diff --git a/Readme.md b/Readme.md
index b179563..f02d49a 100644
--- a/Readme.md
+++ b/Readme.md
@@ -5,6 +5,10 @@ DiffusionMagic focused on the following areas:
- Cross-platform (Windows/Linux/Mac)
- Modular design, latest best optimizations for speed and memory
+## Stable diffusion XL Colab
+We can run StableDiffusion XL 0.9 on Google Colab
+[](https://colab.research.google.com/drive/1KrmcU2gONIQ2WihI1s6uITgDDzkbKaJK?usp=sharing)
+
## Features
- Supports various Stable Diffusion workflows
@@ -113,6 +117,7 @@ Or we can clone the model use the local folder path as model id.
## Linting (Development)
Run the following commands from src folder
`mypy --ignore-missing-imports --explicit-package-bases .`
+
`flake8 --max-line-length=100 .`
## Contribute
Contributions are welcomed.
diff --git a/configs/stable_diffusion_models.txt b/configs/stable_diffusion_models.txt
index 07c4106..f5dd3c0 100644
--- a/configs/stable_diffusion_models.txt
+++ b/configs/stable_diffusion_models.txt
@@ -16,4 +16,5 @@ lllyasviel/sd-controlnet-hed
lllyasviel/sd-controlnet-openpose
lllyasviel/sd-controlnet-depth
lllyasviel/sd-controlnet-scribble
-lllyasviel/sd-controlnet-seg
\ No newline at end of file
+lllyasviel/sd-controlnet-seg
+stabilityai/stable-diffusion-xl-base-1.0
\ No newline at end of file
diff --git a/environment.yml b/environment.yml
index 895a56e..e85d7e9 100644
--- a/environment.yml
+++ b/environment.yml
@@ -12,16 +12,17 @@ dependencies:
- torchvision=0.15.0
- numpy=1.19.2
- pip:
- - accelerate==0.17.1
- - diffusers==0.14.0
- - gradio==3.17.1
- - safetensors==0.2.8
+ - accelerate==0.21.0
+ - diffusers==0.19.3
+ - gradio==3.32.0
+ - safetensors==0.3.1
- scipy==1.10.0
- - transformers==4.26.0
+ - transformers==4.31.0
- pydantic==1.10.4
- mypy==1.0.0
- black==23.1.0
- flake8==6.0.0
- markupsafe==2.0.1
- opencv-contrib-python==4.7.0.72
- - controlnet-aux==0.0.1
\ No newline at end of file
+ - controlnet-aux==0.0.1
+ - invisible-watermark==0.2.0
\ No newline at end of file
diff --git a/src/backend/generate.py b/src/backend/generate.py
index bbae16e..bad9e6d 100644
--- a/src/backend/generate.py
+++ b/src/backend/generate.py
@@ -15,6 +15,7 @@
)
from backend.controlnet.ControlContext import ControlnetContext
from backend.stablediffusion.stablediffusion import StableDiffusion
+from backend.stablediffusion.stablediffusionxl import StableDiffusionXl
from settings import AppSettings
@@ -30,6 +31,7 @@ def __init__(self, compute: Computing):
self.stable_diffusion_depth = StableDiffusionDepthToImage(compute)
self.stable_diffusion_pix_to_pix = StableDiffusionInstructPixToPix(compute)
self.controlnet = ControlnetContext(compute)
+ self.stable_diffusion_xl = StableDiffusionXl(compute)
self.app_settings = AppSettings().get_settings()
self.model_id = self.app_settings.model_settings.model_id
self.low_vram_mode = self.app_settings.low_memory_mode
@@ -78,6 +80,15 @@ def _init_stable_diffusion(self):
)
self.pipe_initialized = True
+ def _init_stable_diffusion_xl(self):
+ if not self.pipe_initialized:
+ print("Initializing stable diffusion xl pipeline")
+ self.stable_diffusion_xl.get_text_to_image_xl_pipleline(
+ self.model_id,
+ self.low_vram_mode,
+ )
+ self.pipe_initialized = True
+
def diffusion_image_to_image(
self,
image,
@@ -355,3 +366,78 @@ def diffusion_control_to_image(
"CannyToImage",
)
return images
+
+ def diffusion_text_to_image_xl(
+ self,
+ prompt,
