File size: 12,652 Bytes
d13a83d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 | #!/usr/bin/env python3
"""Assess, edit, and evaluate one image with the released MR-IQA-2 models.
Actor, Editor, and Judge run sequentially, so all three stages can reuse one
physical GPU without sharing model memory.
"""
from __future__ import annotations
import argparse
import json
import math
import os
import shutil
import subprocess
import sys
from pathlib import Path
from typing import Any, Callable
REPO_ID = "RobinY99/MR-IQA-2"
SEED = 764952063587760
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Run MR-IQA-2 Actor, Editor, and Judge on one GPU."
)
parser.add_argument("image", help="Input image")
parser.add_argument("--output-dir", default="outputs/quick_start")
parser.add_argument("--gpu", default="0", help="Physical GPU index")
parser.add_argument("--actor-env", default="mr_iqa_actor_judge")
parser.add_argument("--editor-env", default="mr_iqa_editor")
parser.add_argument("--actor-python", default="")
parser.add_argument("--editor-python", default="")
parser.add_argument("--actor-model", default=REPO_ID)
parser.add_argument("--actor-subfolder", default="actor")
parser.add_argument("--judge-model", default=REPO_ID)
parser.add_argument("--judge-subfolder", default="judge")
parser.add_argument(
"--editor-model",
default="",
help="Local Editor directory; omitted means REPO_ID/editor",
)
parser.add_argument(
"--revision",
default="",
help="Optional Hugging Face revision; defaults to the current model release",
)
parser.add_argument("--max-pixels", type=int, default=196608)
parser.add_argument("--local-files-only", action="store_true")
parser.add_argument(
"--stage",
choices=("actor", "editor", "judge"),
help=argparse.SUPPRESS,
)
args = parser.parse_args(argv)
if args.max_pixels < 256:
parser.error("--max-pixels must be at least 256")
return args
def _actor_stage(args: argparse.Namespace, image_path: Path, output_dir: Path) -> None:
from actor_to_editor import (
atomic_write_json,
atomic_write_text,
generate_actor_completion,
parse_valid_actor_output,
)
actor_args = argparse.Namespace(
actor_model=args.actor_model,
actor_subfolder=args.actor_subfolder,
actor_revision=(
""
if Path(args.actor_model).expanduser().is_dir()
else args.revision
),
device="cuda:0",
dtype="bfloat16",
attn_implementation="sdpa",
max_new_tokens=512,
max_pixels=args.max_pixels,
seed=SEED,
local_files_only=args.local_files_only,
)
completion = generate_actor_completion(image_path, actor_args)
payload = parse_valid_actor_output(completion)
atomic_write_text(output_dir / "actor_raw.txt", completion)
atomic_write_json(output_dir / "assessment.json", payload)
print(json.dumps(payload, ensure_ascii=False, indent=2))
def _editor_size(width: int, height: int, max_pixels: int) -> tuple[int, int]:
scale = min(1.0, math.sqrt(float(max_pixels) / float(width * height)))
resized_width = max(16, int(width * scale) // 16 * 16)
resized_height = max(16, int(height * scale) // 16 * 16)
return resized_width, resized_height
def _editor_stage(args: argparse.Namespace, image_path: Path, output_dir: Path) -> None:
import torch
from diffusers import Flux2KleinPipeline
from huggingface_hub import snapshot_download
from PIL import Image
from actor_to_editor import atomic_write_json
assessment_path = output_dir / "assessment.json"
assessment = json.loads(assessment_path.read_text(encoding="utf-8"))
reasoning = assessment.get("reasoning")
if not isinstance(reasoning, dict):
raise ValueError("assessment.json has no reasoning object")
solution = reasoning.get("solution")
if not isinstance(solution, str) or not solution.strip():
raise ValueError("assessment.json has no usable solution")
if args.editor_model:
editor_path = Path(args.editor_model).expanduser().resolve(strict=True)
else:
download_kwargs: dict[str, Any] = {
"repo_id": REPO_ID,
"allow_patterns": ["editor/**"],
"local_files_only": args.local_files_only,
}
if args.revision:
download_kwargs["revision"] = args.revision
snapshot = Path(
snapshot_download(**download_kwargs)
)
editor_path = snapshot / "editor"
editor = Flux2KleinPipeline.from_pretrained(
editor_path,
torch_dtype=torch.bfloat16,
local_files_only=True,
).to("cuda")
with Image.open(image_path) as opened:
source = opened.convert("RGB")
edit_width, edit_height = _editor_size(
source.width,
source.height,
args.max_pixels,
)
editor_input = source.resize(
(edit_width, edit_height),
Image.Resampling.LANCZOS,
)
with torch.inference_mode():
edited = editor(
prompt=solution,
image=editor_input,
width=edit_width,
height=edit_height,
num_inference_steps=4,
sigmas=[1.0, 0.75, 0.5, 0.25],
guidance_scale=1.0,
generator=torch.Generator(device="cuda").manual_seed(SEED),
max_sequence_length=512,
text_encoder_out_layers=(9, 18, 27),
).images[0].convert("RGB")
if edited.size != source.size:
edited = edited.resize(source.size, Image.Resampling.LANCZOS)
edited_path = output_dir / "edited.png"
edited.save(edited_path)
result = {
"input_image": str(image_path),
"assessment": assessment,
"edited_image": str(edited_path),
"actor_model": (
f"{args.actor_model}/{args.actor_subfolder}"
if args.actor_subfolder and not Path(args.actor_model).expanduser().is_dir()
else args.actor_model
),
