import torch
from diffusers import StableDiffusion3Pipeline, StableDiffusionPipeline, StableDiffusionXLPipeline, DPMSolverSinglestepScheduler, StableCascadePriorPipeline, StableCascadeDecoderPipeline
import gradio as gr
import os
import random
import numpy as np
from PIL import Image
import spaces
HF_TOKEN = os.getenv("HF_TOKEN") # login with hf token to access sd gated models
if torch.cuda.is_available():
device = "cuda"
print("Using GPU")
else:
device = "cpu"
print("Using CPU")
MAX_SEED = np.iinfo(np.int32).max
# Initialize the pipelines for each sd model
sd3_medium_pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16)
sd3_medium_pipe.enable_model_cpu_offload()
sd2_1_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16)
sd2_1_pipe.enable_model_cpu_offload()
sdxl_pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16)
sdxl_pipe.enable_model_cpu_offload()
sdxl_flash_pipe = StableDiffusionXLPipeline.from_pretrained("sd-community/sdxl-flash", torch_dtype=torch.float16)
sdxl_flash_pipe.enable_model_cpu_offload()
# Ensure sampler uses "trailing" timesteps for sdxl flash.
sdxl_flash_pipe.scheduler = DPMSolverSinglestepScheduler.from_config(sdxl_flash_pipe.scheduler.config, timestep_spacing="trailing")
stable_cascade_prior_pipe = StableCascadePriorPipeline.from_pretrained("stabilityai/stable-cascade-prior", variant="bf16", torch_dtype=torch.bfloat16)
stable_cascade_decoder_pipe = StableCascadeDecoderPipeline.from_pretrained("stabilityai/stable-cascade", variant="bf16", torch_dtype=torch.float16)
stable_cascade_prior_pipe.enable_model_cpu_offload()
stable_cascade_decoder_pipe.enable_model_cpu_offload()
# Helper function to generate images for a single model
@spaces.GPU(duration=80)
def generate_single_image(
prompt,
negative_prompt,
num_inference_steps,
guidance_scale,
height,
width,
seed,
num_images_per_prompt,
model_choice,
generator,
prior_num_inference_steps=None,
prior_guidance_scale=None,
decoder_num_inference_steps=None,
decoder_guidance_scale=None,
):
# Select the correct pipeline based on the model choice
if model_choice == "sd3 medium":
pipe = sd3_medium_pipe
elif model_choice == "sd2.1":
pipe = sd2_1_pipe
elif model_choice == "sdxl":
pipe = sdxl_pipe
elif model_choice == "sdxl flash":
pipe = sdxl_flash_pipe
elif model_choice == "stable cascade":
pipe = stable_cascade_prior_pipe
else:
raise ValueError(f"Invalid model choice: {model_choice}")
if model_choice == "stable cascade":
prior_output = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=prior_num_inference_steps,
guidance_scale=prior_guidance_scale,
height=height,
width=width,
generator=generator,
num_images_per_prompt=num_images_per_prompt,
)
output = stable_cascade_decoder_pipe(
image_embeddings=prior_output.image_embeddings.to(torch.float16),
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=decoder_num_inference_steps,
guidance_scale=decoder_guidance_scale,
).images
else:
output = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
height=height,
width=width,
generator=generator,
num_images_per_prompt=num_images_per_prompt,
).images
return output
# Define the image generation function for the Arena tab
@spaces.GPU(duration=80)
def generate_arena_images(
prompt,
negative_prompt,
num_inference_steps_a,
guidance_scale_a,
num_inference_steps_b,
guidance_scale_b,
height,
width,
seed,
num_images_per_prompt,
model_choice_a,
model_choice_b,
prior_num_inference_steps_a,
prior_guidance_scale_a,
decoder_num_inference_steps_a,
decoder_guidance_scale_a,
prior_num_inference_steps_b,
prior_guidance_scale_b,
decoder_num_inference_steps_b,
decoder_guidance_scale_b,
progress=gr.Progress(track_tqdm=True),
):
if seed == 0:
seed = random.randint(1, 2**32 - 1)
generator = torch.Generator().manual_seed(seed)
# Generate images for both models
images_a = generate_single_image(
prompt,
negative_prompt,
num_inference_steps_a,
guidance_scale_a,
height,
width,
seed,
num_images_per_prompt,
model_choice_a,
