Merge branch 'main' into space-txt2img

README.md

@@ -6,6 +6,7 @@ colorTo: indigo
 sdk: docker
 pinned: false
 suggested_hardware: a10g-small
+disable_embedding: true
 ---
 
 # Real-Time Latent Consistency Model
@@ -28,8 +29,9 @@ python -m venv venv
 source venv/bin/activate
 pip3 install -r requirements.txt
 cd frontend && npm install && npm run build && cd ..
-python run.py --reload --pipeline controlnet
-```
+# fastest pipeline
+python run.py --reload --pipeline img2imgSD21Turbo 
+ ```
 
 # Pipelines
 You can build your own pipeline following examples here [here](pipelines),

app.py

@@ -12,6 +12,10 @@ print("TORCH_DTYPE:", torch_dtype)
 print("PIPELINE:", args.pipeline)
 print("SAFETY_CHECKER:", args.safety_checker)
 print("TORCH_COMPILE:", args.torch_compile)
+print("SFast:", args.sfast)
+print("USE_TAESD:", args.taesd)
+print("COMPEL:", args.compel)
+print("DEBUG:", args.debug)
 
 
 app = FastAPI()

app_init.py

@@ -15,6 +15,9 @@ from types import SimpleNamespace
 from util import pil_to_frame, bytes_to_pil, is_firefox
 import asyncio
 import os
+import time
+
+THROTTLE = 1.0 / 120
 
 
 def init_app(app: FastAPI, user_data: UserData, args: Args, pipeline):
@@ -26,7 +29,7 @@ def init_app(app: FastAPI, user_data: UserData, args: Args, pipeline):
         allow_headers=["*"],
     )
 
-    @app.websocket("/ws")
+    @app.websocket("/api/ws")
     async def websocket_endpoint(websocket: WebSocket):
         await websocket.accept()
         user_count = user_data.get_user_count()
@@ -60,7 +63,7 @@ def init_app(app: FastAPI, user_data: UserData, args: Args, pipeline):
             while True:
                 data = await websocket.receive_json()
                 if data["status"] != "next_frame":
-                    asyncio.sleep(1.0 / 24)
+                    asyncio.sleep(THROTTLE)
                     continue
 
                 params = await websocket.receive_json()
@@ -85,18 +88,18 @@ def init_app(app: FastAPI, user_data: UserData, args: Args, pipeline):
                     )
                     await websocket.close()
                     return
-                await asyncio.sleep(1.0 / 24)
+                await asyncio.sleep(THROTTLE)
 
         except Exception as e:
             logging.error(f"Error: {e}")
             traceback.print_exc()
 
-    @app.get("/queue_size")
+    @app.get("/api/queue")
     async def get_queue_size():
         queue_size = user_data.get_user_count()
         return JSONResponse({"queue_size": queue_size})
 
-    @app.get("/stream/{user_id}")
+    @app.get("/api/stream/{user_id}")
     async def stream(user_id: uuid.UUID, request: Request):
         try:
             print(f"New stream request: {user_id}")
@@ -105,14 +108,15 @@ def init_app(app: FastAPI, user_data: UserData, args: Args, pipeline):
                 websocket = user_data.get_websocket(user_id)
                 last_params = SimpleNamespace()
                 while True:
+                    last_time = time.time()
                     params = await user_data.get_latest_data(user_id)
                     if not vars(params) or params.__dict__ == last_params.__dict__:
                         await websocket.send_json({"status": "send_frame"})
-                        await asyncio.sleep(0.1)
                         continue
 
                     last_params = params
                     image = pipeline.predict(params)
+
                     if image is None:
                         await websocket.send_json({"status": "send_frame"})
                         continue
@@ -122,6 +126,8 @@ def init_app(app: FastAPI, user_data: UserData, args: Args, pipeline):
                     if not is_firefox(request.headers["user-agent"]):
                         yield frame
                     await websocket.send_json({"status": "send_frame"})
+                    if args.debug:
+                        print(f"Time taken: {time.time() - last_time}")
 
             return StreamingResponse(
                 generate(),
@@ -134,7 +140,7 @@ def init_app(app: FastAPI, user_data: UserData, args: Args, pipeline):
             return HTTPException(status_code=404, detail="User not found")
 
     # route to setup frontend
-    @app.get("/settings")
+    @app.get("/api/settings")
     async def settings():
         info_schema = pipeline.Info.schema()
         info = pipeline.Info()

build-run.sh

@@ -13,4 +13,4 @@ if [ -z ${PIPELINE+x} ]; then
     PIPELINE="controlnet"
 fi
 echo -e "\033[1;32m\npipeline: $PIPELINE \033[0m"
-python3 run.py --port 7860 --host 0.0.0.0 --pipeline $PIPELINE
+python3 run.py --port 7860 --host 0.0.0.0 --pipeline $PIPELINE --sfast

config.py

@@ -12,17 +12,20 @@ class Args(NamedTuple):
     timeout: float
     safety_checker: bool
     torch_compile: bool
-    use_taesd: bool
+    taesd: bool
     pipeline: str
     ssl_certfile: str
     ssl_keyfile: str
+    sfast: bool
+    compel: bool = False
+    debug: bool = False
 
 
 MAX_QUEUE_SIZE = int(os.environ.get("MAX_QUEUE_SIZE", 0))
 TIMEOUT = float(os.environ.get("TIMEOUT", 0))
 SAFETY_CHECKER = os.environ.get("SAFETY_CHECKER", None) == "True"
 TORCH_COMPILE = os.environ.get("TORCH_COMPILE", None) == "True"
-USE_TAESD = os.environ.get("USE_TAESD", None) == "True"
+USE_TAESD = os.environ.get("USE_TAESD", "True") == "True"
 default_host = os.getenv("HOST", "0.0.0.0")
 default_port = int(os.getenv("PORT", "7860"))
 default_mode = os.getenv("MODE", "default")
@@ -36,7 +39,7 @@ parser.add_argument(
 )
 parser.add_argument(
     "--max-queue-size",
-    "--max_queue_size",
+    dest="max_queue_size",
     type=int,
     default=MAX_QUEUE_SIZE,
     help="Max Queue Size",
@@ -44,23 +47,28 @@ parser.add_argument(
 parser.add_argument("--timeout", type=float, default=TIMEOUT, help="Timeout")
 parser.add_argument(
     "--safety-checker",
-    "--safety_checker",
-    type=bool,
+    dest="safety_checker",
+    action="store_true",
     default=SAFETY_CHECKER,
     help="Safety Checker",
 )
 parser.add_argument(
     "--torch-compile",
-    "--torch_compile",
-    type=bool,
+    dest="torch_compile",
+    action="store_true",
     default=TORCH_COMPILE,
     help="Torch Compile",
 )
 parser.add_argument(
-    "--use-taesd",
-    "--use_taesd",
-    type=bool,
-    default=USE_TAESD,
+    "--taesd",
+    dest="taesd",
+    action="store_true",
+    help="Use Tiny Autoencoder",
+)
+parser.add_argument(
+    "--no-taesd",
+    dest="taesd",
+    action="store_false",
     help="Use Tiny Autoencoder",
 )
 parser.add_argument(
@@ -71,17 +79,36 @@ parser.add_argument(
 )
 parser.add_argument(
     "--ssl-certfile",
-    "--ssl_certfile",
+    dest="ssl_certfile",
     type=str,
     default=None,
     help="SSL certfile",
 )
 parser.add_argument(
     "--ssl-keyfile",
-    "--ssl_keyfile",
+    dest="ssl_keyfile",
     type=str,
     default=None,
     help="SSL keyfile",
 )
+parser.add_argument(
