| 1 | #!/usr/bin/env python3 |
| 2 | """ |
| 3 | Gemini Image Generation Backend |
| 4 | |
| 5 | Generates or edits images via the Google GenAI API (Gemini). |
| 6 | Used by image_gen.py as a backend module. |
| 7 | |
| 8 | Configuration keys: |
| 9 | GEMINI_API_KEY (required) |
| 10 | GEMINI_BASE_URL (optional) Custom API endpoint for proxy services |
| 11 | GEMINI_MODEL (optional) Override default model |
| 12 | |
| 13 | Dependencies: |
| 14 | pip install google-genai Pillow |
| 15 | """ |
| 16 | |
| 17 | import sys |
| 18 | from pathlib import Path |
| 19 | |
| 20 | _SCRIPTS_DIR = Path(__file__).resolve().parents[1] |
| 21 | if str(_SCRIPTS_DIR) not in sys.path: |
| 22 | sys.path.insert(0, str(_SCRIPTS_DIR)) |
| 23 | |
| 24 | from console_encoding import configure_utf8_stdio # noqa: E402 |
| 25 | |
| 26 | configure_utf8_stdio() |
| 27 | |
| 28 | if __name__ == "__main__": |
| 29 | print(__doc__) |
| 30 | print("Use via: python3 skills/ppt-master/scripts/image_gen.py \"prompt\" --backend gemini") |
| 31 | raise SystemExit(0 if any(arg in {"-h", "--help", "help"} for arg in sys.argv[1:]) else 1) |
| 32 | |
| 33 | import os |
| 34 | import time |
| 35 | import threading |
| 36 | from google import genai |
| 37 | from google.genai import types |
| 38 | from image_backends.backend_common import ( |
| 39 | MAX_RETRIES, |
| 40 | is_permanent_error, |
| 41 | is_rate_limit_error, |
| 42 | normalize_image_size, |
| 43 | resolve_output_path, |
| 44 | retry_delay, |
| 45 | save_image_bytes, |
| 46 | ) |
| 47 | |
| 48 | |
| 49 | # ╔══════════════════════════════���═══════════════════════════════════╗ |
| 50 | # ║ Constants ║ |
| 51 | # ╚═══════════��═════════════════════════════��════════════════════════╝ |
| 52 | |
| 53 | VALID_ASPECT_RATIOS = [ |
| 54 | "1:1", "1:4", "1:8", |
| 55 | "2:3", "3:2", "3:4", "4:1", "4:3", |
| 56 | "4:5", "5:4", "8:1", "9:16", "16:9", "21:9" |
| 57 | ] |
| 58 | |
| 59 | VALID_IMAGE_SIZES = ["512px", "1K", "2K", "4K"] |
| 60 | |
| 61 | DEFAULT_MODEL = "gemini-3.1-flash-image" |
| 62 | |
| 63 | GEMINI_2_5_FLASH_IMAGE_MODELS = { |
| 64 | "gemini-2.5-flash-image", |
| 65 | "gemini-2.5-flash-image-preview", |
| 66 | } |
| 67 | GEMINI_2_5_VALID_ASPECT_RATIOS = [ |
| 68 | "1:1", "2:3", "3:2", "3:4", "4:3", |
| 69 | "4:5", "5:4", "9:16", "16:9", "21:9", |
| 70 | ] |
| 71 | GEMINI_2_5_VALID_IMAGE_SIZES = ["1K"] |
| 72 | |
| 73 | MINIMAL_THINKING_MODELS = { |
| 74 | "gemini-3.1-flash-image", |
| 75 | "gemini-3.1-flash-image-preview", |
| 76 | } |
| 77 | |
| 78 | REFERENCE_IMAGE_MIME_TYPES = { |
| 79 | ".heic": "image/heic", |
| 80 | ".heif": "image/heif", |
| 81 | ".jpeg": "image/jpeg", |
| 82 | ".jpg": "image/jpeg", |
| 83 | ".png": "image/png", |
| 84 | ".webp": "image/webp", |
| 85 | } |
| 86 | |
| 87 | # Signals to image_gen.py that this backend accepts a reference_image as |
| 88 | # multimodal input to the same generate_content request used for generation. |
| 89 | SUPPORTS_REFERENCE_IMAGE = True |
| 90 | |
| 91 | |
| 92 | def _model_id(model: str) -> str: |
| 93 | """Return the final model path component for official capability checks.""" |
| 94 | return (model or "").strip().lower().rsplit("/", 1)[-1] |
| 95 | |
| 96 | |
| 97 | def _request_image_size(image_size: str) -> str: |
| 98 | """Map the public 512px preset to Gemini's API value.""" |