generator,
prior_num_inference_steps_a,
prior_guidance_scale_a,
decoder_num_inference_steps_a,
decoder_guidance_scale_a,
)
images_b = generate_single_image(
prompt,
negative_prompt,
num_inference_steps_b,
guidance_scale_b,
height,
width,
seed,
num_images_per_prompt,
model_choice_b,
generator,
prior_num_inference_steps_b,
prior_guidance_scale_b,
decoder_num_inference_steps_b,
decoder_guidance_scale_b,
)
return images_a, images_b
# Define the image generation function for the Individual tab
@spaces.GPU(duration=80)
def generate_individual_image(
prompt,
negative_prompt,
num_inference_steps,
guidance_scale,
height,
width,
seed,
num_images_per_prompt,
model_choice,
prior_num_inference_steps,
prior_guidance_scale,
decoder_num_inference_steps,
decoder_guidance_scale,
progress=gr.Progress(track_tqdm=True),
):
if seed == 0:
seed = random.randint(1, 2**32 - 1)
generator = torch.Generator().manual_seed(seed)
output = generate_single_image(
prompt,
negative_prompt,
num_inference_steps,
guidance_scale,
height,
width,
seed,
num_images_per_prompt,
model_choice,
generator,
prior_num_inference_steps,
prior_guidance_scale,
decoder_num_inference_steps,
decoder_guidance_scale,
)
return output
# Create the Gradio interface
examples_arena = [
[
"A woman in a red dress singing on top of a building.",
"deformed, distorted, disfigured, poorly drawn, bad anatomy, incorrect anatomy, extra limb, missing limb, floating limbs, mutated hands and fingers, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation",
25,
7.5,
25,
7.5,
1024,
1024,
42,
2,
"sd3 medium",
"sdxl",
25, #prior_num_inference_steps_a
4.0, #prior_guidance_scale_a
12, #decoder_num_inference_steps_a
0.0, #decoder_guidance_scale_a
25, #prior_num_inference_steps_b
4.0, #prior_guidance_scale_b
12, #decoder_num_inference_steps_b
0.0 #decoder_guidance_scale_b
],
[
"An astronaut on mars in a futuristic cyborg suit.",
"deformed, distorted, disfigured, poorly drawn, bad anatomy, incorrect anatomy, extra limb, missing limb, floating limbs, mutated hands and fingers, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation",
25,
7.5,
25,
7.5,
1024,
1024,
42,
2,
"sd3 medium",
"sdxl",
25, #prior_num_inference_steps_a
4.0, #prior_guidance_scale_a
12, #decoder_num_inference_steps_a
0.0, #decoder_guidance_scale_a
25, #prior_num_inference_steps_b
4.0, #prior_guidance_scale_b
12, #decoder_num_inference_steps_b
0.0 #decoder_guidance_scale_b
],
]
examples_individual = [
[
"A woman in a red dress singing on top of a building.",
"deformed, distorted, disfigured, poorly drawn, bad anatomy, incorrect anatomy, extra limb, missing limb, floating limbs, mutated hands and fingers, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation",
25,
7.5,
1024,
1024,
42,
2,
"sdxl",
25, #prior_num_inference_steps
4.0, #prior_guidance_scale
12, #decoder_num_inference_steps
0.0 #decoder_guidance_scale
],
[
"An astronaut on mars in a futuristic cyborg suit.",
"deformed, distorted, disfigured, poorly drawn, bad anatomy, incorrect anatomy, extra limb, missing limb, floating limbs, mutated hands and fingers, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation",
25,
7.5,
1024,
1024,
42,
2,
"sdxl",
25, #prior_num_inference_steps
4.0, #prior_guidance_scale
12, #decoder_num_inference_steps
0.0 #decoder_guidance_scale
],
]
css = """
.gradio-container{max-width: 1000px !important}
h1{text-align:center}
"""
with gr.Blocks(css=css) as demo:
with gr.Row():
with gr.Column():
gr.HTML(
"""
Stable Diffusion Arena
"""
)
gr.HTML(
"""
Made by Nick088
"""
)
with gr.Tabs():
with gr.TabItem("Arena"):
with gr.Group():
with gr.Column():
prompt = gr.Textbox(
label="Prompt",
info="Describe the image you want",
placeholder="A cat...",
)
model_choice_a = gr.Dropdown(
label="Stable Diffusion Model A",
choices=["sd3 medium", "sd2.1", "sdxl", "sdxl flash", "stable cascade"],
value="sd3 medium",
)
model_choice_b = gr.Dropdown(
label="Stable Diffusion Model B",
