Embeddings multiselect

app.css

@@ -105,7 +105,7 @@
   border-width: 0px;
 }
 .tabitem {
-  max-height: 516px;
+  max-height: 698px;
   overflow-x: hidden;
   overflow-y: auto;
   padding: 0 0 8px;

app.py

@@ -4,8 +4,7 @@ import random
 
 import gradio as gr
 
-import config as cfg
-from lib import generate
+from lib import Config, generate
 
 # the CSS `content` attribute expects a string so we need to wrap the number in quotes
 refresh_seed_js = """
@@ -73,8 +72,8 @@ with gr.Blocks(
         radius_size=gr.themes.sizes.radius_sm,
         spacing_size=gr.themes.sizes.spacing_md,
         # fonts
-        font=[gr.themes.GoogleFont("Inter"), *cfg.SANS_FONTS],
-        font_mono=[gr.themes.GoogleFont("Ubuntu Mono"), *cfg.MONO_FONTS],
+        font=[gr.themes.GoogleFont("Inter"), *Config.SANS_FONTS],
+        font_mono=[gr.themes.GoogleFont("Ubuntu Mono"), *Config.MONO_FONTS],
     ).set(
         layout_gap="8px",
         block_shadow="0 0 #0000",
@@ -95,51 +94,59 @@ with gr.Blocks(
             with gr.TabItem("⚙️ Settings"):
                 with gr.Group():
                     negative_prompt = gr.Textbox(
-                        value=cfg.NEGATIVE_PROMPT,
+                        value=None,
                         label="Negative Prompt",
                         placeholder="ugly, bad",
                         lines=2,
                     )
-                    model = gr.Dropdown(
-                        choices=cfg.MODELS,
-                        filterable=False,
-                        value=cfg.MODEL,
-                        label="Model",
-                    )
+
+                    with gr.Row():
+                        model = gr.Dropdown(
+                            choices=Config.MODELS,
+                            filterable=False,
+                            value=Config.MODEL,
+                            label="Model",
+                        )
+                        scheduler = gr.Dropdown(
+                            choices=Config.SCHEDULERS,
+                            value=Config.SCHEDULER,
+                            elem_id="scheduler",
+                            label="Scheduler",
+                            filterable=False,
+                        )
 
                     with gr.Row():
                         styles = json.loads(read_file("data/styles.json"))
                         style = gr.Dropdown(
-                            value=cfg.STYLE,
+                            value=Config.STYLE,
                             label="Style",
                             min_width=200,
                             choices=[("None", None)] + [(s["name"], s["id"]) for s in styles],
                         )
-                        scheduler = gr.Dropdown(
-                            choices=cfg.SCHEDULERS,
-                            value=cfg.SCHEDULER,
-                            elem_id="scheduler",
-                            label="Scheduler",
-                            filterable=False,
+                        embeddings = gr.Dropdown(
+                            label="Embeddings",
+                            choices=[(f"<{e}>", e) for e in Config.EMBEDDINGS],
+                            multiselect=True,
+                            value=[Config.EMBEDDING],
                         )
 
                     with gr.Row():
                         guidance_scale = gr.Slider(
-                            value=cfg.GUIDANCE_SCALE,
+                            value=Config.GUIDANCE_SCALE,
                             label="Guidance Scale",
                             minimum=1.0,
                             maximum=15.0,
                             step=0.1,
                         )
                         inference_steps = gr.Slider(
-                            value=cfg.INFERENCE_STEPS,
+                            value=Config.INFERENCE_STEPS,
                             label="Inference Steps",
                             minimum=1,
                             maximum=50,
                             step=1,
                         )
                         seed = gr.Number(
-                            value=cfg.SEED,
+                            value=Config.SEED,
                             label="Seed",
                             minimum=-1,
                             maximum=(2**64) - 1,
@@ -147,14 +154,14 @@ with gr.Blocks(
 
