radames commited on
Commit
ae27e5e
1 Parent(s): d42561e

accepts compel syntax

Browse files
Files changed (3) hide show
  1. app-img2img.py +4 -1
  2. img2img/index.html +7 -4
  3. tailwind.config.js +0 -0
app-img2img.py CHANGED
@@ -10,6 +10,7 @@ from fastapi.responses import StreamingResponse, JSONResponse
10
  from fastapi.staticfiles import StaticFiles
11
 
12
  from diffusers import DiffusionPipeline, AutoencoderTiny
 
13
  import torch
14
  from PIL import Image
15
  import numpy as np
@@ -48,6 +49,7 @@ pipe.set_progress_bar_config(disable=True)
48
  pipe.to(torch_device="cuda", torch_dtype=torch.float16)
49
  pipe.unet.to(memory_format=torch.channels_last)
50
  pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
 
51
  user_queue_map = {}
52
 
53
  # for torch.compile
@@ -55,10 +57,11 @@ pipe(prompt="warmup", image=[Image.new("RGB", (512, 512))])
55
 
56
  def predict(input_image, prompt, guidance_scale=8.0, strength=0.5, seed=2159232):
57
  generator = torch.manual_seed(seed)
 
58
  # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
59
  num_inference_steps = 3
60
  results = pipe(
61
- prompt=prompt,
62
  generator=generator,
63
  image=input_image,
64
  strength=strength,
 
10
  from fastapi.staticfiles import StaticFiles
11
 
12
  from diffusers import DiffusionPipeline, AutoencoderTiny
13
+ from compel import Compel
14
  import torch
15
  from PIL import Image
16
  import numpy as np
 
49
  pipe.to(torch_device="cuda", torch_dtype=torch.float16)
50
  pipe.unet.to(memory_format=torch.channels_last)
51
  pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
52
+ compel_proc = Compel(tokenizer=pipe.tokenizer, text_encoder=pipe.text_encoder, truncate_long_prompts=False)
53
  user_queue_map = {}
54
 
55
  # for torch.compile
 
57
 
58
  def predict(input_image, prompt, guidance_scale=8.0, strength=0.5, seed=2159232):
59
  generator = torch.manual_seed(seed)
60
+ prompt_embeds = compel_proc(prompt)
61
  # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
62
  num_inference_steps = 3
63
  results = pipe(
64
+ prompt_embeds=prompt_embeds,
65
  generator=generator,
66
  image=input_image,
67
  strength=strength,
img2img/index.html CHANGED
@@ -171,23 +171,26 @@
171
  <p class="text-sm">
172
  This demo showcases
173
  <a href="https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7" target="_blank"
174
- class="text-blue-500 hover:underline">LCM</a> Image to Image pipeline
175
  using
176
  <a href="https://github.com/huggingface/diffusers/tree/main/examples/community#latent-consistency-pipeline"
177
- target="_blank" class="text-blue-500 hover:underline">Diffusers</a> with a MJPEG
178
  stream server.
179
  </p>
180
  <p class="text-sm">
181
  There are <span id="queue_size" class="font-bold">0</span> user(s) sharing the same GPU, affecting
182
  real-time performance. Maximum queue size is 4. <a
183
  href="https://huggingface.co/spaces/radames/Real-Time-Latent-Consistency-Model?duplicate=true"
184
- target="_blank" class="text-blue-500 hover:underline">Duplicate</a> and run it on your own GPU.
 
185
  </p>
186
  </article>
187
  <div>
188
  <h2 class="font-medium">Prompt</h2>
189
  <p class="text-sm text-gray-500">
190
- Change the prompt to generate different images.
 
 
191
  </p>
192
  <div class="flex text-normal px-1 py-1 border border-gray-700 rounded-md items-center">
193
  <textarea type="text" id="prompt" class="font-light w-full px-3 py-2 mx-1 outline-none"
 
171
  <p class="text-sm">
172
  This demo showcases
173
  <a href="https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7" target="_blank"
174
+ class="text-blue-500 underline hover:no-underline">LCM</a> Image to Image pipeline
175
  using
176
  <a href="https://github.com/huggingface/diffusers/tree/main/examples/community#latent-consistency-pipeline"
177
+ target="_blank" class="text-blue-500 underline hover:no-underline">Diffusers</a> with a MJPEG
178
  stream server.
179
  </p>
180
  <p class="text-sm">
181
  There are <span id="queue_size" class="font-bold">0</span> user(s) sharing the same GPU, affecting
182
  real-time performance. Maximum queue size is 4. <a
183
  href="https://huggingface.co/spaces/radames/Real-Time-Latent-Consistency-Model?duplicate=true"
184
+ target="_blank" class="text-blue-500 underline hover:no-underline">Duplicate</a> and run it on your
185
+ own GPU.
186
  </p>
187
  </article>
188
  <div>
189
  <h2 class="font-medium">Prompt</h2>
190
  <p class="text-sm text-gray-500">
191
+ Change the prompt to generate different images, accepts <a
192
+ href="https://github.com/damian0815/compel/blob/main/doc/syntax.md" target="_blank"
193
+ class="text-blue-500 underline hover:no-underline">Compel</a> syntax.
194
  </p>
195
  <div class="flex text-normal px-1 py-1 border border-gray-700 rounded-md items-center">
196
  <textarea type="text" id="prompt" class="font-light w-full px-3 py-2 mx-1 outline-none"
tailwind.config.js DELETED
File without changes