Macで画像生成AIを無料で動かす2026〜Stable DiffusionからFLUXまで、ツールとモデルの選び方〜

#!/usr/bin/env python3
"""diffusers で FLUX.2 [klein] 4B を動かし、読み込みと1枚あたりの時間を測る（Mac・MPS）。

  .venv/bin/python bench_diffusers.py --dtype bf16
"""
import argparse
import json
import os
import resource
import sys
import time

import torch
from diffusers import Flux2KleinPipeline

sys.stdout.reconfigure(line_buffering=True)
PROMPTS = [
    "a beautiful sunset over mountains, photorealistic, golden hour",
    "a close-up photo of hands holding a cup of coffee on a wooden table, natural light",
    'a street at night with a wooden shop sign that says "OPEN" and a neon sign that says "MAC 2026"',
    "日本の屋台、夜、赤い提灯に「ラーメン」と書かれている",
]


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
    ap.add_argument("--runs", type=int, default=3)
    ap.add_argument("--out", default="out")
    a = ap.parse_args()
    dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[a.dtype]

    t0 = time.perf_counter()
    pipe = Flux2KleinPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-4B", torch_dtype=dtype).to("mps")
    load = time.perf_counter() - t0
    print(f"load {load:.1f}s  torch {torch.__version__}")
    res = {"tool": "diffusers", "dtype": a.dtype, "load_s": round(load, 2), "runs": []}
    for i, p in enumerate(PROMPTS):
        for r in range(a.runs):
            g = torch.Generator("cpu").manual_seed(42)
            t = time.perf_counter()
            img = pipe(prompt=p, width=1024, height=1024, num_inference_steps=4, guidance_scale=1.0,
                       generator=g).images[0]
            torch.mps.synchronize()
            sec = time.perf_counter() - t
            res["runs"].append({"prompt": i, "run": r, "sec": round(sec, 2)})
            print(f"prompt={i} run={r} {sec:.2f}s")
            if r == 0:
                img.save(f"{a.out}/diffusers_{a.dtype}_p{i}.png")
    res["peak_rss_gb"] = round(resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1e9, 2)
    res["mps_driver_gb"] = round(torch.mps.driver_allocated_memory() / 1e9, 2)
    print(json.dumps(res))
    json.dump(res, open(f"{a.out}/bench_diffusers_{a.dtype}.json", "w"), indent=1)


if __name__ == "__main__":
    os.makedirs("out", exist_ok=True)
    main()

#!/bin/bash
# mflux で FLUX.2 [klein] 4B を動かす。1回の起動で同じ種を3回描き、1枚目＝冷えた状態・2〜3枚目＝温まった状態。
# 使い方: bash bench_mflux.sh [量子化ビット数（省略で量子化なし）]　※ .venv と同じフォルダに置いて実行
cd "$(dirname "$0")" || exit 1
mkdir -p out
Q=${1:-}
TAG=${Q:+q$Q}; TAG=${TAG:-bf16}
QARG=${Q:+--quantize $Q}
P=(
  "a beautiful sunset over mountains, photorealistic, golden hour"
  "a close-up photo of hands holding a cup of coffee on a wooden table, natural light"
  'a street at night with a wooden shop sign that says "OPEN" and a neon sign that says "MAC 2026"'
  "日本の屋台、夜、赤い提灯に「ラーメン」と書かれている"
)
for i in 0 1 2 3; do
  echo "== prompt $i ($TAG)"
  /usr/bin/time -l .venv/bin/mflux-generate-flux2 --model flux2-klein-4b $QARG \
    --prompt "${P[$i]}" --width 1024 --height 1024 --steps 4 --guidance 1.0 \
    --seed 42 42 42 --output "out/mflux_${TAG}_p${i}_{seed}.png" 2>&1 \
    | tr '\r' '\n' | grep -E "100%|real|maximum resident|peak memory|Saved|Error|error" | tail -12
done

#!/usr/bin/env python3
"""モデルの世代で比べる（diffusers・Mac MPS）。同じ4つの文章を、SD 1.5 と SDXL で描く。
FLUX.2 [klein] 4B の結果は bench_diffusers.py のものを使う。