                     with gr.Row():
                         width = gr.Slider(
-                            value=cfg.WIDTH,
+                            value=Config.WIDTH,
                             label="Width",
                             minimum=320,
                             maximum=768,
                             step=16,
                         )
                         height = gr.Slider(
-                            value=cfg.HEIGHT,
+                            value=Config.HEIGHT,
                             label="Height",
                             minimum=320,
                             maximum=768,
@@ -175,45 +182,19 @@ with gr.Blocks(
                             min_width=180,
                         )
                         scale = gr.Dropdown(
-                            choices=[(f"{s}x", s) for s in cfg.SCALES],
+                            choices=[(f"{s}x", s) for s in Config.SCALES],
                             filterable=False,
-                            value=cfg.SCALE,
+                            value=Config.SCALE,
                             label="Scale",
-                            min_width=100,
                         )
+
+                    with gr.Row():
                         num_images = gr.Dropdown(
                             choices=list(range(1, 5)),
-                            value=cfg.NUM_IMAGES,
+                            value=Config.NUM_IMAGES,
                             filterable=False,
                             label="Images",
-                            min_width=60,
                         )
-
-                    with gr.Row():
-                        use_karras = gr.Checkbox(
-                            elem_classes=["checkbox"],
-                            label="Karras σ",
-                            value=True,
-                        )
-                        use_freeu = gr.Checkbox(
-                            elem_classes=["checkbox"],
-                            label="FreeU",
-                            value=False,
-                        )
-                        use_clip_skip = gr.Checkbox(
-                            elem_classes=["checkbox"],
-                            label="Clip skip",
-                            value=False,
-                        )
-                        increment_seed = gr.Checkbox(
-                            elem_classes=["checkbox"],
-                            label="Autoincrement",
-                            value=True,
-                        )
-
-            with gr.TabItem("🛠️ Advanced"):
-                with gr.Group():
-                    with gr.Row():
                         file_format = gr.Dropdown(
                             choices=["png", "jpeg", "webp"],
                             label="File Format",
@@ -221,14 +202,14 @@ with gr.Blocks(
                             value="png",
                         )
                         deepcache_interval = gr.Slider(
-                            value=cfg.DEEPCACHE_INTERVAL,
-                            label="DeepCache Interval",
+                            value=Config.DEEPCACHE_INTERVAL,
+                            label="DeepCache",
                             minimum=1,
                             maximum=4,
                             step=1,
                         )
                         tome_ratio = gr.Slider(
-                            value=cfg.TOME_RATIO,
+                            value=Config.TOME_RATIO,
                             label="ToMe Ratio",
                             minimum=0.0,
                             maximum=0.5,
@@ -236,6 +217,28 @@ with gr.Blocks(
                         )
 
                     with gr.Row():
+                        increment_seed = gr.Checkbox(
+                            elem_classes=["checkbox"],
+                            label="Autoincrement",
+                            value=True,
+                        )
+                        use_freeu = gr.Checkbox(
+                            elem_classes=["checkbox"],
+                            label="FreeU",
+                            value=False,
+                        )
+                        use_clip_skip = gr.Checkbox(
+                            elem_classes=["checkbox"],
+                            label="Clip skip",
+                            value=False,
+                        )
+
+                    with gr.Row():
+                        use_karras = gr.Checkbox(
+                            elem_classes=["checkbox"],
+                            label="Karras σ",
+                            value=True,
+                        )
                         use_taesd = gr.Checkbox(
                             elem_classes=["checkbox"],
                             label="Tiny VAE",
@@ -333,6 +336,7 @@ with gr.Blocks(
         inputs=[
             prompt,
             negative_prompt,
+            embeddings,
             style,
             seed,
             model,

cli.py

@@ -2,8 +2,7 @@
 # usage: python cli.py 'colorful calico cat artstation'
 import argparse
 
-import config as cfg
-from lib import generate
+from lib import Config, generate
 