| 99 | return "512" if image_size == "512px" else image_size |
| 100 | |
| 101 | |
| 102 | def _validate_model_options(model: str, aspect_ratio: str, image_size: str) -> None: |
| 103 | """Enforce limits only for Gemini models with a known narrower contract.""" |
| 104 | if _model_id(model) not in GEMINI_2_5_FLASH_IMAGE_MODELS: |
| 105 | return |
| 106 | if aspect_ratio not in GEMINI_2_5_VALID_ASPECT_RATIOS: |
| 107 | raise ValueError( |
| 108 | f"Invalid aspect ratio '{aspect_ratio}' for {model}. " |
| 109 | f"Valid: {GEMINI_2_5_VALID_ASPECT_RATIOS}" |
| 110 | ) |
| 111 | if image_size not in GEMINI_2_5_VALID_IMAGE_SIZES: |
| 112 | raise ValueError( |
| 113 | f"Invalid image size '{image_size}' for {model}. " |
| 114 | f"Valid: {GEMINI_2_5_VALID_IMAGE_SIZES}" |
| 115 | ) |
| 116 | |
| 117 | |
| 118 | def _reference_image_mime_type(reference_image: str) -> str: |
| 119 | """Validate one local reference image and return its Gemini MIME type.""" |
| 120 | image_path = Path(reference_image) |
| 121 | if not image_path.is_file(): |
| 122 | raise FileNotFoundError(f"Reference image file not found: {reference_image}") |
| 123 | |
| 124 | mime_type = REFERENCE_IMAGE_MIME_TYPES.get(image_path.suffix.lower()) |
| 125 | if mime_type is None: |
| 126 | extensions = ", ".join(sorted(REFERENCE_IMAGE_MIME_TYPES)) |
| 127 | raise ValueError( |
| 128 | f"Unsupported Gemini reference image format '{image_path.suffix or '<none>'}'. " |
| 129 | f"Supported extensions: {extensions}" |
| 130 | ) |
| 131 | return mime_type |
| 132 | |
| 133 | |
| 134 | def _reference_image_part(reference_image: str) -> types.Part: |
| 135 | """Load one supported local image as a Gemini inline-data part.""" |
| 136 | image_path = Path(reference_image) |
| 137 | mime_type = _reference_image_mime_type(reference_image) |
| 138 | |
| 139 | return types.Part.from_bytes( |
| 140 | data=image_path.read_bytes(), |
| 141 | mime_type=mime_type, |
| 142 | ) |
| 143 | |
| 144 | |
| 145 | # ╔══════���═══════════════════════════════════════════════��═══════════╗ |
| 146 | # ║ Image Generation and Editing ║ |
| 147 | # ╚══════════════════════════════════════════════════════════════════╝ |
| 148 | |
| 149 | def _generate_image(api_key: str, prompt: str, |
| 150 | aspect_ratio: str = "1:1", image_size: str = "1K", |
| 151 | output_dir: str = None, filename: str = None, |
| 152 | model: str = DEFAULT_MODEL, base_url: str = None, |
| 153 | reference_image: str = None) -> str: |
| 154 | """ |
| 155 | Image generation or editing via Gemini API (streaming). |
| 156 | |
| 157 | Returns: |
| 158 | Path of the saved image file |
| 159 | |
| 160 | Raises: |
| 161 | RuntimeError: When generation fails |
| 162 | """ |
| 163 | if base_url: |
| 164 | client = genai.Client(api_key=api_key, http_options={'base_url': base_url}) |
| 165 | else: |
| 166 | client = genai.Client(api_key=api_key) |
| 167 | |
| 168 | config_kwargs = { |
| 169 | "response_modalities": ["IMAGE"], |
| 170 | "image_config": types.ImageConfig( |
| 171 | aspect_ratio=aspect_ratio, |
| 172 | image_size=_request_image_size(image_size), |
| 173 | ), |
| 174 | } |
| 175 | if _model_id(model) in MINIMAL_THINKING_MODELS: |
| 176 | config_kwargs["thinking_config"] = types.ThinkingConfig( |
| 177 | thinking_level="MINIMAL", |
| 178 | ) |
| 179 | config = types.GenerateContentConfig(**config_kwargs) |