+ neg_prompt,
+ image_height,
+ image_width,
+ inference_steps,
+ scheduler,
+ guidance_scale,
+ num_images,
+ attention_slicing,
+ vae_slicing,
+ seed,
+ ) -> Any:
+ stable_diffusion_settings = StableDiffusionSetting(
+ prompt=prompt,
+ negative_prompt=neg_prompt,
+ image_height=image_height,
+ image_width=image_width,
+ inference_steps=inference_steps,
+ guidance_scale=guidance_scale,
+ number_of_images=num_images,
+ scheduler=scheduler,
+ seed=seed,
+ attention_slicing=attention_slicing,
+ vae_slicing=vae_slicing,
+ )
+ self._init_stable_diffusion_xl()
+ images = self.stable_diffusion_xl.text_to_image_xl(stable_diffusion_settings)
+ self._save_images(
+ images,
+ "TextToImage",
+ )
+ return images
+
+ def diffusion_image_to_image_xl(
+ self,
+ image,
+ strength,
+ prompt,
+ neg_prompt,
+ image_height,
+ image_width,
+ inference_steps,
+ scheduler,
+ guidance_scale,
+ num_images,
+ attention_slicing,
+ seed,
+ ) -> Any:
+ stable_diffusion_image_settings = StableDiffusionImageToImageSetting(
+ image=image,
+ strength=strength,
+ prompt=prompt,
+ negative_prompt=neg_prompt,
+ image_height=image_height,
+ image_width=image_width,
+ inference_steps=inference_steps,
+ guidance_scale=guidance_scale,
+ number_of_images=num_images,
+ scheduler=scheduler,
+ seed=seed,
+ attention_slicing=attention_slicing,
+ )
+ self._init_stable_diffusion_xl()
+ images = self.stable_diffusion_xl.image_to_image(
+ stable_diffusion_image_settings
+ )
+
+ self._save_images(
+ images,
+ "ImageToImage",
+ )
+ return images
diff --git a/src/backend/stablediffusion/stable_diffusion_types.py b/src/backend/stablediffusion/stable_diffusion_types.py
index e87b132..3a485a6 100644
--- a/src/backend/stablediffusion/stable_diffusion_types.py
+++ b/src/backend/stablediffusion/stable_diffusion_types.py
@@ -16,6 +16,7 @@ class StableDiffusionType(str, Enum):
controlnet_depth = "controlnet_depth"
controlnet_scribble = "controlnet_scribble"
controlnet_seg = "controlnet_seg"
+ stable_diffusion_xl = "StableDiffusionXl"
def get_diffusion_type(
@@ -44,4 +45,6 @@ def get_diffusion_type(
stable_diffusion_type = StableDiffusionType.controlnet_scribble
elif "controlnet-seg" in model_id:
stable_diffusion_type = StableDiffusionType.controlnet_seg
+ elif "stable-diffusion-xl" in model_id:
+ stable_diffusion_type = StableDiffusionType.stable_diffusion_xl
return stable_diffusion_type
diff --git a/src/backend/stablediffusion/stablediffusionxl.py b/src/backend/stablediffusion/stablediffusionxl.py
new file mode 100644
index 0000000..8b1cbbd
--- /dev/null
+++ b/src/backend/stablediffusion/stablediffusionxl.py
@@ -0,0 +1,181 @@
+from time import time
+
+import torch
+from diffusers import (
+ DiffusionPipeline,
+ StableDiffusionXLImg2ImgPipeline,
+)
+from PIL import Image
+
+from backend.computing import Computing
+from backend.stablediffusion.modelmeta import ModelMeta
+from backend.stablediffusion.models.scheduler_types import SchedulerType
+from backend.stablediffusion.models.setting import (
+ StableDiffusionImageToImageSetting,
+ StableDiffusionSetting,
+)
+from backend.stablediffusion.scheduler_mixin import SamplerMixin
+
+
+class StableDiffusionXl(SamplerMixin):
+ def __init__(self, compute: Computing):
+ self.compute = compute
+ self.pipeline = None
+ self.device = self.compute.name
+
+ super().__init__()
+
+ def get_text_to_image_xl_pipleline(
+ self,