+    "--debug",
+    action="store_true",
+    default=False,
+    help="Debug",
+)
+parser.add_argument(
+    "--compel",
+    action="store_true",
+    default=False,
+    help="Compel",
+)
+parser.add_argument(
+    "--sfast",
+    action="store_true",
+    default=False,
+    help="Enable Stable Fast",
+)
+parser.set_defaults(taesd=USE_TAESD)
 
 args = Args(**vars(parser.parse_args()))

frontend/src/lib/components/ImagePlayer.svelte

@@ -26,7 +26,11 @@
 >
   <!-- svelte-ignore a11y-missing-attribute -->
   {#if isLCMRunning}
-    <img bind:this={imageEl} class="aspect-square w-full rounded-lg" src={'/stream/' + $streamId} />
+    <img
+      bind:this={imageEl}
+      class="aspect-square w-full rounded-lg"
+      src={'/api/stream/' + $streamId}
+    />
     <div class="absolute bottom-1 right-1">
       <Button
         on:click={takeSnapshot}

frontend/src/lib/components/VideoInput.svelte

@@ -20,7 +20,7 @@
   let videoFrameCallbackId: number;
 
   // ajust the throttle time to your needs
-  const THROTTLE_TIME = 1000 / 15;
+  const THROTTLE = 1000 / 120;
   let selectedDevice: string = '';
   let videoIsReady = false;
 
@@ -41,7 +41,7 @@
   }
   let lastMillis = 0;
   async function onFrameChange(now: DOMHighResTimeStamp, metadata: VideoFrameCallbackMetadata) {
-    if (now - lastMillis < THROTTLE_TIME) {
+    if (now - lastMillis < THROTTLE) {
       videoFrameCallbackId = videoEl.requestVideoFrameCallback(onFrameChange);
       return;
     }

frontend/src/lib/lcmLive.ts

@@ -20,7 +20,7 @@ export const lcmLiveActions = {
 
             try {
                 const websocketURL = `${window.location.protocol === "https:" ? "wss" : "ws"
-                    }:${window.location.host}/ws`;
+                    }:${window.location.host}/api/ws`;
 
                 websocket = new WebSocket(websocketURL);
                 websocket.onopen = () => {

frontend/src/routes/+page.svelte

@@ -17,28 +17,35 @@
   let isImageMode: boolean = false;
   let maxQueueSize: number = 0;
   let currentQueueSize: number = 0;
+  let queueCheckerRunning: boolean = false;
+
   onMount(() => {
     getSettings();
   });
 
   async function getSettings() {
-    const settings = await fetch('/settings').then((r) => r.json());
+    const settings = await fetch('/api/settings').then((r) => r.json());
     pipelineParams = settings.input_params.properties;
     pipelineInfo = settings.info.properties;
     isImageMode = pipelineInfo.input_mode.default === PipelineMode.IMAGE;
     maxQueueSize = settings.max_queue_size;
     pageContent = settings.page_content;
     console.log(pipelineParams);
-    if (maxQueueSize > 0) {
+    toggleQueueChecker(true);
+  }
+  function toggleQueueChecker(start: boolean) {
+    queueCheckerRunning = start && maxQueueSize > 0;
+    if (start) {
       getQueueSize();
-      setInterval(() => {
-        getQueueSize();
-      }, 2000);
     }
   }
   async function getQueueSize() {
-    const data = await fetch('/queue_size').then((r) => r.json());
+    if (!queueCheckerRunning) {
+      return;
+    }
+    const data = await fetch('/api/queue').then((r) => r.json());
     currentQueueSize = data.queue_size;
+    setTimeout(getQueueSize, 10000);
   }
 
   function getSreamdata() {
@@ -61,11 +68,13 @@
       disabled = true;
       await lcmLiveActions.start(getSreamdata);
       disabled = false;
+      toggleQueueChecker(false);
     } else {
       if (isImageMode) {
         mediaStreamActions.stop();
       }
       lcmLiveActions.stop();
+      toggleQueueChecker(true);
     }
   }
 </script>

frontend/vite.config.ts

@@ -5,8 +5,8 @@ export default defineConfig({
   plugins: [sveltekit()],
   server: {
     proxy: {
-      '^/settings|/queue_size|/stream': 'http://localhost:7860',
-      '/ws': {
+      '/api': 'http://localhost:7860',
+      '/api/ws': {
         target: 'ws://localhost:7860',
         ws: true
       }

pipelines/controlnet.py

@@ -69,18 +69,18 @@ class Pipeline:
             2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
         )
         steps: int = Field(
-            4, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
+            2, min=1, max=6, title="Steps", field="range", hide=True, id="steps"
         )
         width: int = Field(
-            768, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
+            512, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
         )
         height: int = Field(
-            768, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
+            512, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
         )
         guidance_scale: float = Field(
-            0.2,
+            0.0,
             min=0,
-            max=5,
+            max=2,
             step=0.001,
             title="Guidance Scale",
             field="range",
@@ -169,20 +169,29 @@ class Pipeline:
                 safety_checker=None,
                 controlnet=controlnet_canny,
             )
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
+
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
         self.canny_torch = SobelOperator(device=device)
         self.pipe.set_progress_bar_config(disable=True)
         self.pipe.to(device=device, dtype=torch_dtype)
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        # check if computer has less than 64GB of RAM using sys or os
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
-
         if args.torch_compile:
             self.pipe.unet = torch.compile(
                 self.pipe.unet, mode="reduce-overhead", fullgraph=True
@@ -196,16 +205,21 @@ class Pipeline:
                 image=[Image.new("RGB", (768, 768))],
                 control_image=[Image.new("RGB", (768, 768))],
             )
-
-        self.compel_proc = Compel(
-            tokenizer=self.pipe.tokenizer,
-            text_encoder=self.pipe.text_encoder,
-            truncate_long_prompts=False,
-        )
+        if args.compel:
+            self.compel_proc = Compel(
+                tokenizer=self.pipe.tokenizer,
+                text_encoder=self.pipe.text_encoder,
+                truncate_long_prompts=False,
+            )
 
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
-        prompt_embeds = self.compel_proc(params.prompt)
+        prompt_embeds = None
+        prompt = params.prompt
+        if hasattr(self, "compel_proc"):
+            prompt_embeds = self.compel_proc(params.prompt)
+            prompt = None
+
         control_image = self.canny_torch(
             params.image, params.canny_low_threshold, params.canny_high_threshold
         )
@@ -218,6 +232,7 @@ class Pipeline:
             image=params.image,
             control_image=control_image,
             prompt_embeds=prompt_embeds,
+            prompt=prompt,
             generator=generator,
             strength=strength,
             num_inference_steps=steps,