"editor_model": args.editor_model or f"{REPO_ID}/editor",
"revision": args.revision or None,
"seed": SEED,
"original_size": [source.width, source.height],
"inference_size": [edit_width, edit_height],
"solution_forwarded_verbatim": (
solution == assessment["reasoning"]["solution"]
),
"judge": None,
}
atomic_write_json(output_dir / "result.json", result)
def _quality_delta(original: dict[str, Any], edited: dict[str, Any]) -> float:
original_score = original.get("mean")
edited_score = edited.get("mean")
if not isinstance(original_score, (int, float)):
raise ValueError("Judge did not return a valid score for the input image")
if not isinstance(edited_score, (int, float)):
raise ValueError("Judge did not return a valid score for the edited image")
return float(edited_score) - float(original_score)
def _resolve_model_subfolder(
model: str,
subfolder: str,
revision: str,
local_files_only: bool,
) -> Path:
from huggingface_hub import snapshot_download
local_path = Path(model).expanduser()
if local_path.is_dir():
candidate = (
local_path
if (local_path / "config.json").is_file()
else local_path / subfolder
)
return candidate.resolve(strict=True)
download_kwargs: dict[str, Any] = {
"repo_id": model,
"allow_patterns": [f"{subfolder}/**"],
"local_files_only": local_files_only,
}
if revision:
download_kwargs["revision"] = revision
snapshot = Path(snapshot_download(**download_kwargs))
return (snapshot / subfolder).resolve(strict=True)
def _judge_stage(args: argparse.Namespace, image_path: Path, output_dir: Path) -> None:
from actor_to_editor import atomic_write_json
judge_path = _resolve_model_subfolder(
args.judge_model,
args.judge_subfolder,
args.revision,
args.local_files_only,
)
os.environ["VF_JUDGE_MODEL_PATH"] = str(judge_path)
os.environ["VF_JUDGE_MODEL_ID"] = "mr-iqa-2-e5-judge"
os.environ["VF_JUDGE_PROMPT_SCHEMA"] = "e5_training_reasoning_v5"
os.environ["VF_JUDGER_MAX_BATCH_SIZE"] = "1"
from judge.server import FrozenJudger
edited_path = (output_dir / "edited.png").resolve(strict=True)
judger = FrozenJudger(str(judge_path))
original = judger.score_image(str(image_path), repeats=1)
edited = judger.score_image(str(edited_path), repeats=1)
delta = _quality_delta(original, edited)
evaluation = {
"j0": original["mean"],
"j1": edited["mean"],
"j1_minus_j0": delta,
"original": original,
"edited": edited,
"judge_model": str(judge_path),
"revision": args.revision or None,
}
atomic_write_json(output_dir / "evaluation.json", evaluation)
result_path = output_dir / "result.json"
result = json.loads(result_path.read_text(encoding="utf-8"))
result["judge"] = evaluation
atomic_write_json(result_path, result)
def _stage_command(
*,
conda: str,
environment: str,
python_executable: str,
stage: str,
args: argparse.Namespace,
image_path: Path,
output_dir: Path,
) -> list[str]:
command = (
[python_executable]
if python_executable
else [conda, "run", "--no-capture-output", "-n", environment, "python"]
)
command.extend(
[
str(Path(__file__).resolve()),
str(image_path),
"--output-dir",
str(output_dir),
"--gpu",
"0",
"--max-pixels",
str(args.max_pixels),
"--actor-model",
args.actor_model,
"--actor-subfolder",
args.actor_subfolder,
"--judge-model",
args.judge_model,
"--judge-subfolder",
args.judge_subfolder,
"--editor-model",
args.editor_model,
"--stage",
stage,
]
)
if args.local_files_only:
command.append("--local-files-only")
if args.revision:
command.extend(["--revision", args.revision])
return command
def run_sequential(
args: argparse.Namespace,
*,
command_runner: Callable[..., Any] = subprocess.run,
) -> dict[str, Any]:
image_path = Path(args.image).expanduser().resolve(strict=True)
if not image_path.is_file():
raise FileNotFoundError(f"input image is not a file: {image_path}")
output_dir = Path(args.output_dir).expanduser().resolve()
output_dir.mkdir(parents=True, exist_ok=True)
conda = shutil.which("conda") or ""
if not conda and (not args.actor_python or not args.editor_python):
raise RuntimeError("conda is required to run the two validated environments")
child_environment = os.environ.copy()
child_environment["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
child_environment["PYTHONUNBUFFERED"] = "1"
for stage, environment, python_executable in (
("actor", args.actor_env, args.actor_python),
("editor", args.editor_env, args.editor_python),
("judge", args.actor_env, args.actor_python),
):
command_runner(
_stage_command(
conda=conda,
environment=environment,
python_executable=python_executable,
stage=stage,
args=args,
image_path=image_path,
output_dir=output_dir,
),
check=True,
env=child_environment,
)
return json.loads((output_dir / "result.json").read_text(encoding="utf-8"))
def main(argv: list[str] | None = None) -> int:
args = parse_args(argv)
image_path = Path(args.image).expanduser().resolve(strict=True)
output_dir = Path(args.output_dir).expanduser().resolve()
output_dir.mkdir(parents=True, exist_ok=True)
if args.stage == "actor":
_actor_stage(args, image_path, output_dir)
return 0
if args.stage == "editor":
_editor_stage(args, image_path, output_dir)
return 0
if args.stage == "judge":
_judge_stage(args, image_path, output_dir)
return 0
result = run_sequential(args)
print(json.dumps(result, ensure_ascii=False, indent=2))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|