choices=["sd3 medium", "sd2.1", "sdxl", "sdxl flash", "stable cascade"],
value="sdxl",
)
run_button = gr.Button("Run")
result_1 = gr.Gallery(label="Generated Images (Model A)", elem_id="gallery_1")
result_2 = gr.Gallery(label="Generated Images (Model B)", elem_id="gallery_2")
with gr.Accordion("Advanced options", open=False):
negative_prompt = gr.Textbox(
label="Negative Prompt",
info="Describe what you don't want in the image",
value="deformed, distorted, disfigured, poorly drawn, bad anatomy, incorrect anatomy, extra limb, missing limb, floating limbs, mutated hands and fingers, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation",
placeholder="Ugly, bad anatomy...",
)
with gr.Row():
with gr.Column():
num_inference_steps_a = gr.Slider(
label="Inference Steps (Model A)",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=50,
value=25,
step=1,
visible=True
)
guidance_scale_a = gr.Slider(
label="Guidance Scale (Model A)",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=7.5,
step=0.1,
visible=True
)
prior_num_inference_steps_a = gr.Slider(
label="Prior Inference Steps (Model A)",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=50,
value=25,
step=1,
visible=False
)
prior_guidance_scale_a = gr.Slider(
label="Prior Guidance Scale (Model A)",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=4.0,
step=0.1,
visible=False
)
decoder_num_inference_steps_a = gr.Slider(
label="Decoder Inference Steps (Model A)",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=15,
value=15,
step=1,
visible=False
)
decoder_guidance_scale_a = gr.Slider(
label="Decoder Guidance Scale (Model A)",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=0.0,
step=0.1,
visible=False
)
with gr.Column():
num_inference_steps_b = gr.Slider(
label="Inference Steps (Model B)",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=50,
value=25,
step=1,
visible=True
)
guidance_scale_b = gr.Slider(
label="Guidance Scale (Model B)",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=7.5,
step=0.1,
visible=True
)
prior_num_inference_steps_b = gr.Slider(
label="Prior Inference Steps (Model B)",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=50,
value=25,
step=1,
visible=False
)
prior_guidance_scale_b = gr.Slider(
label="Prior Guidance Scale (Model B)",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=4.0,
step=0.1,
visible=False
)
decoder_num_inference_steps_b = gr.Slider(
label="Decoder Inference Steps (Model B)",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=15,
value=12,
step=1,
visible=False
)
decoder_guidance_scale_b = gr.Slider(
label="Decoder Guidance Scale (Model B)",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=0.0,
step=0.1,
visible=False
)
with gr.Row():
width = gr.Slider(
label="Width",
info="Width of the Image",
minimum=256,
maximum=1344,
step=32,
value=1024,
)
height = gr.Slider(
label="Height",
info="Height of the Image",
minimum=256,
maximum=1344,
step=32,
value=1024,
)
with gr.Row():
seed = gr.Slider(
value=42,
minimum=0,
maximum=MAX_SEED,
step=1,
label="Seed",
info="A starting point to initiate the generation process, put 0 for a random one",
)
num_images_per_prompt = gr.Slider(
label="Images Per Prompt",
info="Number of Images to generate with the settings",
minimum=1,
maximum=4,
step=1,
value=2,
)
def toggle_visibility_arena_a(model_choice_a):
if model_choice_a == "stable cascade":
return {
num_inference_steps_a: gr.update(visible=False),
guidance_scale_a: gr.update(visible=False),
prior_num_inference_steps_a: gr.update(visible=True),
prior_guidance_scale_a: gr.update(visible=True),
decoder_num_inference_steps_a: gr.update(visible=True),
decoder_guidance_scale_a: gr.update(visible=True),
}
elif model_choice_a == "sdxl flash":
return {
num_inference_steps_a: gr.update(visible=True, maximum=15, value=8),
guidance_scale_a: gr.update(visible=True, maximum=6.0, value=3.5),
prior_num_inference_steps_a: gr.update(visible=False),