pipelines/controlnetLoraSD15.py

@@ -81,7 +81,7 @@ class Pipeline:
             2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
         )
         steps: int = Field(
-            4, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
+            1, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
         )
         width: int = Field(
             768, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
@@ -90,7 +90,7 @@ class Pipeline:
             768, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
         )
         guidance_scale: float = Field(
-            0.2,
+            1.0,
             min=0,
             max=2,
             step=0.001,
@@ -180,7 +180,7 @@ class Pipeline:
                     base_model_id,
                     controlnet=controlnet_canny,
                 )
-            self.pipes[base_model_id] = pipe
+                self.pipes[base_model_id] = pipe
         else:
             for base_model_id in base_models.keys():
                 pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
@@ -195,25 +195,23 @@ class Pipeline:
         for pipe in self.pipes.values():
             pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
             pipe.set_progress_bar_config(disable=True)
-            pipe.to(device=device, dtype=torch_dtype).to(device)
             if device.type != "mps":
                 pipe.unet.to(memory_format=torch.channels_last)
 
-            if psutil.virtual_memory().total < 64 * 1024**3:
-                pipe.enable_attention_slicing()
-
-            if args.use_taesd:
+            if args.taesd:
                 pipe.vae = AutoencoderTiny.from_pretrained(
                     taesd_model, torch_dtype=torch_dtype, use_safetensors=True
                 ).to(device)
 
             # Load LCM LoRA
             pipe.load_lora_weights(lcm_lora_id, adapter_name="lcm")
-            pipe.compel_proc = Compel(
-                tokenizer=pipe.tokenizer,
-                text_encoder=pipe.text_encoder,
-                truncate_long_prompts=False,
-            )
+            pipe.to(device=device, dtype=torch_dtype).to(device)
+            if args.compel:
+                self.compel_proc = Compel(
+                    tokenizer=pipe.tokenizer,
+                    text_encoder=pipe.text_encoder,
+                    truncate_long_prompts=False,
+                )
             if args.torch_compile:
                 pipe.unet = torch.compile(
                     pipe.unet, mode="reduce-overhead", fullgraph=True
@@ -233,7 +231,12 @@ class Pipeline:
 
         activation_token = base_models[params.base_model_id]
         prompt = f"{activation_token} {params.prompt}"
-        prompt_embeds = pipe.compel_proc(prompt)
+        prompt_embeds = None
+        prompt = params.prompt
+        if hasattr(self, "compel_proc"):
+            prompt_embeds = self.compel_proc(prompt)
+            prompt = None
+
         control_image = self.canny_torch(
             params.image, params.canny_low_threshold, params.canny_high_threshold
         )
@@ -245,6 +248,7 @@ class Pipeline:
         results = pipe(
             image=params.image,
             control_image=control_image,
+            prompt=prompt,
             prompt_embeds=prompt_embeds,
             generator=generator,
             strength=strength,

pipelines/controlnetLoraSDXL.py

@@ -80,7 +80,7 @@ class Pipeline:
             2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
         )
         steps: int = Field(
-            2, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
+            1, min=1, max=10, title="Steps", field="range", hide=True, id="steps"
         )
         width: int = Field(
             1024, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
@@ -91,7 +91,7 @@ class Pipeline:
         guidance_scale: float = Field(
             1.0,
             min=0,
-            max=20,
+            max=2.0,
             step=0.001,
             title="Guidance Scale",
             field="range",
@@ -199,19 +199,31 @@ class Pipeline:
         self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
         self.pipe.set_progress_bar_config(disable=True)
         self.pipe.to(device=device, dtype=torch_dtype).to(device)
+
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
+        if args.compel:
+            self.pipe.compel_proc = Compel(
+                tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
+                text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
+                returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
+                requires_pooled=[False, True],
+            )
 
-        self.pipe.compel_proc = Compel(
-            tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
-            text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
-            returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
-            requires_pooled=[False, True],
-        )
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
@@ -232,9 +244,23 @@ class Pipeline:
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
 
-        prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
-            [params.prompt, params.negative_prompt]
-        )
+        prompt = params.prompt
+        negative_prompt = params.negative_prompt
+        prompt_embeds = None
+        pooled_prompt_embeds = None
+        negative_prompt_embeds = None
+        negative_pooled_prompt_embeds = None
+        if hasattr(self.pipe, "compel_proc"):
+            _prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
+                [params.prompt, params.negative_prompt]
+            )
+            prompt = None
+            negative_prompt = None
+            prompt_embeds = _prompt_embeds[0:1]
+            pooled_prompt_embeds = pooled_prompt_embeds[0:1]
+            negative_prompt_embeds = _prompt_embeds[1:2]
+            negative_pooled_prompt_embeds = pooled_prompt_embeds[1:2]
+
         control_image = self.canny_torch(
             params.image, params.canny_low_threshold, params.canny_high_threshold
         )
@@ -246,10 +272,12 @@ class Pipeline:
         results = self.pipe(
             image=params.image,
             control_image=control_image,
-            prompt_embeds=prompt_embeds[0:1],
-            pooled_prompt_embeds=pooled_prompt_embeds[0:1],
-            negative_prompt_embeds=prompt_embeds[1:2],
-            negative_pooled_prompt_embeds=pooled_prompt_embeds[1:2],
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
             generator=generator,
             strength=strength,
             num_inference_steps=steps,

pipelines/{controlnelSD21Turbo.py → controlnetSDTurbo.py}

@@ -100,7 +100,7 @@ class Pipeline:
             id="strength",
         )
         controlnet_scale: float = Field(
-            0.2,
+            0.325,
             min=0,
             max=1.0,
