nielsgl's picture
Update app.py
bd83a96
from huggingface_hub import from_pretrained_keras
from keras_cv import models
import gradio as gr
from tensorflow import keras
from diffusers import StableDiffusionPipeline
keras.mixed_precision.set_global_policy("mixed_float16")
# prepare model
resolution = 512
# sd_dreambooth_model = models.StableDiffusion(
# img_width=resolution, img_height=resolution
# )
# db_diffusion_model = from_pretrained_keras("keras-dreambooth/dreambooth_diffusion_model")
# sd_dreambooth_model._diffusion_model = db_diffusion_model
# checkpoint of the converted Stable Diffusion from KerasCV
model_ckpt = "nielsgl/dreambooth-keras-pug-ace-sd2.1"
pipeline = StableDiffusionPipeline.from_pretrained(model_ckpt)
pipeline = pipeline.to("cuda")
unique_id = "puggieace"
class_label = "dog"
prompt = f"A photo of {unique_id} {class_label} on the beach"
image = pipeline(prompt, num_inference_steps=50).images[0]
# generate images
def infer(prompt, negative_prompt, guidance_scale=10, num_inference_steps=50):
neg = negative_prompt if negative_prompt else None
imgs = []
while len(imgs) != 4:
next_prompt = pipeline(prompt, negative_prompt=neg, guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, num_images_per_prompt=5)
for img, is_neg in zip(next_prompt.images, next_prompt.nsfw_content_detected):
if not is_neg:
imgs.append(img)
if len(imgs) == 4:
break
return imgs
output = gr.Gallery(label="Outputs").style(grid=(1,2))
# customize interface
title = "KerasCV Stable Diffusion Demo on images of Ace."
description = "This is a dreambooth model fine-tuned on images of my pug named Ace. To try it, input the concept with `puggieace dog`."
examples=[
["Portrait photo of puggieace dog on a beachtowel wearing sunglasses on the beach, sunset in background, golden hour", "", 12, 50],
["A photo of a cute puggieace dog getting a haircut in a barbershop, ultra realistic, intricate details, highly detailed, photorealistic, octane render, 8 k, unreal engine. art by artgerm and greg rutkowski and charlie bowater and magali villeneuve and alphonse mucha", "", 12, 75],
["Portrait of puggieace dog as a Roman Emperor, city in background, ultra realistic, intricate details, eerie, highly detailed, photorealistic, octane render, 8 k, unreal engine. art by artgerm and greg rutkowski and charlie bowater and magali villeneuve and alphonse mucha", "", 15, 75],
["Photo of cute puggieace dog as an astronaut, space and planet in background, ultra realistic, concept art, intricate details, highly detailed, photorealistic, octane render, 8 k, unreal engine. trending on artstation", "", 15, 75],
["Photo of cute puggieace dog as super hero, futuristic city in background, cinematic light, high dynamic range, insane intricate details, stunning cinema effects, aesthetic, masterpiece, trending on artstation, cartoon art", "", 12, 75],
]
base_14 = "https://huggingface.co/nielsgl/dreambooth-pug-ace-sd1.4/resolve/main/"
base_21 = "https://huggingface.co/nielsgl/dreambooth-keras-pug-ace-sd2.1/resolve/main/"
model_card_1 = f"""
# KerasCV Stable Diffusion in Diffusers πŸ§¨πŸ€—
DreamBooth model for the `puggieace` concept trained by nielsgl on the [nielsgl/dreambooth-ace](https://huggingface.co/datasets/nielsgl/dreambooth-ace) dataset.
It can be used by modifying the `instance_prompt`: **a photo of puggieace**.
The examples are from 2 different Keras CV models (`StableDiffusion` and `StableDiffusionV2`, corresponding to Stable Diffusion V1.4 and V2.1, respectively) trained on the same dataset (`nielsgl/dreambooth-ace`).
## Description
The Stable Diffusion V2 pipeline contained in the corresponding repository (`nielsgl/dreambooth-keras-pug-ace-sd2.1`) was created using a modified version of [this Space](https://huggingface.co/spaces/sayakpaul/convert-kerascv-sd-diffusers) for StableDiffusionV2 from KerasCV. The purpose is to convert the KerasCV Stable Diffusion weights in a way that is compatible with [Diffusers](https://github.com/huggingface/diffusers). This allows users to fine-tune using KerasCV and use the fine-tuned weights in Diffusers taking advantage of its nifty features (like [schedulers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/schedulers), [fast attention](https://huggingface.co/docs/diffusers/optimization/fp16), etc.).
This model was created as part of the Keras DreamBooth Sprint πŸ”₯. Visit the [organisation page](https://huggingface.co/keras-dreambooth) for instructions on how to take part!