+ model_id: str = "stabilityai/stable-diffusion-xl-base-1.0",
+ low_vram_mode: bool = False,
+ sampler: str = SchedulerType.DPMSolverMultistepScheduler.value,
+ ):
+ repo_id = model_id
+ model_meta = ModelMeta(repo_id)
+ is_lora_model = model_meta.is_loramodel()
+ if is_lora_model:
+ print("LoRA model detected")
+ self.model_id = model_meta.get_lora_base_model()
+ print(f"LoRA base model - {self.model_id}")
+ else:
+ self.model_id = model_id
+
+ self.low_vram_mode = low_vram_mode
+ print(f"StableDiffusion - {self.compute.name},{self.compute.datatype}")
+ print(f"using model {model_id}")
+ self.default_sampler = self.find_sampler(
+ sampler,
+ self.model_id,
+ )
+ tic = time()
+ self._load_model()
+ delta = time() - tic
+ print(f"Model loaded in {delta:.2f}s ")
+
+ if self.pipeline is None:
+ raise Exception("Text to image pipeline not initialized")
+ if is_lora_model:
+ self.pipeline.unet.load_attn_procs(repo_id)
+ self._pipeline_to_device()
+ components = self.pipeline.components
+ self.img_to_img_pipeline = StableDiffusionXLImg2ImgPipeline(**components)
+
+ def text_to_image_xl(self, setting: StableDiffusionSetting):
+ if self.pipeline is None:
+ raise Exception("Text to image pipeline not initialized")
+
+ self.pipeline.scheduler = self.find_sampler(
+ setting.scheduler,
+ self.model_id,
+ )
+ generator = None
+ if setting.seed != -1:
+ print(f"Using seed {setting.seed}")
+ generator = torch.Generator(self.device).manual_seed(setting.seed)
+
+ # if setting.attention_slicing:
+ # self.pipeline.enable_attention_slicing()
+ # else:
+ # self.pipeline.disable_attention_slicing()
+
+ if setting.vae_slicing:
+ self.pipeline.enable_vae_slicing()
+ else:
+ self.pipeline.disable_vae_slicing()
+
+ images = self.pipeline(
+ setting.prompt,
+ guidance_scale=setting.guidance_scale,
+ num_inference_steps=setting.inference_steps,
+ height=setting.image_height,
+ width=setting.image_width,
+ negative_prompt=setting.negative_prompt,
+ num_images_per_prompt=setting.number_of_images,
+ generator=generator,
+ ).images
+
+ # self.pipeline.unet = torch.compile(
+ # self.pipeline.unet,
+ # mode="reduce-overhead",
+ # fullgraph=True,
+ # )
+ return images
+
+ def _pipeline_to_device(self):
+ if self.low_vram_mode:
+ print("Running in low VRAM mode,slower to generate images")
+ self.pipeline.enable_sequential_cpu_offload()
+ else:
+ if self.compute.name == "cuda":
+ self.pipeline = self.pipeline.to("cuda")
+ elif self.compute.name == "mps":
+ self.pipeline = self.pipeline.to("mps")
+
+ def _load_full_precision_model(self):
+ self.pipeline = DiffusionPipeline.from_pretrained(
+ self.model_id,
+ torch_dtype=self.compute.datatype,
+ scheduler=self.default_sampler,
+ )
+
+ def _load_model(self):
+ if self.compute.name == "cuda":
+ try:
+ self.pipeline = DiffusionPipeline.from_pretrained(
+ self.model_id,
+ torch_dtype=self.compute.datatype,
+ scheduler=self.default_sampler,
+ use_safetensors=True,
+ variant="fp16",
+ )
+ except Exception as ex:
+ print(
+ f" The fp16 of the model not found using full precision model, {ex}"
+ )
+ self._load_full_precision_model()
+ else:
+ self._load_full_precision_model()
+
+ def image_to_image(self, setting: StableDiffusionImageToImageSetting):
+ if setting.scheduler is None:
+ raise Exception("Scheduler cannot be empty")
+
+ print("Running image to image pipeline")
+ self.img_to_img_pipeline.scheduler = self.find_sampler( # type: ignore