| 180 | |
| 181 | contents = [prompt] |
| 182 | if reference_image is not None: |
| 183 | contents.append(_reference_image_part(reference_image)) |
| 184 | |
| 185 | mode_label = "Proxy Mode" if base_url else "Official Mode" |
| 186 | print(f"[Gemini - {mode_label}]") |
| 187 | if base_url: |
| 188 | print(f" Base URL: {base_url}") |
| 189 | print(f" Model: {model}") |
| 190 | print(f" Prompt: {prompt[:120]}{'...' if len(prompt) > 120 else ''}") |
| 191 | if reference_image is not None: |
| 192 | print(f" Reference: {reference_image}") |
| 193 | print(f" Aspect Ratio: {aspect_ratio}") |
| 194 | print(f" Image Size: {image_size}") |
| 195 | print() |
| 196 | |
| 197 | start_time = time.time() |
| 198 | operation = "Editing" if reference_image is not None else "Generating" |
| 199 | print(f" [..] {operation}...", end="", flush=True) |
| 200 | |
| 201 | heartbeat_stop = threading.Event() |
| 202 | |
| 203 | def _heartbeat(): |
| 204 | while not heartbeat_stop.is_set(): |
| 205 | heartbeat_stop.wait(5) |
| 206 | if not heartbeat_stop.is_set(): |
| 207 | elapsed = time.time() - start_time |
| 208 | print(f" {elapsed:.0f}s...", end="", flush=True) |
| 209 | |
| 210 | hb_thread = threading.Thread(target=_heartbeat, daemon=True) |
| 211 | hb_thread.start() |
| 212 | |
| 213 | last_image_data = None |
| 214 | chunk_count = 0 |
| 215 | total_bytes = 0 |
| 216 | |
| 217 | for chunk in client.models.generate_content_stream( |
| 218 | model=model, |
| 219 | contents=contents, |
| 220 | config=config, |
| 221 | ): |
| 222 | elapsed = time.time() - start_time |
| 223 | |
| 224 | if chunk.parts is None: |
| 225 | continue |
| 226 | |
| 227 | for part in chunk.parts: |
| 228 | if part.text is not None: |
| 229 | print(f"\n Model says: {part.text}", end="", flush=True) |
| 230 | elif part.inline_data is not None: |
| 231 | chunk_count += 1 |
| 232 | data_size = len(part.inline_data.data) if part.inline_data.data else 0 |
| 233 | total_bytes += data_size |
| 234 | size_str = f"{data_size / 1024:.0f}KB" if data_size < 1048576 else f"{data_size / 1048576:.1f}MB" |
| 235 | print(f"\n [OK] Chunk #{chunk_count} received ({size_str}, {elapsed:.1f}s)", end="", flush=True) |
| 236 | last_image_data = part |
| 237 | |
| 238 | heartbeat_stop.set() |
| 239 | hb_thread.join(timeout=1) |
| 240 | |
| 241 | elapsed = time.time() - start_time |
| 242 | print(f"\n [DONE] Stream complete ({elapsed:.1f}s, {chunk_count} chunk(s), {total_bytes / 1024:.0f}KB total)") |
| 243 | |
| 244 | if last_image_data is not None and last_image_data.inline_data is not None: |
| 245 | if chunk_count > 1: |
| 246 | print(f" Keeping the final chunk (highest quality).") |
| 247 | path = resolve_output_path(prompt, output_dir, filename, ".png") |
| 248 | return save_image_bytes( |
| 249 | last_image_data.inline_data.data, |
| 250 | path, |
| 251 | content_type=getattr(last_image_data.inline_data, "mime_type", None), |
| 252 | ) |
| 253 | |
| 254 | raise RuntimeError("No image was generated. The server may have refused the request.") |
| 255 | |
| 256 | |
| 257 | # ╔���═══════════════════════════════════���═════════════════════════════╗ |
| 258 | # ║ Public Entry Point ║ |
| 259 | # ╚═════════════���═══════════════════════════���════════════════════════╝ |
| 260 | |
| 261 | def generate(prompt: str, |
| 262 | aspect_ratio: str = "1:1", image_size: str = "1K", |