## Demo
"""
model_card_2 = f"""
## Examples
### Stable Diffusion V1.4
> Portrait of puggieace dog as a Roman Emperor, city in background
![Portrait of puggieace dog as a Roman Emperor, city in background, ultra realistic, intricate details, eerie, highly detailed, photorealistic, octane render, 8 k, unreal engine. art by artgerm and greg rutkowski and charlie bowater and magali villeneuve and alphonse mucha]({base_14}examples/emperor-1.4.jpeg)
> Photo of puggieace dog wearing sunglasses on the beach, sunset in background, golden hour
![Photo of puggieace dog wearing sunglasses on the beach, sunset in background, golden hour]({base_14}examples/beach-1.4.jpg)
> Photo of cute puggieace dog as an astronaut, planet and spaceship in background
![Photo of cute puggieace dog as an astronaut, planet and spaceship in background, ultra realistic, intricate details, highly detailed, photorealistic, octane render, 8 k, unreal engine. trending on artstation]({base_14}examples/astronaut-1.4.jpg)
### Stable Diffusion V2.1
> Portrait painting of a cute puggieace dog as a samurai
![Portrait painting of a cute puggieace dog as a samurai, ultra realistic, concept art, intricate details, eerie, highly detailed, photorealistic, octane render, 8 k, unreal engine. art by artgerm and greg rutkowski and charlie bowater and magali villeneuve and alphonse mucha]({base_21}examples/samurai-2.1.jpg)
> Photo of cute puggieace dog as an astronaut, space and planet in background
![Photo of cute puggieace dog as an astronaut, space and planet in background, ultra realistic, concept art, intricate details, highly detailed, photorealistic, octane render, 8 k, unreal engine. art by artgerm and greg rutkowski and charlie bowater, trending on artstation]({base_21}examples/astronaut-2.1.jpg)
> A photo of a cute puggieace dog getting a haircut in a barbershop
![A photo of a cute puggieace dog getting a haircut in a barbershop, ultra realistic, intricate details, highly detailed, photorealistic, octane render, 8 k, unreal engine. art by artgerm and greg rutkowski and charlie bowater and magali villeneuve and alphonse mucha]({base_21}examples/haircut-2.1.jpg)
> Portrait photo of puggieace dog in New York
![Portrait photo of puggieace dog in New York, city and skyscrapers in background, highly detailed, photorealistic, hdr, 4k]({base_21}examples/ny-2.1.jpg)
> Portrait of puggieace dog as a Roman Emperor, city in background
![Portrait of puggieace dog as a Roman Emperor, city in background, ultra realistic, intricate details, eerie, highly detailed, photorealistic, octane render, 8 k, unreal engine. art by artgerm and greg rutkowski and charlie bowater and magali villeneuve and alphonse mucha]({base_21}examples/emperor-2.1.jpg)
## Usage with Stable Diffusion V1.4
```python
from huggingface_hub import from_pretrained_keras
import keras_cv
import matplotlib.pyplot as plt
model = keras_cv.models.StableDiffusion(img_width=512, img_height=512, jit_compile=True)
model._diffusion_model = from_pretrained_keras("nielsgl/dreambooth-pug-ace")
model._text_encoder = from_pretrained_keras("nielsgl/dreambooth-pug-ace-text-encoder")
images = model.text_to_image("a photo of puggieace dog on the beach", batch_size=3)
plt.imshow(image[0])
```
## Usage with Stable Diffusion V2.1
```python
from diffusers import StableDiffusionPipeline
pipeline = StableDiffusionPipeline.from_pretrained('nielsgl/dreambooth-keras-pug-ace-sd2.1')
image = pipeline().images[0]
image
```
### Training hyperparameters
The following hyperparameters were used during training for Stable Diffusion v1.4:
| Hyperparameters | Value |
| :-- | :-- |
| name | RMSprop |
| weight_decay | None |
| clipnorm | None |
| global_clipnorm | None |
| clipvalue | None |
| use_ema | False |
| ema_momentum | 0.99 |
| ema_overwrite_frequency | 100 |
| jit_compile | True |
| is_legacy_optimizer | False |
| learning_rate | 0.0010000000474974513 |
| rho | 0.9 |
| momentum | 0.0 |
| epsilon | 1e-07 |
| centered | False |
| training_precision | float32 |
"""
with gr.Blocks() as demo:
with gr.Row():
gr.Markdown(model_card_1)
with gr.Row():
with gr.Column():
prompt_pos = gr.Textbox(label="Positive Prompt", value="a photo of puggieace dog getting a haircut")
prompt_neg = gr.Textbox(label="Negative Prompt", value="bad anatomy, blurry")
# gr.Slider(label='Number of gen image', minimum=1, maximum=4, value=2, step=1),
prompt_gs = gr.Number(label='Guidance scale', value=12)
prompt_steps = gr.Slider(label="Inference Steps",value=50)
prompt_btn = gr.Button("Generate")
with gr.Column():
output = gr.Gallery(label="Outputs").style(grid=(1,2))
prompt_btn.click(infer, inputs=[prompt_pos, prompt_neg, prompt_gs, prompt_steps], outputs=[output])
with gr.Row():
gr.Examples(examples, inputs=[prompt_pos, prompt_neg, prompt_gs, prompt_steps], outputs=output, fn=infer, cache_examples=True)
# gr.Interface(infer, inputs=[gr.Textbox(label="Positive Prompt", value="a photo of puggieace dog getting a haircut"),
# gr.Textbox(label="Negative Prompt", value="bad anatomy, blurry"),
# # gr.Slider(label='Number of gen image', minimum=1, maximum=4, value=2, step=1),
# gr.Number(label='Guidance scale', value=12),
# gr.Slider(label="Inference Steps",value=50),
# ], outputs=[output], title=title, description=description, examples=examples).queue()
with gr.Row():
with gr.Column():
gr.Markdown(model_card_2)
with gr.Column():
gr.Markdown(" ")
demo.queue().launch()
# with gr.Blocks() as card_interface:
# gr.Markdown(model_card)
# demo_interface = gr.Interface(infer, inputs=[gr.Textbox(label="Positive Prompt", value="a photo of puggieace dog getting a haircut"),
# gr.Textbox(label="Negative Prompt", value="bad anatomy, blurry"),
# # gr.Slider(label='Number of gen image', minimum=1, maximum=4, value=2, step=1),
# gr.Number(label='Guidance scale', value=12),
# gr.Slider(label="Inference Steps",value=50),
# ], outputs=[output], title=title, description=description, examples=examples).queue()
# gr.TabbedInterface([card_interface, demo_interface], ["Model Card", "Demo πŸ€—"]).launch()