prior_guidance_scale_a: gr.update(visible=False),
decoder_num_inference_steps_a: gr.update(visible=False),
decoder_guidance_scale_a: gr.update(visible=False),
}
else:
return {
num_inference_steps_a: gr.update(visible=True, maximum=50, value=25),
guidance_scale_a: gr.update(visible=True, maximum=10.0, value=7.5),
prior_num_inference_steps_a: gr.update(visible=False),
prior_guidance_scale_a: gr.update(visible=False),
decoder_num_inference_steps_a: gr.update(visible=False),
decoder_guidance_scale_a: gr.update(visible=False),
}
def toggle_visibility_arena_b(model_choice_b):
if model_choice_b == "stable cascade":
return {
num_inference_steps_b: gr.update(visible=False),
guidance_scale_b: gr.update(visible=False),
prior_num_inference_steps_b: gr.update(visible=True),
prior_guidance_scale_b: gr.update(visible=True),
decoder_num_inference_steps_b: gr.update(visible=True),
decoder_guidance_scale_b: gr.update(visible=True),
}
elif model_choice_b == "sdxl flash":
return {
num_inference_steps_b: gr.update(visible=True, maximum=15, value=8),
guidance_scale_b: gr.update(visible=True, maximum=6.0, value=3.5),
prior_num_inference_steps_b: gr.update(visible=False),
prior_guidance_scale_b: gr.update(visible=False),
decoder_num_inference_steps_b: gr.update(visible=False),
decoder_guidance_scale_b: gr.update(visible=False),
}
else:
return {
num_inference_steps_b: gr.update(visible=True, maximum=50, value=25),
guidance_scale_b: gr.update(visible=True, maximum=10.0, value=7.5),
prior_num_inference_steps_b: gr.update(visible=False),
prior_guidance_scale_b: gr.update(visible=False),
decoder_num_inference_steps_b: gr.update(visible=False),
decoder_guidance_scale_b: gr.update(visible=False),
}
model_choice_a.change(
toggle_visibility_arena_a,
inputs=[model_choice_a],
outputs=[
num_inference_steps_a,
guidance_scale_a,
prior_num_inference_steps_a,
prior_guidance_scale_a,
decoder_num_inference_steps_a,
decoder_guidance_scale_a
]
)
model_choice_b.change(
toggle_visibility_arena_b,
inputs=[model_choice_b],
outputs=[
num_inference_steps_b,
guidance_scale_b,
prior_num_inference_steps_b,
prior_guidance_scale_b,
decoder_num_inference_steps_b,
decoder_guidance_scale_b
]
)
gr.Examples(
examples=examples_arena,
inputs=[
prompt,
negative_prompt,
num_inference_steps_a,
guidance_scale_a,
num_inference_steps_b,
guidance_scale_b,
height,
width,
seed,
num_images_per_prompt,
model_choice_a,
model_choice_b,
prior_num_inference_steps_a,
prior_guidance_scale_a,
decoder_num_inference_steps_a,
decoder_guidance_scale_a,
prior_num_inference_steps_b,
prior_guidance_scale_b,
decoder_num_inference_steps_b,
decoder_guidance_scale_b,
],
outputs=[result_1, result_2],
fn=generate_arena_images,
)
gr.on(
triggers=[
prompt.submit,
run_button.click,
],
fn=generate_arena_images,
inputs=[
prompt,
negative_prompt,
num_inference_steps_a,
guidance_scale_a,
num_inference_steps_b,
guidance_scale_b,
height,
width,
seed,
num_images_per_prompt,
model_choice_a,
model_choice_b,
prior_num_inference_steps_a,
prior_guidance_scale_a,
decoder_num_inference_steps_a,
decoder_guidance_scale_a,
prior_num_inference_steps_b,
prior_guidance_scale_b,
decoder_num_inference_steps_b,
decoder_guidance_scale_b,
],
outputs=[result_1, result_2],
)
with gr.TabItem("Individual"):
with gr.Group():
with gr.Column():
prompt = gr.Textbox(
label="Prompt",
info="Describe the image you want",
placeholder="A cat...",
)
model_choice = gr.Dropdown(
label="Stable Diffusion Model",
choices=["sd3 medium", "sd2.1", "sdxl", "sdxl flash", "stable cascade"],
value="sd3 medium",
)
run_button = gr.Button("Run")
result = gr.Gallery(label="Generated AI Images", elem_id="gallery")
with gr.Accordion("Advanced options", open=False):
with gr.Row():
negative_prompt = gr.Textbox(
label="Negative Prompt",
info="Describe what you don't want in the image",
value="deformed, distorted, disfigured, poorly drawn, bad anatomy, incorrect anatomy, extra limb, missing limb, floating limbs, mutated hands and fingers, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation",