             step=0.001,
@@ -176,10 +176,23 @@ class Pipeline:
                 safety_checker=None,
             )
 
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
+
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
         self.canny_torch = SobelOperator(device=device)
 
         self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
@@ -188,15 +201,16 @@ class Pipeline:
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
+        if args.compel:
+            from compel import Compel
 
-        self.pipe.compel_proc = Compel(
-            tokenizer=self.pipe.tokenizer,
-            text_encoder=self.pipe.text_encoder,
-            truncate_long_prompts=True,
-        )
-        if args.use_taesd:
+            self.pipe.compel_proc = Compel(
+                tokenizer=self.pipe.tokenizer,
+                text_encoder=self.pipe.text_encoder,
+                truncate_long_prompts=True,
+            )
+
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
@@ -216,7 +230,13 @@ class Pipeline:
 
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
-        prompt_embeds = self.pipe.compel_proc(params.prompt)
+        prompt = params.prompt
+        prompt_embeds = None
+        if hasattr(self.pipe, "compel_proc"):
+            prompt_embeds = self.pipe.compel_proc(
+                [params.prompt, params.negative_prompt]
+            )
+            prompt = None
         control_image = self.canny_torch(
             params.image, params.canny_low_threshold, params.canny_high_threshold
         )
@@ -224,10 +244,10 @@ class Pipeline:
         strength = params.strength
         if int(steps * strength) < 1:
             steps = math.ceil(1 / max(0.10, strength))
-        last_time = time.time()
         results = self.pipe(
             image=params.image,
             control_image=control_image,
+            prompt=prompt,
             prompt_embeds=prompt_embeds,
             generator=generator,
             strength=strength,
@@ -240,8 +260,6 @@ class Pipeline:
             control_guidance_start=params.controlnet_start,
             control_guidance_end=params.controlnet_end,
         )
-        print(f"Time taken: {time.time() - last_time}")
-
         nsfw_content_detected = (
             results.nsfw_content_detected[0]
             if "nsfw_content_detected" in results

pipelines/controlnetSDXLTurbo.py

@@ -185,21 +185,32 @@ class Pipeline:
             )
         self.canny_torch = SobelOperator(device=device)
 
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
         self.pipe.set_progress_bar_config(disable=True)
         self.pipe.to(device=device, dtype=torch_dtype).to(device)
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
+        if args.compel:
+            self.pipe.compel_proc = Compel(
+                tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
+                text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
+                returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
+                requires_pooled=[False, True],
+            )
 
-        self.pipe.compel_proc = Compel(
-            tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
-            text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
-            returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
-            requires_pooled=[False, True],
-        )
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
@@ -220,9 +231,23 @@ class Pipeline:
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
 
-        prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
-            [params.prompt, params.negative_prompt]
-        )
+        prompt = params.prompt
+        negative_prompt = params.negative_prompt
+        prompt_embeds = None
+        pooled_prompt_embeds = None
+        negative_prompt_embeds = None
+        negative_pooled_prompt_embeds = None
+        if hasattr(self.pipe, "compel_proc"):
+            _prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
+                [params.prompt, params.negative_prompt]
+            )
+            prompt = None
+            negative_prompt = None
+            prompt_embeds = _prompt_embeds[0:1]
+            pooled_prompt_embeds = pooled_prompt_embeds[0:1]
+            negative_prompt_embeds = _prompt_embeds[1:2]
+            negative_pooled_prompt_embeds = pooled_prompt_embeds[1:2]
+
         control_image = self.canny_torch(
             params.image, params.canny_low_threshold, params.canny_high_threshold
         )
@@ -234,10 +259,12 @@ class Pipeline:
         results = self.pipe(
             image=params.image,
             control_image=control_image,
-            prompt_embeds=prompt_embeds[0:1],
-            pooled_prompt_embeds=pooled_prompt_embeds[0:1],
-            negative_prompt_embeds=prompt_embeds[1:2],
-            negative_pooled_prompt_embeds=pooled_prompt_embeds[1:2],
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
             generator=generator,
             strength=strength,
             num_inference_steps=steps,

pipelines/controlnetSegmindVegaRT.py

@@ -0,0 +1,303 @@
+from diffusers import (
+    StableDiffusionXLControlNetImg2ImgPipeline,
+    ControlNetModel,
+    AutoencoderKL,
+    AutoencoderTiny,
+    LCMScheduler,
+)
+from compel import Compel, ReturnedEmbeddingsType
+import torch
+from pipelines.utils.canny_gpu import SobelOperator
+
+try:
+    import intel_extension_for_pytorch as ipex  # type: ignore
+except:
+    pass
+
+import psutil
+from config import Args
+from pydantic import BaseModel, Field
+from PIL import Image
+import math
+
+controlnet_model = "diffusers/controlnet-canny-sdxl-1.0"
+base_model = "segmind/Segmind-Vega"
+lora_model = "segmind/Segmind-VegaRT"
+taesd_model = "madebyollin/taesdxl"
+
+default_prompt = "Portrait of The Terminator with , glare pose, detailed, intricate, full of colour, cinematic lighting, trending on artstation, 8k, hyperrealistic, focused, extreme details, unreal engine 5 cinematic, masterpiece"
+default_negative_prompt = "blurry, low quality, render, 3D, oversaturated"
+page_content = """
+<h1 class="text-3xl font-bold">Real-Time SegmindVegaRT</h1>
+<h3 class="text-xl font-bold">Image-to-Image ControlNet</h3>
+<p class="text-sm">
+    This demo showcases
+    <a
+    href="https://huggingface.co/segmind/Segmind-VegaRT"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">Segmind-VegaRT</a>
+Image to Image pipeline using
+    <a
+    href="https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl_turbo"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">Diffusers</a
+    > with a MJPEG stream server.
+</p>
+<p class="text-sm text-gray-500">
+    Change the prompt to generate different images, accepts <a
+    href="https://github.com/damian0815/compel/blob/main/doc/syntax.md"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">Compel</a
+    > syntax.