 
 def save_images(images, filename="image.png"):
@@ -16,20 +15,21 @@ def main():
     # fmt: off
     parser = argparse.ArgumentParser(add_help=False, allow_abbrev=False)
     parser.add_argument("prompt", type=str, metavar="PROMPT")
-    parser.add_argument("-n", "--negative", type=str, metavar="STR", default=cfg.NEGATIVE_PROMPT)
-    parser.add_argument("-s", "--seed", type=int, metavar="INT", default=cfg.SEED)
+    parser.add_argument("-n", "--negative", type=str, metavar="STR", default="")
+    parser.add_argument("-e", "--embedding", type=str, metavar="STR", default=[], action="append")
+    parser.add_argument("-s", "--seed", type=int, metavar="INT", default=Config.SEED)
     parser.add_argument("-i", "--images", type=int, metavar="INT", default=1)
     parser.add_argument("-f", "--filename", type=str, metavar="STR", default="image.png")
-    parser.add_argument("-w", "--width", type=int, metavar="INT", default=cfg.WIDTH)
-    parser.add_argument("-h", "--height", type=int, metavar="INT", default=cfg.HEIGHT)
-    parser.add_argument("-m", "--model", type=str, metavar="STR", default=cfg.MODEL)
-    parser.add_argument("-d", "--deepcache", type=int, metavar="INT", default=cfg.DEEPCACHE_INTERVAL)
-    parser.add_argument("--scale", type=int, metavar="INT", choices=cfg.SCALES, default=cfg.SCALE)
-    parser.add_argument("--style", type=str, metavar="STR", default=cfg.STYLE)
-    parser.add_argument("--scheduler", type=str, metavar="STR", default=cfg.SCHEDULER)
-    parser.add_argument("--guidance", type=float, metavar="FLOAT", default=cfg.GUIDANCE_SCALE)
-    parser.add_argument("--steps", type=int, metavar="INT", default=cfg.INFERENCE_STEPS)
-    parser.add_argument("--tome", type=float, metavar="FLOAT", default=cfg.TOME_RATIO)
+    parser.add_argument("-w", "--width", type=int, metavar="INT", default=Config.WIDTH)
+    parser.add_argument("-h", "--height", type=int, metavar="INT", default=Config.HEIGHT)
+    parser.add_argument("-m", "--model", type=str, metavar="STR", default=Config.MODEL)
+    parser.add_argument("-d", "--deepcache", type=int, metavar="INT", default=Config.DEEPCACHE_INTERVAL)
+    parser.add_argument("--scale", type=int, metavar="INT", choices=Config.SCALES, default=Config.SCALE)
+    parser.add_argument("--style", type=str, metavar="STR", default=Config.STYLE)
+    parser.add_argument("--scheduler", type=str, metavar="STR", default=Config.SCHEDULER)
+    parser.add_argument("--guidance", type=float, metavar="FLOAT", default=Config.GUIDANCE_SCALE)
+    parser.add_argument("--steps", type=int, metavar="INT", default=Config.INFERENCE_STEPS)
+    parser.add_argument("--tome", type=float, metavar="FLOAT", default=Config.TOME_RATIO)
     parser.add_argument("--taesd", action="store_true")
     parser.add_argument("--clip-skip", action="store_true")
     parser.add_argument("--truncate", action="store_true")
@@ -42,6 +42,7 @@ def main():
     images = generate(
         args.prompt,
         args.negative,
+        args.embedding,
         args.style,
         args.seed,
         args.model,

config.py

@@ -1,57 +0,0 @@
-MONO_FONTS = ["monospace"]
-
-SANS_FONTS = [
-    "sans-serif",
-    "Apple Color Emoji",
-    "Segoe UI Emoji",
-    "Segoe UI Symbol",
-    "Noto Color Emoji",
-]
-
-# embeddings loaded and renamed in generate.py
-NEGATIVE_PROMPT = "<fast_negative>"
-
-MODEL = "Lykon/dreamshaper-8"
-
-MODELS = [
-    "fluently/Fluently-v4",
-    "Linaqruf/anything-v3-1",
-    "Lykon/dreamshaper-8",
-    "prompthero/openjourney-v4",
-    "runwayml/stable-diffusion-v1-5",
-    "SG161222/Realistic_Vision_V5.1_Novae",
-]
-
-SCHEDULER = "DEIS 2M"
-
-SCHEDULERS = [
-    "DEIS 2M",
-    "DPM++ 2M",
-    "DPM2 a",
-    "Euler a",
-    "Heun",
-    "LMS",
-    "PNDM",
-]
-
-STYLE = "sai-enhance"
-
-WIDTH = 448
-
-HEIGHT = 576
-
-NUM_IMAGES = 4
-
-SEED = -1
-
-GUIDANCE_SCALE = 7
-
-INFERENCE_STEPS = 30
-
-DEEPCACHE_INTERVAL = 2
-
-TOME_RATIO = 0.0
-
-SCALE = 2
-
-SCALES = [1, 2, 4]

lib/__init__.py

@@ -1,3 +1,6 @@
+from .config import Config
 from .inference import generate
 from .loader import Loader
 from .upscaler import RealESRGAN
+
+__all__ = ["Config", "Loader", "RealESRGAN", "generate"]