+ setting.scheduler,
+ self.model_id,
+ )
+ generator = None
+ if setting.seed != -1 and setting.seed:
+ print(f"Using seed {setting.seed}")
+ generator = torch.Generator(self.device).manual_seed(setting.seed)
+
+ if setting.attention_slicing:
+ self.img_to_img_pipeline.enable_attention_slicing() # type: ignore
+ else:
+ self.img_to_img_pipeline.disable_attention_slicing() # type: ignore
+
+ if setting.vae_slicing:
+ self.pipeline.enable_vae_slicing() # type: ignore
+ else:
+ self.pipeline.disable_vae_slicing() # type: ignore
+
+ init_image = setting.image.resize(
+ (
+ setting.image_width,
+ setting.image_height,
+ ),
+ Image.Resampling.LANCZOS,
+ )
+ images = self.img_to_img_pipeline( # type: ignore
+ image=init_image,
+ strength=setting.strength,
+ prompt=setting.prompt,
+ guidance_scale=setting.guidance_scale,
+ num_inference_steps=setting.inference_steps,
+ negative_prompt=setting.negative_prompt,
+ num_images_per_prompt=setting.number_of_images,
+ generator=generator,
+ ).images
+ return images
diff --git a/src/constants.py b/src/constants.py
index 79f2bca..465b7e0 100644
--- a/src/constants.py
+++ b/src/constants.py
@@ -1,4 +1,4 @@
-VERSION = "2.0.0-beta.0"
+VERSION = "3.0.0"
STABLE_DIFFUSION_MODELS_FILE = "stable_diffusion_models.txt"
APP_SETTINGS_FILE = "settings.yaml"
CONFIG_DIRECTORY = "configs"
diff --git a/src/frontend/web/depth_to_image_ui.py b/src/frontend/web/depth_to_image_ui.py
index a4c7cd4..367e68d 100644
--- a/src/frontend/web/depth_to_image_ui.py
+++ b/src/frontend/web/depth_to_image_ui.py
@@ -114,7 +114,7 @@ def random_seed():
show_label=True,
elem_id="gallery",
).style(
- grid=2,
+ columns=2,
)
generate_btn.click(
fn=generate_callback_fn,
diff --git a/src/frontend/web/image_inpainting_ui.py b/src/frontend/web/image_inpainting_ui.py
index edd9b48..3a1f05d 100644
--- a/src/frontend/web/image_inpainting_ui.py
+++ b/src/frontend/web/image_inpainting_ui.py
@@ -69,7 +69,7 @@ def random_seed():
label="Number of images to generate",
)
attn_slicing = gr.Checkbox(
- label="Attention slicing (Enable if low VRAM)",
+ label="Attention slicing (Not used)",
value=True,
)
seed = gr.Number(
@@ -105,7 +105,7 @@ def random_seed():
show_label=True,
elem_id="gallery",
).style(
- grid=2,
+ columns=2,
)
generate_btn.click(
fn=generate_callback_fn,
diff --git a/src/frontend/web/image_inpainting_xl_ui.py b/src/frontend/web/image_inpainting_xl_ui.py
new file mode 100644
index 0000000..be77595
--- /dev/null
+++ b/src/frontend/web/image_inpainting_xl_ui.py
@@ -0,0 +1,115 @@
+from typing import Any
+
+import gradio as gr
+
+from backend.stablediffusion.models.scheduler_types import (
+ SchedulerType,
+ get_sampler_names,
+)
+
+random_enabled = True
+
+
+def get_image_inpainting_xl_ui(generate_callback_fn: Any) -> None:
+ with gr.Blocks():
+ with gr.Row():
+ with gr.Column():
+
+ def random_seed():
+ global random_enabled
+ random_enabled = not random_enabled
+ seed_val = -1
+ if not random_enabled:
+ seed_val = 42
+
+ return gr.Number.update(
+ interactive=not random_enabled, value=seed_val
+ )
+
+ input_image = gr.Image(label="Input image", type="pil", tool="sketch")
+ prompt = gr.Textbox(
+ label="Describe the image you'd like to see",
+ lines=3,
+ placeholder="A fantasy landscape",
+ )
+
+ neg_prompt = gr.Textbox(
+ label="Don't want to see",
+ lines=1,
+ placeholder="",
+ value="bad, deformed, ugly, bad anatomy",