+</p>
+"""
+
+
+class Pipeline:
+    class Info(BaseModel):
+        name: str = "controlnet+SegmindVegaRT"
+        title: str = "SegmindVegaRT + Controlnet"
+        description: str = "Generates an image from a text prompt"
+        input_mode: str = "image"
+        page_content: str = page_content
+
+    class InputParams(BaseModel):
+        prompt: str = Field(
+            default_prompt,
+            title="Prompt",
+            field="textarea",
+            id="prompt",
+        )
+        negative_prompt: str = Field(
+            default_negative_prompt,
+            title="Negative Prompt",
+            field="textarea",
+            id="negative_prompt",
+            hide=True,
+        )
+        seed: int = Field(
+            2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
+        )
+        steps: int = Field(
+            2, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
+        )
+        width: int = Field(
+            1024, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
+        )
+        height: int = Field(
+            1024, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
+        )
+        guidance_scale: float = Field(
+            0.0,
+            min=0,
+            max=1,
+            step=0.001,
+            title="Guidance Scale",
+            field="range",
+            hide=True,
+            id="guidance_scale",
+        )
+        strength: float = Field(
+            0.5,
+            min=0.25,
+            max=1.0,
+            step=0.001,
+            title="Strength",
+            field="range",
+            hide=True,
+            id="strength",
+        )
+        controlnet_scale: float = Field(
+            0.5,
+            min=0,
+            max=1.0,
+            step=0.001,
+            title="Controlnet Scale",
+            field="range",
+            hide=True,
+            id="controlnet_scale",
+        )
+        controlnet_start: float = Field(
+            0.0,
+            min=0,
+            max=1.0,
+            step=0.001,
+            title="Controlnet Start",
+            field="range",
+            hide=True,
+            id="controlnet_start",
+        )
+        controlnet_end: float = Field(
+            1.0,
+            min=0,
+            max=1.0,
+            step=0.001,
+            title="Controlnet End",
+            field="range",
+            hide=True,
+            id="controlnet_end",
+        )
+        canny_low_threshold: float = Field(
+            0.31,
+            min=0,
+            max=1.0,
+            step=0.001,
+            title="Canny Low Threshold",
+            field="range",
+            hide=True,
+            id="canny_low_threshold",
+        )
+        canny_high_threshold: float = Field(
+            0.125,
+            min=0,
+            max=1.0,
+            step=0.001,
+            title="Canny High Threshold",
+            field="range",
+            hide=True,
+            id="canny_high_threshold",
+        )
+        debug_canny: bool = Field(
+            False,
+            title="Debug Canny",
+            field="checkbox",
+            hide=True,
+            id="debug_canny",
+        )
+
+    def __init__(self, args: Args, device: torch.device, torch_dtype: torch.dtype):
+        controlnet_canny = ControlNetModel.from_pretrained(
+            controlnet_model,
+            torch_dtype=torch_dtype,
+        ).to(device)
+        vae = AutoencoderKL.from_pretrained(
+            "madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch_dtype
+        )
+        if args.safety_checker:
+            self.pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
+                base_model,
+                controlnet=controlnet_canny,
+                vae=vae,
+            )
+        else:
+            self.pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
+                base_model,
+                safety_checker=None,
+                controlnet=controlnet_canny,
+                vae=vae,
+            )
+        self.canny_torch = SobelOperator(device=device)
+
+        self.pipe.load_lora_weights(lora_model)
+        self.pipe.fuse_lora()
+        self.pipe.scheduler = LCMScheduler.from_pretrained(
+            base_model, subfolder="scheduler"
+        )
+
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
+        self.pipe.set_progress_bar_config(disable=True)
+        self.pipe.to(device=device, dtype=torch_dtype).to(device)
+        if device.type != "mps":
+            self.pipe.unet.to(memory_format=torch.channels_last)
+
+        if args.compel:
+            self.pipe.compel_proc = Compel(
+                tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
+                text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
+                returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
+                requires_pooled=[False, True],
+            )
+        if args.taesd:
+            self.pipe.vae = AutoencoderTiny.from_pretrained(
+                taesd_model, torch_dtype=torch_dtype, use_safetensors=True
+            ).to(device)
+
+        if args.torch_compile:
+            self.pipe.unet = torch.compile(
+                self.pipe.unet, mode="reduce-overhead", fullgraph=True
+            )
+            self.pipe.vae = torch.compile(
+                self.pipe.vae, mode="reduce-overhead", fullgraph=True
+            )
+            self.pipe(
+                prompt="warmup",
+                image=[Image.new("RGB", (768, 768))],
+                control_image=[Image.new("RGB", (768, 768))],
+            )
+
+    def predict(self, params: "Pipeline.InputParams") -> Image.Image:
+        generator = torch.manual_seed(params.seed)
+
+        prompt = params.prompt
+        negative_prompt = params.negative_prompt
+        prompt_embeds = None
+        pooled_prompt_embeds = None
+        negative_prompt_embeds = None
+        negative_pooled_prompt_embeds = None
+        if hasattr(self.pipe, "compel_proc"):
+            _prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
+                [params.prompt, params.negative_prompt]
+            )
+            prompt = None
+            negative_prompt = None
+            prompt_embeds = _prompt_embeds[0:1]
+            pooled_prompt_embeds = pooled_prompt_embeds[0:1]
+            negative_prompt_embeds = _prompt_embeds[1:2]
+            negative_pooled_prompt_embeds = pooled_prompt_embeds[1:2]
+
+        control_image = self.canny_torch(
+            params.image, params.canny_low_threshold, params.canny_high_threshold
+        )
+        steps = params.steps
+        strength = params.strength
+        if int(steps * strength) < 1:
+            steps = math.ceil(1 / max(0.10, strength))
+
+        results = self.pipe(
+            image=params.image,
+            control_image=control_image,
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            generator=generator,
+            strength=strength,
+            num_inference_steps=steps,
+            guidance_scale=params.guidance_scale,
+            width=params.width,
+            height=params.height,
+            output_type="pil",
+            controlnet_conditioning_scale=params.controlnet_scale,
+            control_guidance_start=params.controlnet_start,
+            control_guidance_end=params.controlnet_end,
+        )
+
+        nsfw_content_detected = (
+            results.nsfw_content_detected[0]
+            if "nsfw_content_detected" in results
+            else False
+        )
+        if nsfw_content_detected:
+            return None
+        result_image = results.images[0]
+        if params.debug_canny:
+            # paste control_image on top of result_image
+            w0, h0 = (200, 200)
+            control_image = control_image.resize((w0, h0))
+            w1, h1 = result_image.size
+            result_image.paste(control_image, (w1 - w0, h1 - h0))
+
+        return result_image

pipelines/img2img.py

@@ -102,20 +102,28 @@ class Pipeline:
                 base_model,
                 safety_checker=None,
             )
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
 
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
         self.pipe.set_progress_bar_config(disable=True)
         self.pipe.to(device=device, dtype=torch_dtype)
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        # check if computer has less than 64GB of RAM using sys or os
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
-
         if args.torch_compile:
             print("Running torch compile")
             self.pipe.unet = torch.compile(
@@ -130,15 +138,20 @@ class Pipeline:
                 image=[Image.new("RGB", (768, 768))],
             )
 
-        self.compel_proc = Compel(
-            tokenizer=self.pipe.tokenizer,
-            text_encoder=self.pipe.text_encoder,
-            truncate_long_prompts=False,
-        )
+        if args.compel:
+            self.compel_proc = Compel(
+                tokenizer=self.pipe.tokenizer,
+                text_encoder=self.pipe.text_encoder,
+                truncate_long_prompts=False,
+            )
 
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
-        prompt_embeds = self.compel_proc(params.prompt)
+        prompt_embeds = None
+        prompt = params.prompt
+        if hasattr(self, "compel_proc"):
+            prompt_embeds = self.compel_proc(params.prompt)
+            prompt = None
 
         steps = params.steps
         strength = params.strength
@@ -147,6 +160,7 @@ class Pipeline:
 
         results = self.pipe(
             image=params.image,
+            prompt=prompt,
             prompt_embeds=prompt_embeds,
             generator=generator,
             strength=strength,

pipelines/img2imgSDTurbo.py

@@ -0,0 +1,188 @@
+from diffusers import (
+    AutoPipelineForImage2Image,
+    AutoencoderTiny,
+)
+import torch
+
+try:
+    import intel_extension_for_pytorch as ipex  # type: ignore
+except:
+    pass
+
+import psutil
+from config import Args
+from pydantic import BaseModel, Field
+from PIL import Image
+import math
+from sfast.compilers.stable_diffusion_pipeline_compiler import (
+    compile,
+    CompilationConfig,
+)
+
+base_model = "stabilityai/sd-turbo"
+taesd_model = "madebyollin/taesd"
+
+default_prompt = "close-up photography of old man standing in the rain at night, in a street lit by lamps, leica 35mm summilux"
+default_negative_prompt = "blurry, low quality, render, 3D, oversaturated"
+page_content = """
+<h1 class="text-3xl font-bold">Real-Time SD-Turbo</h1>
+<h3 class="text-xl font-bold">Image-to-Image</h3>
+<p class="text-sm">
+    This demo showcases
+    <a
+    href="https://huggingface.co/stabilityai/sdxl-turbo"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">SDXL Turbo</a>
+Image to Image pipeline using
+    <a
+    href="https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl_turbo"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">Diffusers</a
+    > with a MJPEG stream server.
