examples/context_ir.py
3.8 KB · 92 lines · python Raw
1 #!/usr/bin/env python3
2 """Enhance a prompt with the MiniMax H3-Image-Context-IR API.
3
4 export MINIMAX_API_BASE=https://api.minimax.cn # CN; Global: https://api.minimax.io
5 export MINIMAX_API_KEY=<API key>
6 python3 context_ir.py t2i # instruction.t2i.txt
7 python3 context_ir.py edit # instruction.edit.txt with reference1.png
8
9 Sends the instruction (and, for edit, the reference images) to POST /v1/h3_image_context_ir
10 and writes the enhanced prompt to revised_prompt.txt. Unless --size is given, the API also
11 picks the aspect ratio; the matching size of about 16 megapixels is written to
12 revised_size.txt. Pass the files to t2i.py / edit.py with --prompt and --size. Uses only the
13 Python standard library.
14 """
15
16 import argparse
17 import base64
18 import json
19 import math
20 import mimetypes
21 import os
22 import urllib.request
23 from pathlib import Path
24
25 # absolute(), not resolve(): in the Hugging Face cache this file is a symlink into blobs/.
26 HERE = Path(__file__).absolute().parent
27 AREA = 4096 * 4096
28
29
30 def data_uri(path: Path) -> str:
31 content_type = mimetypes.guess_type(path.name)[0]
32 if content_type is None:
33 raise ValueError(f"unknown image type: {path}")
34 return f"data:{content_type};base64,{base64.b64encode(path.read_bytes()).decode()}"
35
36
37 def size_for(aspect_ratio: str) -> str:
38 """About 16 megapixels at the ratio, both sides multiples of 32 (matches the README Sizes table)."""
39 x, y = (int(v) for v in aspect_ratio.split(":"))
40 width = round(math.sqrt(AREA * x / y) / 32) * 32
41 height = round(math.sqrt(AREA * y / x) / 32) * 32
42 return f"{width}x{height}"
43
44
45 def post(path: str, body: dict) -> dict:
46 request = urllib.request.Request(
47 f"{os.environ['MINIMAX_API_BASE'].rstrip('/')}{path}",
48 data=json.dumps(body).encode(),
49 headers={"Authorization": f"Bearer {os.environ['MINIMAX_API_KEY']}", "Content-Type": "application/json"},
50 )
51 with urllib.request.urlopen(request, timeout=600) as response:
52 return json.load(response)
53
54
55 def request_body(task: str, instruction: Path, images: list, size) -> dict:
56 body = {"model": "MiniMax-H3-Image", "prompt": instruction.read_text().strip()}
57 if task == "edit":
58 body["images"] = [{"image_url": data_uri(path)} for path in images or [HERE / "reference1.png"]]
59 if size:
60 body["size"] = size
61 return body
62
63
64 def add_task_arguments(parser: argparse.ArgumentParser) -> None:
65 parser.add_argument("task", choices=["t2i", "edit"])
66 parser.add_argument("--instruction", type=Path, help="default: instruction.<task>.txt")
67 parser.add_argument("--image", action="append", type=Path,
68 help="edit only: reference image; repeat for <Picture 2>, ... (default: reference1.png)")
69 parser.add_argument("--size", help="WIDTHxHEIGHT of the image you will generate (default: chosen by the API)")
70
71
72 def main() -> None:
73 parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
74 add_task_arguments(parser)
75 parser.add_argument("--out", default="revised_prompt.txt")
76 parser.add_argument("--size-out", default="revised_size.txt", help="written when the API picks the aspect ratio")
77 args = parser.parse_args()
78
79 instruction = args.instruction or HERE / f"instruction.{args.task}.txt"
80 result = post("/v1/h3_image_context_ir", request_body(args.task, instruction, args.image, args.size))
81 Path(args.out).write_text(result["revised_prompt"])
82 summary = {"request_id": result.get("request_id"), "usage": result.get("usage")}
83 if "aspect_ratio" in result:
84 summary["aspect_ratio"] = result["aspect_ratio"]
85 summary["size"] = size_for(result["aspect_ratio"])
86 Path(args.size_out).write_text(summary["size"])
87 print(json.dumps(summary))
88
89
90 if __name__ == "__main__":
91 main()
92