placeholder="Ugly, bad anatomy...",
)
with gr.Row():
num_inference_steps = gr.Slider(
label="Number of Inference Steps",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=50,
value=25,
step=1,
visible=True
)
guidance_scale = gr.Slider(
label="Guidance Scale",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=7.5,
step=0.1,
visible=True
)
prior_num_inference_steps = gr.Slider(
label="Prior Inference Steps",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=50,
value=25,
step=1,
visible=False
)
prior_guidance_scale = gr.Slider(
label="Prior Guidance Scale",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=4.0,
step=0.1,
visible=False
)
decoder_num_inference_steps = gr.Slider(
label="Decoder Inference Steps",
info="The number of denoising steps of the image. More denoising steps usually lead to a higher quality image at the cost of slower inference",
minimum=1,
maximum=15,
value=12,
step=1,
visible=False
)
decoder_guidance_scale = gr.Slider(
label="Decoder Guidance Scale",
info="Controls how much the image generation process follows the text prompt. Higher values make the image stick more closely to the input text.",
minimum=0.0,
maximum=10.0,
value=0.0,
step=0.1,
visible=False
)
with gr.Row():
width = gr.Slider(
label="Width",
info="Width of the Image",
minimum=256,
maximum=1344,
step=32,
value=1024,
)
height = gr.Slider(
label="Height",
info="Height of the Image",
minimum=256,
maximum=1344,
step=32,
value=1024,
)
with gr.Row():
seed = gr.Slider(
value=42,
minimum=0,
maximum=MAX_SEED,
step=1,
label="Seed",
info="A starting point to initiate the generation process, put 0 for a random one",
)
num_images_per_prompt = gr.Slider(
label="Images Per Prompt",
info="Number of Images to generate with the settings",
minimum=1,
maximum=4,
step=1,
value=2,
)
def toggle_visibility_individual(model_choice):
if model_choice == "stable cascade":
return {
num_inference_steps: gr.update(visible=False),
guidance_scale: gr.update(visible=False),
prior_num_inference_steps: gr.update(visible=True),
prior_guidance_scale: gr.update(visible=True),
decoder_num_inference_steps: gr.update(visible=True),
decoder_guidance_scale: gr.update(visible=True),
}
elif model_choice == "sdxl flash":
return {
num_inference_steps: gr.update(visible=True, maximum=15, value=8),
guidance_scale: gr.update(visible=True, maximum=6.0, value=3.5),
prior_num_inference_steps: gr.update(visible=False),
prior_guidance_scale: gr.update(visible=False),
decoder_num_inference_steps: gr.update(visible=False),
decoder_guidance_scale: gr.update(visible=False),
}
else:
return {
num_inference_steps: gr.update(visible=True, maximum=50, value=25),
guidance_scale: gr.update(visible=True, maximum=10.0, value=7.5),
prior_num_inference_steps: gr.update(visible=False),
prior_guidance_scale: gr.update(visible=False),
decoder_num_inference_steps: gr.update(visible=False),
decoder_guidance_scale: gr.update(visible=False),
}
model_choice.change(
toggle_visibility_individual,
inputs=[model_choice],
outputs=[
num_inference_steps,
guidance_scale,
prior_num_inference_steps,
prior_guidance_scale,
decoder_num_inference_steps,
decoder_guidance_scale
]
)
gr.Examples(
examples=examples_individual,
inputs=[
prompt,
negative_prompt,
num_inference_steps,
guidance_scale,
height,
width,
seed,
num_images_per_prompt,
model_choice,
prior_num_inference_steps,
prior_guidance_scale,
decoder_num_inference_steps,
decoder_guidance_scale,
],
outputs=[result],
fn=generate_individual_image,
)
gr.on(
triggers=[
prompt.submit,
run_button.click,
],
fn=generate_individual_image,
inputs=[
prompt,
negative_prompt,
num_inference_steps,
guidance_scale,
height,
width,
seed,
num_images_per_prompt,
model_choice,
prior_num_inference_steps,
prior_guidance_scale,
decoder_num_inference_steps,
decoder_guidance_scale,
],
outputs=[result],
)
demo.queue().launch(share=False)