+ )
+ with gr.Accordion("Advanced options", open=False):
+ image_height = gr.Slider(
+ 1024, 2048, value=1024, step=64, label="Image Height"
+ )
+ image_width = gr.Slider(
+ 1024, 2048, value=1024, step=64, label="Image Width"
+ )
+ num_inference_steps = gr.Slider(
+ 1, 100, value=20, step=1, label="Inference Steps"
+ )
+ scheduler = gr.Dropdown(
+ get_sampler_names(),
+ value=SchedulerType.DPMSolverMultistepScheduler.value,
+ label="Sampler",
+ )
+ guidance_scale = gr.Slider(
+ 1.0,
+ 30.0,
+ value=7.5,
+ step=0.5,
+ label="Guidance Scale",
+ )
+ num_images = gr.Slider(
+ 1,
+ 50,
+ value=1,
+ step=1,
+ label="Number of images to generate",
+ )
+ attn_slicing = gr.Checkbox(
+ label="Attention slicing (Enable if low VRAM)",
+ value=True,
+ )
+ seed = gr.Number(
+ label="Seed",
+ value=-1,
+ precision=0,
+ interactive=False,
+ )
+ seed_checkbox = gr.Checkbox(
+ label="Use random seed",
+ value=True,
+ interactive=True,
+ )
+
+ input_params = [
+ input_image,
+ prompt,
+ neg_prompt,
+ image_height,
+ image_width,
+ num_inference_steps,
+ scheduler,
+ guidance_scale,
+ num_images,
+ attn_slicing,
+ seed,
+ ]
+
+ with gr.Column():
+ generate_btn = gr.Button("Inpaint!", elem_id="generate_button")
+ output = gr.Gallery(
+ label="Generated images",
+ show_label=True,
+ elem_id="gallery",
+ ).style(
+ columns=2,
+ )
+ generate_btn.click(
+ fn=generate_callback_fn,
+ inputs=input_params,
+ outputs=output,
+ )
+ seed_checkbox.change(fn=random_seed, outputs=seed)
diff --git a/src/frontend/web/image_to_image_ui.py b/src/frontend/web/image_to_image_ui.py
index 619e497..f9975fb 100644
--- a/src/frontend/web/image_to_image_ui.py
+++ b/src/frontend/web/image_to_image_ui.py
@@ -114,7 +114,7 @@ def random_seed():
show_label=True,
elem_id="gallery",
).style(
- grid=2,
+ columns=2,
)
generate_btn.click(
fn=generate_callback_fn,
diff --git a/src/frontend/web/image_to_image_xl_ui.py b/src/frontend/web/image_to_image_xl_ui.py
new file mode 100644
index 0000000..b1bfd48
--- /dev/null
+++ b/src/frontend/web/image_to_image_xl_ui.py
@@ -0,0 +1,125 @@
+from typing import Any
+
+import gradio as gr
+
+from backend.stablediffusion.models.scheduler_types import (
+ SchedulerType,
+ get_sampler_names,
+)
+
+random_enabled = True
+
+
+def get_image_to_image_xl_ui(generate_callback_fn: Any) -> None:
+ with gr.Blocks():
+ with gr.Row():
+ with gr.Column():
+
+ def random_seed():
+ global random_enabled
+ random_enabled = not random_enabled
+ seed_val = -1
+ if not random_enabled:
+ seed_val = 42
+
+ return gr.Number.update(
+ interactive=not random_enabled, value=seed_val
+ )
+
+ input_image = gr.Image(label="Input image", type="pil")
+ strength = gr.Slider(
+ 0.0,
+ 1.0,
+ value=0.75,
+ step=0.05,
+ label="Strength",
+ )
+
+ prompt = gr.Textbox(
+ label="Describe the image you'd like to see",
+ lines=3,
+ placeholder="A fantasy landscape",
+ )
+
+ neg_prompt = gr.Textbox(
+ label="Don't want to see",
+ lines=1,
+ placeholder="",
+ value="bad, deformed, ugly, bad anatomy",
+ )
+ with gr.Accordion("Advanced options", open=False):
+ image_height = gr.Slider(
+ 768, 2048, value=1024, step=64, label="Image Height"
+ )
+ image_width = gr.Slider(
+ 768, 2048, value=1024, step=64, label="Image Width"
+ )
+ num_inference_steps = gr.Slider(
+ 1, 100, value=20, step=1, label="Inference Steps"
+ )
+ scheduler = gr.Dropdown(
+ get_sampler_names(),
