accepts compel syntax

app-img2img.py

@@ -10,6 +10,7 @@ from fastapi.responses import StreamingResponse, JSONResponse
 from fastapi.staticfiles import StaticFiles
 
 from diffusers import DiffusionPipeline, AutoencoderTiny
+from compel import Compel
 import torch
 from PIL import Image
 import numpy as np
@@ -48,6 +49,7 @@ pipe.set_progress_bar_config(disable=True)
 pipe.to(torch_device="cuda", torch_dtype=torch.float16)
 pipe.unet.to(memory_format=torch.channels_last)
 pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+compel_proc = Compel(tokenizer=pipe.tokenizer, text_encoder=pipe.text_encoder, truncate_long_prompts=False)
 user_queue_map = {}
 
 # for torch.compile
@@ -55,10 +57,11 @@ pipe(prompt="warmup", image=[Image.new("RGB", (512, 512))])
 
 def predict(input_image, prompt, guidance_scale=8.0, strength=0.5, seed=2159232):
     generator = torch.manual_seed(seed)
+    prompt_embeds = compel_proc(prompt)
     # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
     num_inference_steps = 3
     results = pipe(
-        prompt=prompt,
+        prompt_embeds=prompt_embeds,
         generator=generator,
         image=input_image,
         strength=strength,

img2img/index.html

@@ -171,23 +171,26 @@
             <p class="text-sm">
                 This demo showcases
                 <a href="https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7" target="_blank"
-                    class="text-blue-500 hover:underline">LCM</a> Image to Image pipeline
+                    class="text-blue-500 underline hover:no-underline">LCM</a> Image to Image pipeline
                 using
                 <a href="https://github.com/huggingface/diffusers/tree/main/examples/community#latent-consistency-pipeline"
-                    target="_blank" class="text-blue-500 hover:underline">Diffusers</a> with a MJPEG
+                    target="_blank" class="text-blue-500 underline hover:no-underline">Diffusers</a> with a MJPEG
                 stream server.
             </p>
             <p class="text-sm">
                 There are <span id="queue_size" class="font-bold">0</span> user(s) sharing the same GPU, affecting
                 real-time performance. Maximum queue size is 4. <a
                     href="https://huggingface.co/spaces/radames/Real-Time-Latent-Consistency-Model?duplicate=true"
-                    target="_blank" class="text-blue-500 hover:underline">Duplicate</a> and run it on your own GPU.
+                    target="_blank" class="text-blue-500 underline hover:no-underline">Duplicate</a> and run it on your
+                own GPU.
             </p>
         </article>
         <div>
             <h2 class="font-medium">Prompt</h2>
             <p class="text-sm text-gray-500">
-                Change the prompt to generate different images.
+                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>
             <div class="flex text-normal px-1 py-1 border border-gray-700 rounded-md items-center">
                 <textarea type="text" id="prompt" class="font-light w-full px-3 py-2 mx-1 outline-none"