File size: 11,732 Bytes
f02144e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 | {
"nbformat": 4,
"nbformat_minor": 5,
"metadata": {
"colab": {
"provenance": [],
"gpuType": "T4"
},
"kernelspec": {
"name": "python3",
"display_name": "Python 3"
},
"language_info": {
"name": "python"
},
"accelerator": "GPU"
},
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Aniimage-1 Colab Generator\n",
"\n",
"Hugging Face's default `DiffusionPipeline.from_pretrained(...)` snippet does not work for this repo because `8BitStudio/Aniimage-1` is not packaged as a full Diffusers pipeline. It ships the Aniimage UNet weights/config, so this notebook manually loads:\n",
"\n",
"- Aniimage-1 UNet from `8BitStudio/Aniimage-1`\n",
"- SD VAE from `stabilityai/sd-vae-ft-mse`\n",
"- CLIP text encoder/tokenizer from `openai/clip-vit-large-patch14`\n",
"\n",
"Use Colab GPU runtime: `Runtime > Change runtime type > T4 GPU` or better.\n",
"If you have any errors or issues, open a discussion on the Hugging Face repo."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!pip -q install -U diffusers transformers accelerate safetensors huggingface_hub pillow"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import json\n",
"from pathlib import Path\n",
"\n",
"import numpy as np\n",
"import torch\n",
"import torch.nn.functional as F\n",
"from PIL import Image, ImageEnhance, ImageFilter\n",
"from IPython.display import display\n",
"from huggingface_hub import hf_hub_download\n",
"from safetensors.torch import load_file\n",
"from diffusers import (\n",
" AutoencoderKL,\n",
" DDIMScheduler,\n",
" DPMSolverMultistepScheduler,\n",
" EulerAncestralDiscreteScheduler,\n",
" EulerDiscreteScheduler,\n",
" UNet2DConditionModel,\n",
")\n",
"from transformers import CLIPTextModel, CLIPTokenizer\n",
"\n",
"REPO_ID = \"8BitStudio/Aniimage-1\"\n",
"VAE_ID = \"stabilityai/sd-vae-ft-mse\"\n",
"CLIP_ID = \"openai/clip-vit-large-patch14\"\n",
"MODEL_DIR = Path(\"/content/aniimage_1\")\n",
"\n",
"device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
"autocast_dtype = torch.float16 if device == \"cuda\" else torch.float32\n",
"\n",
"if device == \"cuda\":\n",
" torch.backends.cuda.matmul.allow_tf32 = True\n",
" torch.backends.cudnn.allow_tf32 = True\n",
"\n",
"print(\"Device:\", device)\n",
"if device == \"cuda\":\n",
" print(\"GPU:\", torch.cuda.get_device_name(0))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def make_scheduler(name=\"DPM++ 2M Karras\"):\n",
" base = dict(\n",
" num_train_timesteps=1000,\n",
" beta_schedule=\"scaled_linear\",\n",
" prediction_type=\"epsilon\",\n",
" )\n",
" if name == \"DPM++ 2M Karras\":\n",
" return DPMSolverMultistepScheduler(\n",
" **base,\n",
" algorithm_type=\"dpmsolver++\",\n",
" solver_order=2,\n",
" use_karras_sigmas=True,\n",
" )\n",
" if name == \"DPM++ SDE Karras\":\n",
" return DPMSolverMultistepScheduler(\n",
" **base,\n",
" algorithm_type=\"sde-dpmsolver++\",\n",
" solver_order=2,\n",
" use_karras_sigmas=True,\n",
" )\n",
" if name == \"Euler a\":\n",
" return EulerAncestralDiscreteScheduler(**base)\n",
" if name == \"Euler\":\n",
" return EulerDiscreteScheduler(**base)\n",
" if name == \"DDIM\":\n",
" return DDIMScheduler(**base, clip_sample=False, set_alpha_to_one=False)\n",
" raise ValueError(f\"Unknown scheduler: {name}\")\n",
"\n",
"\n",
"def load_aniimage1():\n",
" MODEL_DIR.mkdir(parents=True, exist_ok=True)\n",
"\n",
" print(\"Downloading Aniimage-1 UNet config/weights if needed...\")\n",
" config_path = hf_hub_download(\n",
" repo_id=REPO_ID,\n",
" filename=\"config.json\",\n",
" local_dir=str(MODEL_DIR),\n",
" )\n",
" weights_path = hf_hub_download(\n",
" repo_id=REPO_ID,\n",
" filename=\"diffusion_pytorch_model.safetensors\",\n",
" local_dir=str(MODEL_DIR),\n",
" )\n",
"\n",
" print(\"Loading VAE...\")\n",
" vae = AutoencoderKL.from_pretrained(VAE_ID).to(device).eval()\n",
" vae.requires_grad_(False)\n",
"\n",
" print(\"Loading CLIP tokenizer/text encoder...\")\n",
" tokenizer = CLIPTokenizer.from_pretrained(CLIP_ID)\n",
" text_encoder = CLIPTextModel.from_pretrained(CLIP_ID).to(device).eval()\n",
" text_encoder.requires_grad_(False)\n",
"\n",
" print(\"Loading Aniimage-1 UNet...\")\n",
" with open(config_path, \"r\", encoding=\"utf-8\") as f:\n",
" unet_config = json.load(f)\n",
" unet = UNet2DConditionModel.from_config(unet_config)\n",
" state = load_file(weights_path, device=\"cpu\")\n",