  .venv/bin/python bench_generations.py sd15
  .venv/bin/python bench_generations.py sdxl
  .venv/bin/python bench_generations.py sd15 --dtype fp16   # 実験4：6月の「float32が安定」は今も必要か
"""
import argparse
import json
import os
import sys
import time

import numpy as np
import torch
from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline

sys.stdout.reconfigure(line_buffering=True)
PROMPTS = [
    "a beautiful sunset over mountains, photorealistic, golden hour",
    "a close-up photo of hands holding a cup of coffee on a wooden table, natural light",
    'a street at night with a wooden shop sign that says "OPEN" and a neon sign that says "MAC 2026"',
    "日本の屋台、夜、赤い提灯に「ラーメン」と書かれている",
]
MODELS = {
    # 各モデルの標準的な使い方（公式の説明に合わせる）
    "sd15": dict(repo="stable-diffusion-v1-5/stable-diffusion-v1-5", cls=StableDiffusionPipeline, size=512, steps=20, cfg=7.5),
    "sdxl": dict(repo="stabilityai/stable-diffusion-xl-base-1.0", cls=StableDiffusionXLPipeline, size=1024, steps=30, cfg=7.0),
}


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("model", choices=MODELS)
    ap.add_argument("--dtype", choices=["fp32", "fp16"], default=None)
    ap.add_argument("--runs", type=int, default=2)
    a = ap.parse_args()
    m = MODELS[a.model]
    dtype_name = a.dtype or ("fp32" if a.model == "sd15" else "fp16")
    dtype = torch.float32 if dtype_name == "fp32" else torch.float16
    kw = dict(torch_dtype=dtype)
    if a.model == "sd15":
        kw["safety_checker"] = None
    if dtype == torch.float16 and a.model == "sdxl":
        kw.update(variant="fp16", use_safetensors=True)
    t0 = time.perf_counter()
    pipe = m["cls"].from_pretrained(m["repo"], **kw).to("mps")
    load = time.perf_counter() - t0
    print(f"{a.model} {dtype_name} load {load:.1f}s")
    res = {"model": a.model, "dtype": dtype_name, "load_s": round(load, 2), "size": m["size"], "steps": m["steps"], "runs": []}
    for i, p in enumerate(PROMPTS):
        for r in range(a.runs):
            g = torch.Generator("cpu").manual_seed(42)
            t = time.perf_counter()
            img = pipe(prompt=p, width=m["size"], height=m["size"], num_inference_steps=m["steps"],
                       guidance_scale=m["cfg"], generator=g).images[0]
            torch.mps.synchronize()
            sec = time.perf_counter() - t
            arr = np.asarray(img).astype(np.float32)
            black = float(arr.mean()) < 2.0          # 真っ黒（NaNで潰れた）かどうか
            res["runs"].append({"prompt": i, "run": r, "sec": round(sec, 2), "mean": round(float(arr.mean()), 1), "black": black})
            print(f"prompt={i} run={r} {sec:.2f}s mean={arr.mean():.1f}{' ★真っ黒' if black else ''}")
            if r == 0:
                img.save(f"out/gen_{a.model}_{dtype_name}_p{i}.png")
    json.dump(res, open(f"out/bench_gen_{a.model}_{dtype_name}.json", "w"), indent=1)


if __name__ == "__main__":
    os.makedirs("out", exist_ok=True)
    main()

#!/usr/bin/env python3
"""ComfyUI に「文章から画像を1枚」の手順（ワークフロー）を送り、描き終わるまでの時間を測る。標準ライブラリだけ。

  python3 comfy_run.py --port 8189 --ckpt v1-5-pruned-emaonly.safetensors --out exp1_comfy.png

手順はComfyUIの初期画面と同じ7つの部品：チェックポイント読込 → 文章×2 → 空の画像 → サンプラー → 画像に戻す → 保存。
"""
import argparse
import json
import sys
import time
import urllib.parse
import urllib.request

sys.stdout.reconfigure(encoding="utf-8", line_buffering=True)


