |
-
|
| 92 |
-
# Process latents to match text embedding format
|
| 93 |
-
latents = process_latents(latents, height, width)
|
| 94 |
-
|
| 95 |
-
# Get text embeddings
|
| 96 |
-
text_embeddings = pipe.transformer.text_encoder([prompt])
|
| 97 |
-
print(f"Text embedding shape: {text_embeddings.shape}")
|
| 98 |
-
|
| 99 |
-
# Combine image latents and text embeddings
|
| 100 |
-
combined_embeddings = torch.cat([latents, text_embeddings], dim=1)
|
| 101 |
-
print(f"Combined embedding shape: {combined_embeddings.shape}")
|
| 102 |
-
|
| 103 |
-
image = pipe(
|
| 104 |
-
prompt=prompt,
|
| 105 |
-
height=height,
|
| 106 |
-
width=width,
|
| 107 |
-
num_inference_steps=num_inference_steps,
|
| 108 |
-
generator=generator,
|
| 109 |
-
guidance_scale=0.0,
|
| 110 |
-
latents=combined_embeddings
|
| 111 |
-
).images[0]
|
| 112 |
-
|
| 113 |
return image, seed
|
| 114 |
except Exception as e:
|
| 115 |
-
print(f"Error during
|
| 116 |
import traceback
|
| 117 |
traceback.print_exc()
|
| 118 |
-
return
|
| 119 |
|
| 120 |
-
# Gradio interface
|
| 121 |
with gr.Blocks() as demo:
|
| 122 |
with gr.Row():
|
| 123 |
prompt = gr.Textbox(label="Prompt")
|
| 124 |
-
|
| 125 |
-
|
| 126 |
with gr.Row():
|
| 127 |
generate = gr.Button("Generate")
|
| 128 |
-
|
| 129 |
with gr.Row():
|
| 130 |
-
result = gr.Image(label="
|
| 131 |
-
seed_output = gr.Number(label="Seed")
|
| 132 |
-
|
| 133 |
with gr.Accordion("Advanced Settings", open=False):
|
| 134 |
-
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42)
|
| 135 |
-
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
| 136 |
width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
|
| 137 |
height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
|
| 138 |
num_inference_steps = gr.Slider(label="Number of inference steps", minimum=1, maximum=50, step=1, value=4)
|
| 139 |
|
| 140 |
generate.click(
|
| 141 |
-
|
| 142 |
-
inputs=[prompt,
|
| 143 |
outputs=[result, seed_output]
|
| 144 |
)
|
| 145 |
|
|
|
|
| 1 |
import spaces
|
| 2 |
import gradio as gr
|
|
|
|
|
|
|
| 3 |
import torch
|
|
|
|
|
|
|
|
|
|
| 4 |
from diffusers import DiffusionPipeline
|
| 5 |
|
| 6 |
# Constants
|
| 7 |
+
MAX_SEED = 2**32 - 1
|
|
|
|
|
|
|
| 8 |
MAX_IMAGE_SIZE = 2048
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
# Load FLUX model
|
| 11 |
+
pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
|
| 12 |
+
pipe = pipe.to("cuda")
|
| 13 |
pipe.enable_model_cpu_offload()
|
| 14 |
pipe.vae.enable_slicing()
|
| 15 |
pipe.vae.enable_tiling()
|
| 16 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
@spaces.GPU()
|
| 18 |
+
def generate_image(prompt, seed, width, height, num_inference_steps):
|
| 19 |
+
generator = torch.Generator(device="cuda").manual_seed(seed)
|
| 20 |
+
|
|
|
|
|
|
|
| 21 |
try:
|
| 22 |
+
image = pipe(
|
| 23 |
+
prompt=prompt,
|
| 24 |
+
height=height,
|
| 25 |
+
width=width,
|
| 26 |
+
num_inference_steps=num_inference_steps,
|
| 27 |
+
generator=generator,
|
| 28 |
+
guidance_scale=0.0
|
| 29 |
+
).images[0]
|
| 30 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
return image, seed
|
| 32 |
except Exception as e:
|
| 33 |
+
print(f"Error during image generation: {e}")
|
| 34 |
import traceback
|
| 35 |
traceback.print_exc()
|
| 36 |
+
return None, seed
|
| 37 |
|
| 38 |
+
# Gradio interface
|
| 39 |
with gr.Blocks() as demo:
|
| 40 |
with gr.Row():
|
| 41 |
prompt = gr.Textbox(label="Prompt")
|
| 42 |
+
|
|
|
|
| 43 |
with gr.Row():
|
| 44 |
generate = gr.Button("Generate")
|
| 45 |
+
|
| 46 |
with gr.Row():
|
| 47 |
+
result = gr.Image(label="Generated Image")
|
| 48 |
+
seed_output = gr.Number(label="Seed Used")
|
| 49 |
+
|
| 50 |
with gr.Accordion("Advanced Settings", open=False):
|
| 51 |
+
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, randomize=True)
|
|
|
|
| 52 |
width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
|
| 53 |
height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
|
| 54 |
num_inference_steps = gr.Slider(label="Number of inference steps", minimum=1, maximum=50, step=1, value=4)
|
| 55 |
|
| 56 |
generate.click(
|
| 57 |
+
generate_image,
|
| 58 |
+
inputs=[prompt, seed, width, height, num_inference_steps],
|
| 59 |
outputs=[result, seed_output]
|
| 60 |
)
|
| 61 |
|