Simplify textual inversion embeddings

DOCS.md

@@ -41,16 +41,6 @@ Apply up to 2 LoRA (low-rank adaptation) adapters with adjustable strength:
 
 > NB: The trigger words are automatically appended to the positive prompt for you.
 
-### Embeddings
-
-Select one or more [textual inversion](https://huggingface.co/docs/diffusers/en/using-diffusers/textual_inversion_inference) embeddings:
-
-* [`fast_negative`](https://civitai.com/models/71961?modelVersionId=94057): all-purpose (default, **recommended**)
-* [`cyberrealistic_negative`](https://civitai.com/models/77976?modelVersionId=82745): realistic add-on (for CyberRealistic)
-* [`unrealistic_dream`](https://civitai.com/models/72437?modelVersionId=77173): realistic add-on (for RealisticVision)
-
-> NB: The trigger token is automatically appended to the negative prompt for you.
-
 ### Styles
 
 [Styles](https://huggingface.co/spaces/adamelliotfields/diffusion/blob/main/data/styles.json) are prompt templates that wrap your positive and negative prompts. They were originally derived from the [twri/sdxl_prompt_styler](https://github.com/twri/sdxl_prompt_styler) Comfy node, but have since been entirely rewritten.
@@ -83,7 +73,7 @@ Initial image strength (known as _denoising strength_) is essentially how much t
 
 #### ControlNet
 
-In [ControlNet](https://github.com/lllyasviel/ControlNet), the input image is used to get a feature map from an _annotator_. These are computer vision models used for tasks like edge detection and pose estimation. ControlNet models are trained to understand these feature maps. Read the [Diffusers docs](https://huggingface.co/docs/diffusers/using-diffusers/controlnet) to learn more.
+In [ControlNet](https://github.com/lllyasviel/ControlNet), the input image is used to get a feature map from an _annotator_. These are computer vision models used for tasks like edge detection and pose estimation. ControlNet models are trained to understand these feature maps. Read the [docs](https://huggingface.co/docs/diffusers/using-diffusers/controlnet) to learn more.
 
 Currently, the only annotator available is [Canny](https://huggingface.co/lllyasviel/control_v11p_sd15_canny) (edge detection).
 
@@ -95,6 +85,10 @@ For capturing faces, enable `IP-Adapter Face` to use the full-face model. You sh
 
 ### 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.
+
 #### DeepCache
 
 [DeepCache](https://github.com/horseee/DeepCache) caches lower UNet layers and reuses them every `Interval` steps. Trade quality for speed:

app.py

@@ -215,15 +215,6 @@ with gr.Blocks(
                     label="Scheduler",
                     filterable=False,
                 )
-            with gr.Row():
-                embeddings = gr.Dropdown(
-                    elem_id="embeddings",
-                    label="Embeddings",
-                    choices=[(f"<{e}>", e) for e in Config.EMBEDDINGS],
-                    multiselect=True,
-                    value=[Config.EMBEDDING],
-                    min_width=240,
-                )
             with gr.Row():
                 with gr.Group(elem_classes=["gap-0"]):
                     lora_1 = gr.Dropdown(
@@ -315,7 +306,7 @@ with gr.Blocks(
             with gr.Row():
                 file_format = gr.Dropdown(
                     choices=["png", "jpeg", "webp"],
-                    label="File Format",
+                    label="Format",
                     filterable=False,
                     value="png",
                 )
@@ -343,6 +334,11 @@ with gr.Blocks(
                     label="Karras σ",
                     value=True,
                 )
+                use_negative_embedding = gr.Checkbox(
+                    elem_classes=["checkbox"],
+                    label="Use negative TI",
+                    value=False,
+                )
                 use_taesd = gr.Checkbox(
                     elem_classes=["checkbox"],
                     label="Tiny VAE",
@@ -487,7 +483,6 @@ with gr.Blocks(
             lora_1_weight,
             lora_2,
             lora_2_weight,
-            embeddings,
             style,
             seed,
             model,
@@ -506,6 +501,7 @@ with gr.Blocks(
             use_freeu,
             use_clip_skip,
             use_ip_face,
+            use_negative_embedding,
             DISABLE_IMAGE_PROMPT,
             DISABLE_CONTROL_IMAGE_PROMPT,
             DISABLE_IP_IMAGE_PROMPT,

embeddings/cyberrealistic_negative.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:65f3ea567c04c22f92024c5b55cbeca580bc330c4290aeb647ebd86273b3ffb8
-size 197662

embeddings/unrealistic_dream.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a77451e7ea075c7f72d488d2b740b3d3970c671c0ac39dd3155f3c3b129df959
-size 114539

lib/config.py

@@ -140,12 +140,7 @@ Config = SimpleNamespace(
     ANNOTATORS={
         "canny": "lllyasviel/control_v11p_sd15_canny",
     },
-    EMBEDDING="fast_negative",
-    EMBEDDINGS=[
-        "cyberrealistic_negative",
-        "fast_negative",
-        "unrealistic_dream",
-    ],
+    NEGATIVE_EMBEDDING="fast_negative",
     STYLE="enhance",
     WIDTH=512,
     HEIGHT=512,

lib/inference.py

@@ -70,7 +70,6 @@ def generate(
     lora_1_weight=0.0,
     lora_2=None,
     lora_2_weight=0.0,
-    embeddings=[],
     style=None,
     seed=None,
     model="Lykon/dreamshaper-8",
@@ -89,6 +88,7 @@ def generate(
     freeu=False,
     clip_skip=False,
     ip_face=False,
+    negative_embedding=False,
     Error=Exception,
     Info=None,
     progress=None,
@@ -193,11 +193,13 @@ def generate(
         pipe.unload_lora_weights()
         raise Error("Error setting LoRA weights")
 
-    # load embeddings
-    embeddings_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "embeddings"))
-    for embedding in embeddings:
+    # Load negative embedding if requested
+    if negative_embedding:
+        embeddings_dir = os.path.abspath(
+            os.path.join(os.path.dirname(__file__), "..", "embeddings")
+        )
+        embedding = Config.NEGATIVE_EMBEDDING
         try:
-            # wrap embeddings in angle brackets
             pipe.load_textual_inversion(
                 pretrained_model_name_or_path=f"{embeddings_dir}/{embedding}.pt",
                 token=f"<{embedding}>",
@@ -219,6 +221,7 @@ def generate(
     images = []
     current_seed = seed
     safe_progress(progress, 0, num_images, f"Generating image 0/{num_images}")
+
     for i in range(num_images):
         try:
             generator = torch.Generator(device=pipe.device).manual_seed(current_seed)
@@ -228,12 +231,12 @@ def generate(
             if negative_styled.startswith("(), "):
                 negative_styled = negative_styled[4:]
 
+            if negative_embedding:
+                negative_styled += f", <{Config.NEGATIVE_EMBEDDING}>"
+
             for lora in loras:
                 positive_styled += f", {Config.CIVIT_LORAS[lora]['trigger']}"
 
-            for embedding in embeddings:
-                negative_styled += f", <{embedding}>"
-
             positive_embeds, negative_embeds = compel.pad_conditioning_tensors_to_same_length(
                 [compel(positive_styled), compel(negative_styled)]
             )
@@ -273,7 +276,7 @@ def generate(
             images.append((image, str(current_seed)))
             current_seed += 1
         finally:
-            if embeddings:
+            if negative_embedding:
                 pipe.unload_textual_inversion()
             if loras:
                 pipe.unload_lora_weights()