Add SDXS-512-0.9 support

rupeshs · Mar 27, 2024 · 797d84b · 797d84b
1 parent 256e1f6
commit 797d84b
Show file tree

Hide file tree

Showing 6 changed files with 31 additions and 10 deletions.
diff --git a/Readme.md b/Readme.md
@@ -9,7 +9,7 @@ The following interfaces are available :
 - WebUI
 - CLI (CommandLine Interface)
 
-🚀 Using __OpenVINO(SD Turbo)__, it took __1.7 seconds__ to create a single 512x512 image on a __Core i7-12700__.
+🚀 Using __OpenVINO(SDXS-512-0.9)__, it took __0.82 seconds__ (820 milliseconds) to create a single 512x512 image on a __Core i7-12700__.
 
 ## Supported platforms⚡️
 
@@ -19,6 +19,20 @@ The following interfaces are available :
 - Android + Termux
 - Raspberry PI 4
 
+## 🚀 Fast 1 step inference (SDXS-512-0.9)
+
+### Inference Speed
+
+Tested on Core i7-12700 to generate __512x512__ image(1 step).
+
+__SDXS-512-0.9__
+
+| Diffusion Pipeline    | Latency       |
+| --------------------- | ------------- |
+| Pytorch               | 4.8s          |
+| OpenVINO              | 3.8s          |
+| OpenVINO + TAESD      | __0.82s__     |
+
 ## 🚀 Fast 1 step inference (SD/SDXL Turbo - Adversarial Diffusion Distillation,ADD)
 
 Added support for ultra fast 1 step inference using [sdxl-turbo](https://huggingface.co/stabilityai/sdxl-turbo) model
@@ -125,6 +139,10 @@ If we enable Tiny decoder(TAESD) we can save some memory(2GB approx) for example
 - Add multilora support,thanks [monstruosoft](https://github.com/monstruosoft)
 - Add basic ControlNet v1.1 support(LCM-LoRA mode),thanks [monstruosoft](https://github.com/monstruosoft)
 - Add ControlNet annotators(Canny,Depth,LineArt,MLSD,NormalBAE,Pose,SoftEdge,Shuffle)
+- Add SDXS-512 0.9 support
+- Add SDXS-512 0.9 OpenVINO,fast 1 step inference (0.8 seconds to generate 512x512 image)
+- Default model changed to SDXS-512-0.9
+- Faster realtime image generation
 
 ## 2 Steps fast inference (LCM)
 

diff --git a/configs/lcm-models.txt b/configs/lcm-models.txt
@@ -1,3 +1,4 @@
+rupeshs/sdxs-512-0.9-orig-vae
 stabilityai/sd-turbo
 rupeshs/SDXL-Lightning-2steps
 stabilityai/sdxl-turbo

diff --git a/configs/openvino-lcm-models.txt b/configs/openvino-lcm-models.txt
@@ -1,3 +1,4 @@
+rupeshs/sdxs-512-0.9-openvino
 rupeshs/sd-turbo-openvino
 rupeshs/SDXL-Lightning-2steps-openvino-int8
 rupeshs/sdxl-turbo-openvino-int8

diff --git a/src/constants.py b/src/constants.py
@@ -1,8 +1,8 @@
 from os import environ
 
-APP_VERSION = "v1.0.0 beta 27"
-LCM_DEFAULT_MODEL = "stabilityai/sd-turbo"
-LCM_DEFAULT_MODEL_OPENVINO = "rupeshs/sd-turbo-openvino"
+APP_VERSION = "v1.0.0 beta 28"
+LCM_DEFAULT_MODEL = "rupeshs/sdxs-512-0.9-orig-vae"
+LCM_DEFAULT_MODEL_OPENVINO = "rupeshs/sdxs-512-0.9-openvino"
 APP_NAME = "FastSD CPU"
 APP_SETTINGS_FILE = "settings.yaml"
 RESULTS_DIRECTORY = "results"

diff --git a/src/frontend/gui/app_window.py b/src/frontend/gui/app_window.py
@@ -538,10 +538,10 @@ def generate_image(self):
 
         if self.config.settings.lcm_diffusion_setting.use_openvino:
             model_id = self.openvino_lcm_model_id.currentText()
+            self.config.settings.lcm_diffusion_setting.openvino_lcm_model_id = model_id
         else:
             model_id = self.lcm_model.currentText()
-
-        self.config.settings.lcm_diffusion_setting.lcm_model_id = model_id
+            self.config.settings.lcm_diffusion_setting.lcm_model_id = model_id
 
         reshape_required = False
         if self.config.settings.lcm_diffusion_setting.use_openvino:

diff --git a/src/frontend/webui/realtime_ui.py b/src/frontend/webui/realtime_ui.py
@@ -36,15 +36,16 @@ def predict(
     seed,
 ):
     lcm_diffusion_setting = LCMDiffusionSetting()
-    lcm_diffusion_setting.openvino_lcm_model_id = "rupeshs/LCM-dreamshaper-v7-openvino"
+    lcm_diffusion_setting.openvino_lcm_model_id = "rupeshs/sdxs-512-0.9-openvino"
     lcm_diffusion_setting.prompt = prompt
     lcm_diffusion_setting.guidance_scale = 1.0
     lcm_diffusion_setting.inference_steps = steps
     lcm_diffusion_setting.seed = seed
     lcm_diffusion_setting.use_seed = True
-    lcm_diffusion_setting.image_width = 256 if is_openvino_device() else 512
-    lcm_diffusion_setting.image_height = 256 if is_openvino_device() else 512
+    lcm_diffusion_setting.image_width = 512
+    lcm_diffusion_setting.image_height = 512
     lcm_diffusion_setting.use_openvino = True if is_openvino_device() else False
+    lcm_diffusion_setting.use_tiny_auto_encoder = True
     lcm_text_to_image.init(
         DEVICE,
         lcm_diffusion_setting,
@@ -116,7 +117,7 @@ def _get_footer_message() -> str:
         with gr.Accordion("Advanced options", open=False):
             steps = gr.Slider(
                 label="Steps",
-                value=4 if is_openvino_device() else 3,
+                value=1,
                 minimum=1,
                 maximum=6,
                 step=1,