+</p>
+<p class="text-sm text-gray-500">
+    Change the prompt to generate different images, accepts <a
+    href="https://github.com/damian0815/compel/blob/main/doc/syntax.md"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">Compel</a
+    > syntax.
+</p>
+"""
+
+
+class Pipeline:
+    class Info(BaseModel):
+        name: str = "img2img"
+        title: str = "Image-to-Image SDXL"
+        description: str = "Generates an image from a text prompt"
+        input_mode: str = "image"
+        page_content: str = page_content
+
+    class InputParams(BaseModel):
+        prompt: str = Field(
+            default_prompt,
+            title="Prompt",
+            field="textarea",
+            id="prompt",
+        )
+        negative_prompt: str = Field(
+            default_negative_prompt,
+            title="Negative Prompt",
+            field="textarea",
+            id="negative_prompt",
+            hide=True,
+        )
+        seed: int = Field(
+            2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
+        )
+        steps: int = Field(
+            1, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
+        )
+        width: int = Field(
+            512, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
+        )
+        height: int = Field(
+            512, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
+        )
+        strength: float = Field(
+            0.5,
+            min=0.25,
+            max=1.0,
+            step=0.001,
+            title="Strength",
+            field="range",
+            hide=True,
+            id="strength",
+        )
+
+    def __init__(self, args: Args, device: torch.device, torch_dtype: torch.dtype):
+        if args.safety_checker:
+            self.pipe = AutoPipelineForImage2Image.from_pretrained(base_model)
+        else:
+            self.pipe = AutoPipelineForImage2Image.from_pretrained(
+                base_model,
+                safety_checker=None,
+            )
+        if args.taesd:
+            self.pipe.vae = AutoencoderTiny.from_pretrained(
+                taesd_model, torch_dtype=torch_dtype, use_safetensors=True
+            ).to(device)
+
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
+        self.pipe.set_progress_bar_config(disable=True)
+        self.pipe.to(device=device, dtype=torch_dtype)
+        if device.type != "mps":
+            self.pipe.unet.to(memory_format=torch.channels_last)
+
+        if args.torch_compile:
+            print("Running torch compile")
+            self.pipe.unet = torch.compile(
+                self.pipe.unet, mode="reduce-overhead", fullgraph=True
+            )
+            self.pipe.vae = torch.compile(
+                self.pipe.vae, mode="reduce-overhead", fullgraph=True
+            )
+
+            self.pipe(
+                prompt="warmup",
+                image=[Image.new("RGB", (768, 768))],
+            )
+        if args.compel:
+            from compel import Compel
+
+            self.pipe.compel_proc = Compel(
+                tokenizer=self.pipe.tokenizer,
+                text_encoder=self.pipe.text_encoder,
+                truncate_long_prompts=True,
+            )
+
+    def predict(self, params: "Pipeline.InputParams") -> Image.Image:
+        generator = torch.manual_seed(params.seed)
+        steps = params.steps
+        strength = params.strength
+        if int(steps * strength) < 1:
+            steps = math.ceil(1 / max(0.10, strength))
+
+        prompt = params.prompt
+        prompt_embeds = None
+        if hasattr(self.pipe, "compel_proc"):
+            prompt_embeds = self.pipe.compel_proc(
+                [params.prompt, params.negative_prompt]
+            )
+            prompt = None
+
+        results = self.pipe(
+            image=params.image,
+            prompt_embeds=prompt_embeds,
+            prompt=prompt,
+            negative_prompt=params.negative_prompt,
+            generator=generator,
+            strength=strength,
+            num_inference_steps=steps,
+            guidance_scale=1.1,
+            width=params.width,
+            height=params.height,
+            output_type="pil",
+        )
+
+        nsfw_content_detected = (
+            results.nsfw_content_detected[0]
+            if "nsfw_content_detected" in results
+            else False
+        )
+        if nsfw_content_detected:
+            return None
+        result_image = results.images[0]
+
+        return result_image

pipelines/img2imgSDXLTurbo.py

@@ -73,18 +73,18 @@ class Pipeline:
             2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
         )
         steps: int = Field(
-            4, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
+            1, min=1, max=10, title="Steps", field="range", hide=True, id="steps"
         )
         width: int = Field(
-            512, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
+            768, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
         )
         height: int = Field(
-            512, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
+            768, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
         )
         guidance_scale: float = Field(
-            0.2,
+            1.0,
             min=0,
-            max=20,
+            max=1,
             step=0.001,
             title="Guidance Scale",
             field="range",
@@ -110,20 +110,28 @@ class Pipeline:
                 base_model,
                 safety_checker=None,
             )
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
 
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
         self.pipe.set_progress_bar_config(disable=True)
         self.pipe.to(device=device, dtype=torch_dtype)
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        # check if computer has less than 64GB of RAM using sys or os
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
-
         if args.torch_compile:
             print("Running torch compile")
             self.pipe.unet = torch.compile(
@@ -132,24 +140,38 @@ class Pipeline:
             self.pipe.vae = torch.compile(
                 self.pipe.vae, mode="reduce-overhead", fullgraph=True
             )
-
             self.pipe(
                 prompt="warmup",
                 image=[Image.new("RGB", (768, 768))],
             )
 
-        self.pipe.compel_proc = Compel(
-            tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
-            text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
-            returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
-            requires_pooled=[False, True],
-        )
+        if args.compel:
+            self.pipe.compel_proc = Compel(
+                tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
+                text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
+                returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
+                requires_pooled=[False, True],
+            )
 
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
-        prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
-            [params.prompt, params.negative_prompt]
-        )
+        prompt = params.prompt
+        negative_prompt = params.negative_prompt
+        prompt_embeds = None
+        pooled_prompt_embeds = None
+        negative_prompt_embeds = None
+        negative_pooled_prompt_embeds = None
+        if hasattr(self.pipe, "compel_proc"):
+            _prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