| 263 | output_dir: str = None, filename: str = None, |
| 264 | model: str = None, max_retries: int = MAX_RETRIES, |
| 265 | reference_image: str = None) -> str: |
| 266 | """ |
| 267 | Gemini image generation or image-to-image editing with automatic retry. |
| 268 | |
| 269 | Reads credentials from the current process environment or a `.env` file: |
| 270 | GEMINI_API_KEY |
| 271 | GEMINI_BASE_URL |
| 272 | GEMINI_MODEL (optional override) |
| 273 | |
| 274 | Args: |
| 275 | prompt: Prompt text (edit instruction when reference_image is set) |
| 276 | aspect_ratio: Aspect ratio (e.g. "16:9", "1:1") |
| 277 | image_size: Image size ("512px", "1K", "2K", "4K", case-insensitive) |
| 278 | output_dir: Output directory |
| 279 | filename: Output filename (without extension) |
| 280 | model: Model name (default: gemini-3.1-flash-image) |
| 281 | max_retries: Maximum number of retries |
| 282 | reference_image: Optional source image path. When set, the image and |
| 283 | edit instruction are sent together as multimodal input. |
| 284 | |
| 285 | Returns: |
| 286 | Path of the saved image file |
| 287 | """ |
| 288 | api_key = os.environ.get("GEMINI_API_KEY") |
| 289 | base_url = os.environ.get("GEMINI_BASE_URL") |
| 290 | |
| 291 | if not api_key: |
| 292 | raise ValueError( |
| 293 | "No API key found. Set GEMINI_API_KEY in the current environment or a .env file." |
| 294 | ) |
| 295 | |
| 296 | if model is None: |
| 297 | model = os.environ.get("GEMINI_MODEL") or DEFAULT_MODEL |
| 298 | |
| 299 | image_size = normalize_image_size(image_size) |
| 300 | |
| 301 | if aspect_ratio not in VALID_ASPECT_RATIOS: |
| 302 | raise ValueError(f"Invalid aspect ratio '{aspect_ratio}'. Valid: {VALID_ASPECT_RATIOS}") |
| 303 | |
| 304 | if image_size not in VALID_IMAGE_SIZES: |
| 305 | raise ValueError(f"Invalid image size '{image_size}'. Valid: {VALID_IMAGE_SIZES}") |
| 306 | |
| 307 | _validate_model_options(model, aspect_ratio, image_size) |
| 308 | |
| 309 | if reference_image is not None: |
| 310 | # Validate local input before entering the retry loop. API failures can |
| 311 | # be transient; an unsupported or missing source image cannot be fixed |
| 312 | # by resending the same request. |
| 313 | _reference_image_mime_type(reference_image) |
| 314 | |
| 315 | last_error = None |
| 316 | for attempt in range(max_retries + 1): |
| 317 | try: |
| 318 | return _generate_image(api_key, prompt, |
| 319 | aspect_ratio, image_size, output_dir, |
| 320 | filename, model, base_url, reference_image) |
| 321 | except Exception as e: |
| 322 | last_error = e |
| 323 | if is_permanent_error(e): |
| 324 | raise |
| 325 | if attempt < max_retries and is_rate_limit_error(e): |
| 326 | delay = retry_delay(attempt, rate_limited=True) |
| 327 | print(f"\n [WARN] Rate limit hit (attempt {attempt + 1}/{max_retries + 1}). " |
| 328 | f"Waiting {delay}s before retry...") |
| 329 | time.sleep(delay) |
| 330 | elif attempt < max_retries: |
| 331 | delay = retry_delay(attempt, rate_limited=False) |
| 332 | print(f"\n [WARN] Error (attempt {attempt + 1}/{max_retries + 1}): {e}. " |
| 333 | f"Retrying in {delay}s...") |
| 334 | time.sleep(delay) |
| 335 | else: |
| 336 | break |
| 337 | |
| 338 | raise RuntimeError(f"Failed after {max_retries + 1} attempts. Last error: {last_error}") |
| 339 |