Remove negative embedding checkbox

DOCS.md

@@ -49,13 +49,13 @@ Currently, the only annotator available is [Canny](https://huggingface.co/lllyas
 
 In an image-to-image pipeline, the input image is used as the initial latent representation. With [IP-Adapter](https://github.com/tencent-ailab/IP-Adapter), the image is processed by a separate image encoder and the encoded features are used as conditioning along with the text prompt.
 
-For capturing faces, enable `IP-Adapter Face` to use the full-face model. You should use an input image that is mostly a face and it should be high quality. You can generate fake portraits with Realistic Vision to experiment.
+For capturing faces, enable `IP-Adapter Face` to use the full-face model. You should use an input image that is mostly a face and it should be high quality.
 
 ### Advanced
 
 #### Textual Inversion
 
-Enable `Use negative TI` to append [`fast_negative`](https://civitai.com/models/71961?modelVersionId=94057) to your negative prompt. Read [An Image is Worth One Word](https://huggingface.co/papers/2208.01618) to learn more.
+Add `<fast_negative>` anywhere in your negative prompt to apply the [`fast_negative`](https://civitai.com/models/71961?modelVersionId=94057) textual inversion embedding. Read [An Image is Worth One Word](https://huggingface.co/papers/2208.01618) to learn more.
 
 #### DeepCache
 

app.py

@@ -184,7 +184,7 @@ with gr.Blocks(
             with gr.Row():
                 negative_prompt = gr.Textbox(
                     label="Negative Prompt",
-                    value="nsfw",
+                    value="nsfw, <fast_negative>",
                     lines=1,
                 )
 
@@ -289,11 +289,6 @@ with gr.Blocks(
                     label="Karras σ",
                     value=True,
                 )
-                use_negative_embedding = gr.Checkbox(
-                    elem_classes=["checkbox"],
-                    label="Use negative TI",
-                    value=False,
-                )
 
             # Image-to-Image settings
             gr.HTML("<h3>Image-to-Image</h3>")
@@ -434,7 +429,6 @@ with gr.Blocks(
             num_images,
             use_karras,
             use_ip_face,
-            use_negative_embedding,
             DISABLE_IMAGE_PROMPT,
             DISABLE_CONTROL_IMAGE_PROMPT,
             DISABLE_IP_IMAGE_PROMPT,

lib/config.py

@@ -121,14 +121,13 @@ Config = SimpleNamespace(
     ANNOTATORS={
         "canny": "lllyasviel/control_v11p_sd15_canny",
     },
-    NEGATIVE_EMBEDDING="fast_negative",
     WIDTH=512,
     HEIGHT=512,
     NUM_IMAGES=1,
     SEED=-1,
-    GUIDANCE_SCALE=7.5,
+    GUIDANCE_SCALE=6,
     INFERENCE_STEPS=40,
-    DENOISING_STRENGTH=0.7,
+    DENOISING_STRENGTH=0.8,
     DEEPCACHE_INTERVAL=1,
     SCALE=1,
     SCALES=[1, 2, 4],

lib/inference.py

@@ -50,7 +50,7 @@ def generate(
     annotator="canny",
     width=512,
     height=512,
-    guidance_scale=7.5,
+    guidance_scale=6.0,
     inference_steps=40,
     denoising_strength=0.8,
     deepcache=1,
@@ -58,7 +58,6 @@ def generate(
     num_images=1,
     karras=False,
     ip_face=False,
-    negative_embedding=False,
     Error=Exception,
     Info=None,
     progress=None,
@@ -85,6 +84,8 @@ def generate(
 
     EMBEDDINGS_TYPE = ReturnedEmbeddingsType.LAST_HIDDEN_STATES_NORMALIZED
 
+    FAST_NEGATIVE = "<fast_negative>" in negative_prompt
+
     if ip_image_prompt:
         IP_ADAPTER = "full-face" if ip_face else "plus"
     else:
@@ -123,19 +124,15 @@ def generate(
     pipe = loader.pipe
     upscaler = loader.upscaler
 
-    # Load negative embedding if requested
-    if negative_embedding:
+    # Load fast negative embedding
+    if FAST_NEGATIVE:
         embeddings_dir = os.path.abspath(
             os.path.join(os.path.dirname(__file__), "..", "embeddings")
         )
-        embedding = Config.NEGATIVE_EMBEDDING
-        try:
-            pipe.load_textual_inversion(
-                pretrained_model_name_or_path=f"{embeddings_dir}/{embedding}.pt",
-                token=f"<{embedding}>",
-            )
-        except (EnvironmentError, HFValidationError, RepositoryNotFoundError):
-            raise Error(f"Invalid embedding: {embedding}")
+        pipe.load_textual_inversion(
+            pretrained_model_name_or_path=f"{embeddings_dir}/fast_negative.pt",
+            token="<fast_negative>",
+        )
 
     # Embed prompts with weights
     compel = Compel(
@@ -155,10 +152,6 @@ def generate(
     for i in range(num_images):
         try:
             generator = torch.Generator(device=pipe.device).manual_seed(current_seed)
-
-            if negative_embedding:
-                negative_prompt += f", <{Config.NEGATIVE_EMBEDDING}>"
-
             positive_embeds, negative_embeds = compel.pad_conditioning_tensors_to_same_length(
                 [compel(positive_prompt), compel(negative_prompt)]
             )
@@ -198,8 +191,9 @@ def generate(
             images.append((image, str(current_seed)))
             current_seed += 1
         finally:
-            if negative_embedding:
+            if FAST_NEGATIVE:
                 pipe.unload_textual_inversion()
+
             CURRENT_STEP = 0
             CURRENT_IMAGE += 1