+                [params.prompt, params.negative_prompt]
+            )
+            prompt = None
+            negative_prompt = None
+            prompt_embeds = _prompt_embeds[0:1]
+            pooled_prompt_embeds = pooled_prompt_embeds[0:1]
+            negative_prompt_embeds = _prompt_embeds[1:2]
+            negative_pooled_prompt_embeds = pooled_prompt_embeds[1:2]
+
         steps = params.steps
         strength = params.strength
         if int(steps * strength) < 1:
@@ -157,10 +179,12 @@ class Pipeline:
 
         results = self.pipe(
             image=params.image,
-            prompt_embeds=prompt_embeds[0:1],
-            pooled_prompt_embeds=pooled_prompt_embeds[0:1],
-            negative_prompt_embeds=prompt_embeds[1:2],
-            negative_pooled_prompt_embeds=pooled_prompt_embeds[1:2],
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
             generator=generator,
             strength=strength,
             num_inference_steps=steps,

pipelines/img2imgSegmindVegaRT.py

@@ -0,0 +1,217 @@
+from diffusers import (
+    AutoPipelineForImage2Image,
+    LCMScheduler,
+    AutoencoderTiny,
+)
+from compel import Compel, ReturnedEmbeddingsType
+import torch
+
+try:
+    import intel_extension_for_pytorch as ipex  # type: ignore
+except:
+    pass
+
+import psutil
+from config import Args
+from pydantic import BaseModel, Field
+from PIL import Image
+import math
+
+base_model = "segmind/Segmind-Vega"
+lora_model = "segmind/Segmind-VegaRT"
+taesd_model = "madebyollin/taesdxl"
+
+default_prompt = "close-up photography of old man standing in the rain at night, in a street lit by lamps, leica 35mm summilux"
+default_negative_prompt = "blurry, low quality, render, 3D, oversaturated"
+page_content = """
+<h1 class="text-3xl font-bold">Real-Time SegmindVegaRT</h1>
+<h3 class="text-xl font-bold">Image-to-Image</h3>
+<p class="text-sm">
+    This demo showcases
+    <a
+    href="https://huggingface.co/segmind/Segmind-VegaRT"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">SegmindVegaRT</a>
+Image to Image pipeline using
+    <a
+    href="https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl_turbo"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">Diffusers</a
+    > with a MJPEG stream server.
+</p>
+<p class="text-sm text-gray-500">
+    Change the prompt to generate different images, accepts <a
+    href="https://github.com/damian0815/compel/blob/main/doc/syntax.md"
+    target="_blank"
+    class="text-blue-500 underline hover:no-underline">Compel</a
+    > syntax.
+</p>
+"""
+
+
+class Pipeline:
+    class Info(BaseModel):
+        name: str = "img2img"
+        title: str = "Image-to-Image Playground 256"
+        description: str = "Generates an image from a text prompt"
+        input_mode: str = "image"
+        page_content: str = page_content
+
+    class InputParams(BaseModel):
+        prompt: str = Field(
+            default_prompt,
+            title="Prompt",
+            field="textarea",
+            id="prompt",
+        )
+        negative_prompt: str = Field(
+            default_negative_prompt,
+            title="Negative Prompt",
+            field="textarea",
+            id="negative_prompt",
+            hide=True,
+        )
+        seed: int = Field(
+            2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
+        )
+        steps: int = Field(
+            1, min=1, max=10, title="Steps", field="range", hide=True, id="steps"
+        )
+        width: int = Field(
+            1024, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
+        )
+        height: int = Field(
+            1024, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
+        )
+        guidance_scale: float = Field(
+            0.0,
+            min=0,
+            max=1,
+            step=0.001,
+            title="Guidance Scale",
+            field="range",
+            hide=True,
+            id="guidance_scale",
+        )
+        strength: float = Field(
+            0.5,
+            min=0.25,
+            max=1.0,
+            step=0.001,
+            title="Strength",
+            field="range",
+            hide=True,
+            id="strength",
+        )
+
+    def __init__(self, args: Args, device: torch.device, torch_dtype: torch.dtype):
+        if args.safety_checker:
+            self.pipe = AutoPipelineForImage2Image.from_pretrained(
+                base_model,
+                variant="fp16",
+            )
+        else:
+            self.pipe = AutoPipelineForImage2Image.from_pretrained(
+                base_model,
+                safety_checker=None,
+                variant="fp16",
+            )
+        if args.taesd:
+            self.pipe.vae = AutoencoderTiny.from_pretrained(
+                taesd_model, torch_dtype=torch_dtype, use_safetensors=True
+            ).to(device)
+
+        self.pipe.load_lora_weights(lora_model)
+        self.pipe.fuse_lora()
+        self.pipe.scheduler = LCMScheduler.from_pretrained(
+            base_model, subfolder="scheduler"
+        )
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
+        self.pipe.set_progress_bar_config(disable=True)
+        self.pipe.to(device=device, dtype=torch_dtype)
+        if device.type != "mps":
+            self.pipe.unet.to(memory_format=torch.channels_last)
+
+        if args.torch_compile:
+            print("Running torch compile")
+            self.pipe.unet = torch.compile(
+                self.pipe.unet, mode="reduce-overhead", fullgraph=False
+            )
+            self.pipe.vae = torch.compile(
+                self.pipe.vae, mode="reduce-overhead", fullgraph=False
+            )
+
+            self.pipe(
+                prompt="warmup",
+                image=[Image.new("RGB", (768, 768))],
+            )
+        if args.compel:
+            self.pipe.compel_proc = Compel(
+                tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
+                text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
+                returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
+                requires_pooled=[False, True],
+            )
+
+    def predict(self, params: "Pipeline.InputParams") -> Image.Image:
+        generator = torch.manual_seed(params.seed)
+        prompt = params.prompt
+        negative_prompt = params.negative_prompt
+        prompt_embeds = None
+        pooled_prompt_embeds = None
+        negative_prompt_embeds = None
+        negative_pooled_prompt_embeds = None
+        if hasattr(self.pipe, "compel_proc"):
+            _prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
+                [params.prompt, params.negative_prompt]
+            )
+            prompt = None
+            negative_prompt = None
+            prompt_embeds = _prompt_embeds[0:1]
+            pooled_prompt_embeds = pooled_prompt_embeds[0:1]
+            negative_prompt_embeds = _prompt_embeds[1:2]
+            negative_pooled_prompt_embeds = pooled_prompt_embeds[1:2]
+
+        steps = params.steps
+        strength = params.strength
+        if int(steps * strength) < 1:
+            steps = math.ceil(1 / max(0.10, strength))
+
+        results = self.pipe(
+            image=params.image,
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            generator=generator,
+            strength=strength,
+            num_inference_steps=steps,
+            guidance_scale=params.guidance_scale,
+            width=params.width,
+            height=params.height,