lib/config.py

@@ -0,0 +1,51 @@
+from types import SimpleNamespace
+
+Config = SimpleNamespace(
+    MONO_FONTS=["monospace"],
+    SANS_FONTS=[
+        "sans-serif",
+        "Apple Color Emoji",
+        "Segoe UI Emoji",
+        "Segoe UI Symbol",
+        "Noto Color Emoji",
+    ],
+    MODEL="Lykon/dreamshaper-8",
+    MODELS=[
+        "fluently/Fluently-v4",
+        "Linaqruf/anything-v3-1",
+        "Lykon/dreamshaper-8",
+        "prompthero/openjourney-v4",
+        "runwayml/stable-diffusion-v1-5",
+        "SG161222/Realistic_Vision_V5.1_Novae",
+    ],
+    SCHEDULER="DEIS 2M",
+    SCHEDULERS=[
+        "DEIS 2M",
+        "DPM++ 2M",
+        "DPM2 a",
+        "Euler a",
+        "Heun",
+        "LMS",
+        "PNDM",
+    ],
+    EMBEDDING="fast_negative",
+    EMBEDDINGS={
+        "async_negative",
+        "bad_dream",
+        "cyberrealistic_negative",
+        "deep_negative",
+        "fast_negative",
+        "unrealistic_dream",
+    },
+    STYLE="sai-enhance",
+    WIDTH=448,
+    HEIGHT=576,
+    NUM_IMAGES=1,
+    SEED=-1,
+    GUIDANCE_SCALE=7,
+    INFERENCE_STEPS=30,
+    DEEPCACHE_INTERVAL=2,
+    TOME_RATIO=0.0,
+    SCALE=1,
+    SCALES=[1, 2, 4],
+)

lib/inference.py

@@ -12,6 +12,7 @@ import tomesd
 import torch
 from compel import Compel, DiffusersTextualInversionManager, ReturnedEmbeddingsType
 from compel.prompt_parser import PromptParser
+from huggingface_hub.utils import HFValidationError, RepositoryNotFoundError
 
 from .loader import Loader
 
@@ -75,6 +76,7 @@ def apply_style(prompt, style_id, negative=False):
 def generate(
     positive_prompt,
     negative_prompt="",
+    embeddings=[],
     style=None,
     seed=None,
     model="runwayml/stable-diffusion-v1-5",
@@ -132,6 +134,21 @@ def generate(
             DEVICE,
         )
 
+        # load embeddings and append to negative prompt
+        embeddings_dir = os.path.join(os.path.dirname(__file__), "..", "embeddings")
+        embeddings_dir = os.path.abspath(embeddings_dir)
+        for embedding in embeddings:
+            try:
+                pipe.load_textual_inversion(
+                    pretrained_model_name_or_path=f"{embeddings_dir}/{embedding}.pt",
+                    token=f"<{embedding}>",
+                )
+                negative_prompt = (
+                    f"{negative_prompt}, {embedding}" if negative_prompt else embedding
+                )
+            except (EnvironmentError, HFValidationError, RepositoryNotFoundError):
+                raise Error(f"Invalid embedding: {embedding}")
+
         # prompt embeds
         compel = Compel(
             textual_inversion_manager=DiffusersTextualInversionManager(pipe),
@@ -185,6 +202,7 @@ def generate(
                     images.append((image, str(current_seed)))
                 finally:
                     if not ZERO_GPU:
+                        pipe.unload_textual_inversion()
                         torch.cuda.empty_cache()
 
             if increment_seed:
@@ -193,6 +211,7 @@ def generate(
         if ZERO_GPU:
             # spaces always start fresh
             loader.pipe = None
+            loader.upscaler = None
 
         diff = time.perf_counter() - start
         if Info:

lib/loader.py

@@ -22,14 +22,6 @@ ZERO_GPU = (
     or os.environ.get("SPACES_ZERO_GPU", "") == "1"
 )
 
-EMBEDDINGS = {
-    "./embeddings/bad_prompt_version2.pt": "<bad_prompt>",
-    "./embeddings/BadDream.pt": "<bad_dream>",
-    "./embeddings/FastNegativeV2.pt": "<fast_negative>",
-    "./embeddings/negative_hand.pt": "<negative_hand>",
-    "./embeddings/UnrealisticDream.pt": "<unrealistic_dream>",
-}
-
 