+ value=SchedulerType.UniPCMultistepScheduler.value,
+ label="Sampler",
+ )
+ guidance_scale = gr.Slider(
+ 1.0,
+ 30.0,
+ value=7.5,
+ step=0.5,
+ label="Guidance Scale",
+ )
+ num_images = gr.Slider(
+ 1,
+ 50,
+ value=1,
+ step=1,
+ label="Number of images to generate",
+ )
+ attn_slicing = gr.Checkbox(
+ label="Attention slicing (Not supported)",
+ value=False,
+ interactive=False,
+ )
+ seed = gr.Number(
+ label="Seed",
+ value=-1,
+ precision=0,
+ interactive=False,
+ )
+ seed_checkbox = gr.Checkbox(
+ label="Use random seed",
+ value=True,
+ interactive=True,
+ )
+
+ input_params = [
+ input_image,
+ strength,
+ prompt,
+ neg_prompt,
+ image_height,
+ image_width,
+ num_inference_steps,
+ scheduler,
+ guidance_scale,
+ num_images,
+ attn_slicing,
+ seed,
+ ]
+
+ with gr.Column():
+ generate_btn = gr.Button("Generate", elem_id="generate_button")
+ output = gr.Gallery(
+ label="Generated images",
+ show_label=True,
+ elem_id="gallery",
+ ).style(
+ columns=2,
+ )
+ generate_btn.click(
+ fn=generate_callback_fn,
+ inputs=input_params,
+ outputs=output,
+ )
+ seed_checkbox.change(fn=random_seed, outputs=seed)
diff --git a/src/frontend/web/image_variations_ui.py b/src/frontend/web/image_variations_ui.py
index dc91fbd..0daa00d 100644
--- a/src/frontend/web/image_variations_ui.py
+++ b/src/frontend/web/image_variations_ui.py
@@ -99,7 +99,7 @@ def random_seed():
show_label=True,
elem_id="gallery",
).style(
- grid=2,
+ columns=2,
)
generate_btn.click(
fn=generate_callback_fn,
diff --git a/src/frontend/web/instruct_pix_to_pix_ui.py b/src/frontend/web/instruct_pix_to_pix_ui.py
index ecec6cb..b99b14b 100644
--- a/src/frontend/web/instruct_pix_to_pix_ui.py
+++ b/src/frontend/web/instruct_pix_to_pix_ui.py
@@ -116,7 +116,7 @@ def random_seed():
show_label=True,
elem_id="gallery",
).style(
- grid=2,
+ columns=2,
)
generate_btn.click(
fn=generate_callback_fn,
diff --git a/src/frontend/web/text_to_image_ui.py b/src/frontend/web/text_to_image_ui.py
index 79e5f1d..5d82dd6 100644
--- a/src/frontend/web/text_to_image_ui.py
+++ b/src/frontend/web/text_to_image_ui.py
@@ -104,7 +104,7 @@ def random_seed():
show_label=True,
elem_id="gallery",
).style(
- grid=2,
+ columns=2,
)
seed_checkbox.change(fn=random_seed, outputs=seed)
generate_btn.click(
diff --git a/src/frontend/web/text_to_image_xl_ui.py b/src/frontend/web/text_to_image_xl_ui.py
new file mode 100644
index 0000000..f93cdcc
--- /dev/null
+++ b/src/frontend/web/text_to_image_xl_ui.py
@@ -0,0 +1,115 @@
+from typing import Any
+
+import gradio as gr
+
+from backend.stablediffusion.models.scheduler_types import (
+ SchedulerType,
+ get_sampler_names,
+)
+
+random_enabled = True
+
+
+def get_text_to_image_xl_ui(generate_callback_fn: Any) -> None:
+ with gr.Blocks():
+ with gr.Row():
+ with gr.Column():
+
+ def random_seed():
+ global random_enabled
+ random_enabled = not random_enabled
+ seed_val = -1
+ if not random_enabled:
+ seed_val = 42
+ return gr.Number.update(
+ interactive=not random_enabled, value=seed_val
+ )
+
+ # with gr.Row():
+ prompt = gr.Textbox(
+ label="Describe the image you'd like to see",
+ lines=3,
+ placeholder="A fantasy landscape",
+ )
+ neg_prompt = gr.Textbox(
+ label="Don't want to see",
+ lines=1,
+ placeholder="",
+ value="bad, deformed, ugly, bad anatomy",
+ )
+ with gr.Accordion("Advanced options", open=False):
+ image_height = gr.Slider(