" unet.load_state_dict(state)\n",
" unet.to(device).eval()\n",
" unet.requires_grad_(False)\n",
"\n",
" print(\"Aniimage-1 ready.\")\n",
" return vae, tokenizer, text_encoder, unet\n",
"\n",
"\n",
"vae, tokenizer, text_encoder, unet = load_aniimage1()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def encode_prompt(prompt, negative_prompt=\"\"):\n",
" prompts = [negative_prompt or \"\", prompt]\n",
" tokens = tokenizer(\n",
" prompts,\n",
" padding=\"max_length\",\n",
" max_length=tokenizer.model_max_length,\n",
" truncation=True,\n",
" return_tensors=\"pt\",\n",
" )\n",
" with torch.no_grad():\n",
" return text_encoder(tokens.input_ids.to(device))[0]\n",
"\n",
"\n",
"def sharpen_latents(latents, amount=0.08):\n",
" blurred = F.avg_pool2d(latents, kernel_size=3, stride=1, padding=1)\n",
" return latents + amount * (latents - blurred)\n",
"\n",
"\n",
"def post_process(img):\n",
" img = img.filter(ImageFilter.UnsharpMask(radius=1.5, percent=40, threshold=2))\n",
" img = ImageEnhance.Contrast(img).enhance(1.06)\n",
" img = ImageEnhance.Color(img).enhance(1.10)\n",
" return img\n",
"\n",
"\n",
"@torch.no_grad()\n",
"def decode_latents(latents, polish=True):\n",
" scaled = latents / vae.config.scaling_factor\n",
" # Keep decode in fp32. It is a little slower but avoids washed/banded VAE output.\n",
" image = vae.decode(scaled.float()).sample\n",
" image = (image.float() / 2 + 0.5).clamp(0, 1)\n",
" image = image.cpu().permute(0, 2, 3, 1).numpy()[0]\n",
" image = (image * 255).round().astype(\"uint8\")\n",
" img = Image.fromarray(image)\n",
" return post_process(img) if polish else img\n",
"\n",
"\n",
"@torch.no_grad()\n",
"def generate_aniimage1(\n",
" prompt,\n",
" negative_prompt=\"\",\n",
" scheduler_name=\"DPM++ 2M Karras\",\n",
" steps=25,\n",
" guidance_scale=7.5,\n",
" seed=-1,\n",
" polish=True,\n",
"):\n",
" if seed is None or int(seed) < 0:\n",
" seed = int(torch.randint(0, 2**31 - 1, (1,)).item())\n",
" else:\n",
" seed = int(seed)\n",
"\n",
" generator = torch.Generator(device=device).manual_seed(seed)\n",
" text_embeddings = encode_prompt(prompt, negative_prompt)\n",
"\n",
" scheduler = make_scheduler(scheduler_name)\n",
" scheduler.set_timesteps(int(steps), device=device)\n",
"\n",
" latents = torch.randn(\n",
" (1, 4, 32, 32),\n",
" generator=generator,\n",
" device=device,\n",
" dtype=torch.float32,\n",
" )\n",
" latents = latents * scheduler.init_noise_sigma\n",
"\n",
" for t in scheduler.timesteps:\n",
" latent_input = torch.cat([latents, latents], dim=0)\n",
" latent_input = scheduler.scale_model_input(latent_input, t)\n",
"\n",
" with torch.autocast(\n",
" device_type=\"cuda\",\n",
" dtype=autocast_dtype,\n",
" enabled=(device == \"cuda\"),\n",
" ):\n",
" noise_pred = unet(\n",
" latent_input,\n",
" t,\n",
" encoder_hidden_states=text_embeddings,\n",
" ).sample\n",
"\n",
" noise_uncond, noise_text = noise_pred.chunk(2)\n",
" noise_pred = noise_uncond + float(guidance_scale) * (noise_text - noise_uncond)\n",
" latents = scheduler.step(noise_pred, t, latents).prev_sample\n",
"\n",
" if polish:\n",
" latents = sharpen_latents(latents)\n",
" return decode_latents(latents, polish=polish), seed"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"prompt = \"an anime girl with long blue hair\" #@param {type:\"string\"}\n",
"negative_prompt = \"low quality, ugly, blurry, distorted, deformed, bad anatomy, bad proportions, extra limbs, missing limbs, watermark, text, signature, washed out, flat colors, manga panel, disfigured, poorly drawn, jpeg artifacts, cropped, out of frame\" #@param {type:\"string\"}\n",
"scheduler_name = \"DPM++ 2M Karras\" #@param [\"DPM++ 2M Karras\", \"DPM++ SDE Karras\", \"Euler a\", \"Euler\", \"DDIM\"]\n",
"steps = 25 #@param {type:\"slider\", min:10, max:80, step:1}\n",
"guidance_scale = 7.5 #@param {type:\"slider\", min:1, max:15, step:0.1}\n",
"seed = -1 #@param {type:\"integer\"}\n",
"polish = False #@param {type:\"boolean\"}\n",
"\n",
"image, used_seed = generate_aniimage1(\n",
" prompt=prompt,\n",
" negative_prompt=negative_prompt,\n",
" scheduler_name=scheduler_name,\n",
" steps=steps,\n",
" guidance_scale=guidance_scale,\n",
" seed=seed,\n",
" polish=polish,\n",
")\n",
"\n",
"out_path = \"/content/aniimage1_output.png\"\n",
"image.save(out_path)\n",
"print(\"Seed:\", used_seed)\n",
"print(\"Saved:\", out_path)\n",
"display(image)"
]
}
]
}
|