+            output_type="pil",
+        )
+
+        nsfw_content_detected = (
+            results.nsfw_content_detected[0]
+            if "nsfw_content_detected" in results
+            else False
+        )
+        if nsfw_content_detected:
+            return None
+        result_image = results.images[0]
+
+        return result_image

pipelines/txt2img.py

@@ -85,20 +85,28 @@ class Pipeline:
             self.pipe = DiffusionPipeline.from_pretrained(
                 base_model, safety_checker=None
             )
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
 
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
         self.pipe.set_progress_bar_config(disable=True)
         self.pipe.to(device=device, dtype=torch_dtype)
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        # check if computer has less than 64GB of RAM using sys or os
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
-
         if args.torch_compile:
             self.pipe.unet = torch.compile(
                 self.pipe.unet, mode="reduce-overhead", fullgraph=True
@@ -109,17 +117,24 @@ class Pipeline:
 
             self.pipe(prompt="warmup", num_inference_steps=1, guidance_scale=8.0)
 
-        self.compel_proc = Compel(
-            tokenizer=self.pipe.tokenizer,
-            text_encoder=self.pipe.text_encoder,
-            truncate_long_prompts=False,
-        )
+        if args.compel:
+            self.compel_proc = Compel(
+                tokenizer=self.pipe.tokenizer,
+                text_encoder=self.pipe.text_encoder,
+                truncate_long_prompts=False,
+            )
 
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
-        prompt_embeds = self.compel_proc(params.prompt)
+        prompt_embeds = None
+        prompt = params.prompt
+        if hasattr(self, "compel_proc"):
+            prompt_embeds = self.compel_proc(params.prompt)
+            prompt = None
+
         results = self.pipe(
             prompt_embeds=prompt_embeds,
+            prompt=prompt,
             generator=generator,
             num_inference_steps=params.steps,
             guidance_scale=params.guidance_scale,

pipelines/txt2imgLora.py

@@ -92,20 +92,19 @@ class Pipeline:
             self.pipe = DiffusionPipeline.from_pretrained(
                 base_model, safety_checker=None
             )
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
+
         self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
         self.pipe.set_progress_bar_config(disable=True)
+        self.pipe.load_lora_weights(lcm_lora_id, adapter_name="lcm")
         self.pipe.to(device=device, dtype=torch_dtype)
+
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        # check if computer has less than 64GB of RAM using sys or os
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
-
         if args.torch_compile:
             self.pipe.unet = torch.compile(
                 self.pipe.unet, mode="reduce-overhead", fullgraph=True
@@ -116,18 +115,35 @@ class Pipeline:
 
             self.pipe(prompt="warmup", num_inference_steps=1, guidance_scale=8.0)
 
-        self.pipe.load_lora_weights(lcm_lora_id, adapter_name="lcm")
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
 
-        self.compel_proc = Compel(
-            tokenizer=self.pipe.tokenizer,
-            text_encoder=self.pipe.text_encoder,
-            truncate_long_prompts=False,
-        )
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
+        if args.compel:
+            self.compel_proc = Compel(
+                tokenizer=self.pipe.tokenizer,
+                text_encoder=self.pipe.text_encoder,
+                truncate_long_prompts=False,
+            )
 
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
-        prompt_embeds = self.compel_proc(params.prompt)
+        prompt_embeds = None
+        prompt = params.prompt
+        if hasattr(self, "compel_proc"):
+            prompt_embeds = self.compel_proc(params.prompt)
+            prompt = None
+
         results = self.pipe(
+            prompt=prompt,
             prompt_embeds=prompt_embeds,
             generator=generator,
             num_inference_steps=params.steps,

pipelines/txt2imgLoraSDXL.py

@@ -111,19 +111,29 @@ class Pipeline:
         self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
         self.pipe.set_progress_bar_config(disable=True)
         self.pipe.to(device=device, dtype=torch_dtype).to(device)
+
+        if args.sfast:
+            from sfast.compilers.stable_diffusion_pipeline_compiler import (
+                compile,
+                CompilationConfig,
+            )
+
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.pipe = compile(self.pipe, config=config)
+
         if device.type != "mps":
             self.pipe.unet.to(memory_format=torch.channels_last)
 
-        if psutil.virtual_memory().total < 64 * 1024**3:
-            self.pipe.enable_attention_slicing()
-
         self.pipe.compel_proc = Compel(
             tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
             text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
             returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
             requires_pooled=[False, True],
         )
-        if args.use_taesd:
+        if args.taesd:
             self.pipe.vae = AutoencoderTiny.from_pretrained(
                 taesd_model, torch_dtype=torch_dtype, use_safetensors=True
             ).to(device)
@@ -142,14 +152,30 @@ class Pipeline:
     def predict(self, params: "Pipeline.InputParams") -> Image.Image:
         generator = torch.manual_seed(params.seed)
 
-        prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
-            [params.prompt, params.negative_prompt]
-        )
+        prompt = params.prompt
+        negative_prompt = params.negative_prompt
+        prompt_embeds = None
+        pooled_prompt_embeds = None
+        negative_prompt_embeds = None
+        negative_pooled_prompt_embeds = None
+        if hasattr(self.pipe, "compel_proc"):
+            _prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
+                [params.prompt, params.negative_prompt]
+            )
+            prompt = None
+            negative_prompt = None
+            prompt_embeds = _prompt_embeds[0:1]
+            pooled_prompt_embeds = pooled_prompt_embeds[0:1]
+            negative_prompt_embeds = _prompt_embeds[1:2]
+            negative_pooled_prompt_embeds = pooled_prompt_embeds[1:2]
+
         results = self.pipe(
-            prompt_embeds=prompt_embeds[0:1],
-            pooled_prompt_embeds=pooled_prompt_embeds[0:1],
-            negative_prompt_embeds=prompt_embeds[1:2],
-            negative_pooled_prompt_embeds=pooled_prompt_embeds[1:2],
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
             generator=generator,
             num_inference_steps=params.steps,
             guidance_scale=params.guidance_scale,

requirements.txt

@@ -1,4 +1,4 @@
-git+https://github.com/huggingface/diffusers@dadd55fb36acc862254cf935826d54349b0fcd8c
+git+https://github.com/huggingface/diffusers@2d94c7838e273c40920ffd6d24d724357add7f2d
 transformers==4.35.2
 --extra-index-url https://download.pytorch.org/whl/cu121;
 torch==2.1.0
@@ -10,4 +10,5 @@ compel==2.0.2
 controlnet-aux==0.0.7
 peft==0.6.0
 xformers; sys_platform != 'darwin' or platform_machine != 'arm64'
-markdown2
+markdown2
+stable_fast @ https://github.com/chengzeyi/stable-fast/releases/download/v0.0.15.post1/stable_fast-0.0.15.post1+torch211cu121-cp310-cp310-manylinux2014_x86_64.whl