Upload Pony_Illustrious_SDXL_Text_Embedding_Generator.ipynb

Pony_Illustrious_SDXL_Text_Embedding_Generator.ipynb

@@ -0,0 +1,149 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "a58c955f"
+      },
+      "source": [
+        "# SDXL Text Embedding Generator\n",
+        "This notebook mounts Google Drive, loads CLIP-L and CLIP-G using your Hugging Face token, and generates dual text embeddings formatted for SDXL models saved to Drive as `.safetensors` files."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title SDXL Text Embedding Generator (TI format)\n",
+        "# Mount Google Drive & Install\n",
+        "from google.colab import drive\n",
+        "import os\n",
+        "drive.mount('/content/drive')\n",
+        "\n",
+        "!pip install -q transformers huggingface_hub safetensors torch\n",
+        "\n",
+        "#@title Load Models\n",
+        "import torch\n",
+        "from google.colab import userdata\n",
+        "from huggingface_hub import login\n",
+        "from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextModelWithProjection\n",
+        "from safetensors.torch import save_file\n",
+        "\n",
+        "# Login (optional but recommended)\n",
+        "try:\n",
+        "    hf_token = userdata.get('HF_TOKEN')\n",
+        "    login(token=hf_token)\n",
+        "except Exception as e:\n",
+        "    print(f\"HF token warning: {e}\")\n",
+        "\n",
+        "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
+        "print(f\"Using device: {device}\")\n",
+        "\n",
+        "# Correct model IDs\n",
+        "clip_l_id = \"openai/clip-vit-large-patch14\"\n",
+        "clip_g_id = \"laion/CLIP-ViT-bigG-14-laion2B-39B-b160k\"\n",
+        "\n",
+        "print(\"Loading CLIP-L...\")\n",
+        "tokenizer_l = CLIPTokenizer.from_pretrained(clip_l_id)\n",
+        "text_encoder_l = CLIPTextModel.from_pretrained(clip_l_id).to(device)\n",
+        "text_encoder_l.eval()\n",
+        "\n",
+        "print(\"Loading CLIP-G...\")\n",
+        "tokenizer_g = CLIPTokenizer.from_pretrained(clip_g_id)\n",
+        "text_encoder_g = CLIPTextModelWithProjection.from_pretrained(clip_g_id).to(device)\n",
+        "text_encoder_g.eval()\n",
+        "\n",
+        "print(\"Models loaded.\")"
+      ],
+      "metadata": {
+        "id": "3L8XEw6RoRvF"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title Generate Embedding\n",
+        "#@markdown ### Settings\n",
+        "text_to_encode = \"A cinematic shot of a majestic castle on a mountain\"  #@param {type:\"string\"}\n",
+        "embedding_file_name = \"castle_embedding\"  #@param {type:\"string\"}\n",
+        "keep_special_tokens = False  # Keep false. This setting minimizes the embedding sized. If true , the embedding is 77 tokens long , leaving no room for the user to prompt anything with it inside a 77 token batch encoding. Keep in mind , a 77 token embedding will still sort of work , as the prompt will extend to 154 tokens and take the average of the two 77 token batch encodings\n",
+        "\n",
+        "# Tokenize\n",
+        "inputs_l = tokenizer_l(\n",
+        "    [text_to_encode],\n",
+        "    padding=\"max_length\",\n",
+        "    max_length=77,\n",
+        "    truncation=True,\n",
+        "    return_tensors=\"pt\"\n",
+        ").to(device)\n",
+        "\n",
+        "inputs_g = tokenizer_g(\n",
+        "    [text_to_encode],\n",
+        "    padding=\"max_length\",\n",
+        "    max_length=77,\n",
+        "    truncation=True,\n",
+        "    return_tensors=\"pt\"\n",
+        ").to(device)\n",
+        "\n",
+        "with torch.no_grad():\n",
+        "    # Get INPUT embeddings (token embedding layer) – this is what TI files use\n",
+        "    emb_l = text_encoder_l.get_input_embeddings()(inputs_l.input_ids)  # [1, 77, 768]\n",
+        "    emb_g = text_encoder_g.get_input_embeddings()(inputs_g.input_ids)  # [1, 77, 1280]\n",
+        "\n",
+        "# Optional: strip special tokens (BOS, EOS, padding)\n",
+        "if not keep_special_tokens:\n",
+        "    # Find actual content tokens (non-pad)\n",
+        "    # CLIP pad token is usually 0\n",
+        "    mask_l = inputs_l.input_ids[0] != tokenizer_l.pad_token_id\n",
+        "    mask_g = inputs_g.input_ids[0] != tokenizer_g.pad_token_id\n",
+        "\n",
+        "    # Also drop BOS (usually first token) and EOS if present\n",
+        "    # Safer approach: keep everything that is not pad\n",
+        "    emb_l = emb_l[0, mask_l]   # [N, 768]\n",
+        "    emb_g = emb_g[0, mask_g]   # [N, 1280]\n",
+        "else:\n",
+        "    emb_l = emb_l.squeeze(0)   # [77, 768]\n",
+        "    emb_g = emb_g.squeeze(0)   # [77, 1280]\n",
+        "\n",
+        "print(f\"CLIP-L shape: {emb_l.shape}\")\n",
+        "print(f\"CLIP-G shape: {emb_g.shape}\")\n",
+        "\n",
+        "# Save\n",
+        "save_dir = \"/content/drive/MyDrive/SDXL_Embeddings\"\n",
+        "os.makedirs(save_dir, exist_ok=True)\n",
+        "save_path = os.path.join(save_dir, f\"{embedding_file_name}.safetensors\")\n",
+        "\n",
+        "tensors_to_save = {\n",
+        "    \"clip_l\": emb_l.cpu().contiguous(),\n",
+        "    \"clip_g\": emb_g.cpu().contiguous()\n",
+        "}\n",
+        "\n",
+        "save_file(tensors_to_save, save_path)\n",
+        "print(f\"\\nSuccessfully saved: {save_path}\")\n",
+        "print(\"You can now use it as:  embedding:castle_embedding\")"
+      ],
+      "metadata": {
+        "id": "XquwZtTIow0p"
+      },
+      "execution_count": null,
+      "outputs": []
+    }
+  ],
+  "metadata": {
+    "colab": {
+      "provenance": [],
+      "gpuType": "T4"
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python"
+    },
+    "accelerator": "GPU"
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}