 # inspired by ComfyUI
 # https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/model_management.py
@@ -191,10 +183,6 @@ class Loader:
             device=device,
             dtype=dtype,
         )
-        self.pipe.load_textual_inversion(
-            pretrained_model_name_or_path=list(EMBEDDINGS.keys()),
-            tokens=list(EMBEDDINGS.values()),
-        )
         self._load_vae(model_lower, taesd, variant)
         self._load_freeu(freeu)
         self._load_deepcache(deepcache_interval)

usage.md

@@ -12,33 +12,21 @@ Positive and negative prompts are embedded by [Compel](https://github.com/damian
 
 Note that `++` is `1.1^2` (and so on). See [syntax features](https://github.com/damian0815/compel/blob/main/doc/syntax.md) to learn more and read [Civitai](https://civitai.com)'s guide on [prompting](https://education.civitai.com/civitais-prompt-crafting-guide-part-1-basics/) for best practices.
 
-#### Negative Prompt
-
-Start with a [textual inversion](https://huggingface.co/docs/diffusers/en/using-diffusers/textual_inversion_inference) embedding:
-
-* [`<bad_prompt>`](https://civitai.com/models/55700/badprompt-negative-embedding)
-* [`<negative_hand>`](https://civitai.com/models/56519/negativehand-negative-embedding)
-* [`<fast_negative>`](https://civitai.com/models/71961/fast-negative-embedding-fastnegativev2)
-* [`<bad_dream>`](https://civitai.com/models/72437?modelVersionId=77169)
-* [`<unrealistic_dream>`](https://civitai.com/models/72437?modelVersionId=77173)
-
-And add to it. You can use weighting in the negative prompt as well.
-
 #### Arrays
 
 Arrays allow you to generate different images from a single prompt. For example, `[[cat,corgi]]` will expand into 2 separate prompts. Make sure `Images` is set accordingly (e.g., 2). Only works for the positive prompt. Inspired by [Fooocus](https://github.com/lllyasviel/Fooocus/pull/1503).
 
-### Styles
+### Embeddings
 
-Styles are prompt templates from twri's [sdxl_prompt_styler](https://github.com/twri/sdxl_prompt_styler) Comfy node. Start with a subject like "cat", pick a style, and iterate from there.
+Select multiple negative [textual inversion](https://huggingface.co/docs/diffusers/en/using-diffusers/textual_inversion_inference) embeddings. Fast Negative and Bad Dream can be used standalone or together; Unrealistic Dream should be combined with one of the others:
 
-#### FreeU
-
-[FreeU](https://github.com/ChenyangSi/FreeU) (Si et al. 2023) re-weights the contributions sourced from the U-Net’s skip connections and backbone feature maps to potentially improve image quality.
+* [`<fast_negative>`](https://civitai.com/models/71961/fast-negative-embedding-fastnegativev2): all-purpose (default)
+* [`<bad_dream>`](https://civitai.com/models/72437?modelVersionId=77169): DreamShaper-style
+* [`<unrealistic_dream>`](https://civitai.com/models/72437?modelVersionId=77173): realistic add-on
 
-#### Clip Skip
+### Styles
 
-When enabled, the last CLIP layer is skipped. This _can_ improve image quality with anime models.
+Styles are prompt templates from twri's [sdxl_prompt_styler](https://github.com/twri/sdxl_prompt_styler) Comfy node. Start with a subject like "cat", pick a style, and iterate from there.
 
 ### Scale
 
@@ -55,7 +43,7 @@ Each model checkpoint has a different aesthetic:
 * [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5): base
 * [sg161222/realistic_vision_v5.1](https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE): photorealistic
 
-#### Schedulers
+### Schedulers
 
 Optionally, the [Karras](https://arxiv.org/abs/2206.00364) noise schedule can be used:
 
@@ -81,6 +69,14 @@ Optionally, the [Karras](https://arxiv.org/abs/2206.00364) noise schedule can be
 
 [Token merging](https://github.com/dbolya/tomesd) (Bolya & Hoffman 2023) reduces the number of tokens processed by the model. Set `Ratio` to the desired reduction factor. ToMe's impact is more noticeable on larger images.
 
+#### FreeU
+
+[FreeU](https://github.com/ChenyangSi/FreeU) (Si et al. 2023) re-weights the contributions sourced from the U-Net’s skip connections and backbone feature maps to potentially improve image quality.
+
+#### Clip Skip
+
+When enabled, the last CLIP layer is skipped. This _can_ improve image quality with anime models.
+
 #### Tiny VAE
 
 Enable [madebyollin/taesd](https://github.com/madebyollin/taesd) for almost instant latent decoding with a minor loss in detail. Useful for development.