+ 768, 2048, value=1024, step=64, label="Image Height"
+ )
+ image_width = gr.Slider(
+ 768, 2048, value=1024, step=64, label="Image Width"
+ )
+ num_inference_steps = gr.Slider(
+ 1, 100, value=20, step=1, label="Inference Steps"
+ )
+ scheduler = gr.Dropdown(
+ get_sampler_names(),
+ value=SchedulerType.LMSDiscreteScheduler.value,
+ label="Sampler",
+ )
+ guidance_scale = gr.Slider(
+ 1.0, 30.0, value=7.5, step=0.5, label="Guidance Scale"
+ )
+ num_images = gr.Slider(
+ 1,
+ 50,
+ value=1,
+ step=1,
+ label="Number of images to generate",
+ )
+ attn_slicing = gr.Checkbox(
+ label="Attention slicing (Not supported)",
+ value=False,
+ interactive=False,
+ )
+
+ vae_slicing = gr.Checkbox(
+ label="VAE slicing (Enable if low VRAM)",
+ value=True,
+ )
+ seed = gr.Number(
+ label="Seed",
+ value=-1,
+ precision=0,
+ interactive=False,
+ )
+ seed_checkbox = gr.Checkbox(
+ label="Use random seed",
+ value=True,
+ interactive=True,
+ )
+
+ input_params = [
+ prompt,
+ neg_prompt,
+ image_height,
+ image_width,
+ num_inference_steps,
+ scheduler,
+ guidance_scale,
+ num_images,
+ attn_slicing,
+ vae_slicing,
+ seed,
+ ]
+
+ with gr.Column():
+ generate_btn = gr.Button("Generate", elem_id="generate_button")
+ output = gr.Gallery(
+ label="Generated images",
+ show_label=True,
+ elem_id="gallery",
+ ).style(
+ columns=2,
+ )
+ seed_checkbox.change(fn=random_seed, outputs=seed)
+ generate_btn.click(
+ fn=generate_callback_fn,
+ inputs=input_params,
+ outputs=output,
+ )
diff --git a/src/frontend/web/ui.py b/src/frontend/web/ui.py
index 9b4f88d..10de569 100644
--- a/src/frontend/web/ui.py
+++ b/src/frontend/web/ui.py
@@ -12,6 +12,8 @@
from frontend.web.instruct_pix_to_pix_ui import get_instruct_pix_to_pix_ui
from frontend.web.settings_ui import get_settings_ui
from frontend.web.text_to_image_ui import get_text_to_image_ui
+from frontend.web.text_to_image_xl_ui import get_text_to_image_xl_ui
+from frontend.web.image_to_image_xl_ui import get_image_to_image_xl_ui
from frontend.web.controlnet.controlnet_image_ui import get_controlnet_to_image_ui
from settings import AppSettings
from utils import DiffusionMagicPaths
@@ -34,7 +36,7 @@ def diffusionmagic_web_ui(generate: Generate) -> gr.Blocks:
css=DiffusionMagicPaths.get_css_path(),
title="DiffusionMagic",
) as diffusion_magic_ui:
- gr.HTML("DiffusionMagic 2.0 Beta
")
+ gr.HTML("DiffusionMagic 3.0
")
with gr.Tabs():
if stable_diffusion_type == StableDiffusionType.base:
with gr.TabItem("Text to Image"):
@@ -76,6 +78,11 @@ def diffusionmagic_web_ui(generate: Generate) -> gr.Blocks:
elif stable_diffusion_type == StableDiffusionType.controlnet_seg:
with gr.TabItem("Controlnet Segmentation"):
get_controlnet_to_image_ui(generate.diffusion_control_to_image)
+ elif stable_diffusion_type == StableDiffusionType.stable_diffusion_xl:
+ with gr.TabItem("Text to Image SDXL"):
+ get_text_to_image_xl_ui(generate.diffusion_text_to_image_xl)
+ with gr.TabItem("Image to Image SDXL"):
+ get_image_to_image_xl_ui(generate.diffusion_image_to_image_xl)
with gr.TabItem("Settings"):
get_settings_ui()
diff --git a/src/models/configs.py b/src/models/configs.py
index 41134e8..cc0f3fd 100644
--- a/src/models/configs.py
+++ b/src/models/configs.py
@@ -3,7 +3,7 @@
class StableDiffusionModel(BaseModel):
- model_id: str = "stabilityai/stable-diffusion-2-1-base"
+ model_id: str = "stabilityai/stable-diffusion-xl-base-1.0"
use_local: bool = False