def workflow(ckpt, prompt, negative, steps, cfg, seed, w, h, sampler="euler", scheduler="normal"):
    return {
        "1": {"class_type": "CheckpointLoaderSimple", "inputs": {"ckpt_name": ckpt}},
        "2": {"class_type": "CLIPTextEncode", "inputs": {"text": prompt, "clip": ["1", 1]}},
        "3": {"class_type": "CLIPTextEncode", "inputs": {"text": negative, "clip": ["1", 1]}},
        "4": {"class_type": "EmptyLatentImage", "inputs": {"width": w, "height": h, "batch_size": 1}},
        "5": {"class_type": "KSampler", "inputs": {
            "model": ["1", 0], "positive": ["2", 0], "negative": ["3", 0], "latent_image": ["4", 0],
            "seed": seed, "steps": steps, "cfg": cfg, "sampler_name": sampler, "scheduler": scheduler, "denoise": 1.0}},
        "6": {"class_type": "VAEDecode", "inputs": {"samples": ["5", 0], "vae": ["1", 2]}},
        "7": {"class_type": "SaveImage", "inputs": {"images": ["6", 0], "filename_prefix": "p1000"}},
    }


def call(url, body=None):
    req = urllib.request.Request(url, data=json.dumps(body).encode() if body else None,
                                 headers={"Content-Type": "application/json"})
    return json.loads(urllib.request.urlopen(req, timeout=600).read())


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--port", type=int, default=8188)
    ap.add_argument("--ckpt", required=True)
    ap.add_argument("--prompt", default="a beautiful sunset over mountains, photorealistic, golden hour, 4k")
    ap.add_argument("--negative", default="")
    ap.add_argument("--steps", type=int, default=20)
    ap.add_argument("--cfg", type=float, default=7.5)
    ap.add_argument("--seed", type=int, default=42)
    ap.add_argument("--size", default="512x512")
    ap.add_argument("--out", required=True)
    a = ap.parse_args()
    w, h = map(int, a.size.split("x"))
    base = f"http://127.0.0.1:{a.port}"
    t0 = time.perf_counter()
    pid = call(base + "/prompt", {"prompt": workflow(a.ckpt, a.prompt, a.negative, a.steps, a.cfg, a.seed, w, h)})["prompt_id"]
    while True:
        hist = call(f"{base}/history/{pid}")
        if pid in hist:
            break
        time.sleep(0.5)
    sec = time.perf_counter() - t0
    st = hist[pid].get("status", {})
    outs = [img for n in hist[pid]["outputs"].values() for img in n.get("images", [])]
    if not outs:
        print("失敗:", json.dumps(st, ensure_ascii=False)[:800])
        sys.exit(1)
    img = outs[0]
    q = urllib.parse.urlencode({"filename": img["filename"], "subfolder": img["subfolder"], "type": img["type"]})
    open(a.out, "wb").write(urllib.request.urlopen(f"{base}/view?{q}").read())
    print(f"OK {sec:.1f}秒 → {a.out}  ({st.get('status_str')})")


if __name__ == "__main__":
    main()

//@api-1.0
// p1000 Draw Things の計測：同じ文章・同じ種・同じ大きさで3回ずつ生成し、時間を記録する
const cfg = pipeline.configuration;
const prompts = [
  "a beautiful sunset over mountains, photorealistic, golden hour",
  "a close-up photo of hands holding a cup of coffee on a wooden table, natural light",
  "a street at night with a wooden shop sign that says \"OPEN\" and a neon sign that says \"MAC 2026\"",
  "日本の屋台、夜、赤い提灯に「ラーメン」と書かれている",
];
cfg.width = 1024;
cfg.height = 1024;
cfg.steps = 4;
cfg.guidanceScale = 1;
cfg.batchSize = 1;
cfg.batchCount = 1;
cfg.hiresFix = false;
cfg.seed = 42;
console.log("CONFIG " + JSON.stringify({model: cfg.model, sampler: cfg.sampler, steps: cfg.steps,
  guidanceScale: cfg.guidanceScale, width: cfg.width, height: cfg.height, shift: cfg.shift, seedMode: cfg.seedMode}));
console.log("PICTURES " + filesystem.pictures.path);
const times = [];
for (let i = 0; i < prompts.length; i++) {
  for (let r = 0; r < 3; r++) {
    canvas.clear();
    cfg.seed = 42;
    const t0 = Date.now();
    pipeline.run({configuration: cfg, prompt: prompts[i], negativePrompt: ""});
    const sec = (Date.now() - t0) / 1000;
    times.push(sec);
    console.log("RESULT prompt=" + i + " run=" + r + " sec=" + sec.toFixed(2));
    if (r === 0) {
      canvas.saveImage(filesystem.pictures.path + "/p1000_dt_p" + i + ".png", true);
    }
  }
}
console.log("ALL " + JSON.stringify(times));