Selective Color Portrait with Heavy Brand Text Template

date: '2026-05-01T14:05:58+08:00'
result: pass
prompt:
  text: >-
    Ultra-realistic cinematic portrait of a teenage boy with soft facial features, messy
    medium-length hair, and round black glasses, wearing headphones around his neck and a dark blue
    Nirvana t-shirt with a yellow smiley logo. He is holding two small square photo frames in front
    of his face—one highlighting his eye and the other his lips—in full color, while the rest of the
    image is mostly black and white. Background is a grunge-style wall covered with punk rock
    posters, graffiti, and band logos (punk aesthetic, chaotic layout, high detail). Visible posters
    include iconic rock band styles, spray-painted text like "PUNK NOT DEAD," "RIOT," and raw street
    art textures. Dramatic soft lighting, high contrast, shallow depth of field, sharp focus on face
    and frames, editorial photography style, 4K, hyper-detailed, moody tone.
refs: []
provider:
  id: gpt_image_2
  relay: apimart
  config:
    aspect_ratio: '1:1'
    size: '1:1'
    'n': 1
output:
  path: ./punk_selective_color_v1.png
  bytes: 2203515
  wall_seconds: 36.1
  task_id: task_01KQH28JZ1CP8ZBC3TSX42KBZE
script: experiments/punk_selective_color_test/test_v1.py
cost_yuan: 0.5
notes: >
  7 unique 维度全 PASSED · 36.1s wall · 出片 magazine-cover quality.


  **4 个 framework-级新发现:**


  (a) **Selective color WORKS on gpt-image-2** (我们 0 测过)
      · 主要在 frame 内 (eye + lips) · 但 model 自我延展到 brand items (Nirvana shirt blue + 黄 smiley · RIOT 红字)
      · 解 "rest is mostly black and white" 不是 "exclusively" · 智能 hierarchy
      · 启示: 局部色彩 + 大面 desaturate 是新可控 primitive

  (b) **In-image text 上限再上调 · 6 distinct strings 全 readable**
      · "PUNK NOT DEAD" (12) · "RIOT" (4) · "RAMONES" (7) · "SEX PISTOLS" (11) · "PUNK PHIL" (9) · "NIRVANA" (7)
      · 之前 ladder: Noor 5 (2026-04) → sushi monster 17 (2026-05 上午) → 此条 6 strings × 4-12 chars
      · production 含义: 拥挤背景多 brand text 完全可控 · 适合 punk/streetwear/磁带店/concert 场景

  (c) **Music brand IPs 高准确度** (Nirvana logo + smiley · Ramones · Sex Pistols)
      · 比 Ninja Hattori 的"Shinzo 漂为 Kenichi"准很多 · music brand 在 model 训练中更密集
      · IP 准确度排序假设: music brand > 顶级 anime IP > 副线 anime IP > 小众 IP
      · 启示: brand-heavy 场景前 grep IP 知名度 · 知名 = 准

  (d) **"Photo frames as compositional mask"** · advanced visual concept 完美执行
      · 模型理解 frame 不只是 prop · 而是 viewport / 局部色彩窗口
      · 类似的 "object-as-frame" 范式可拓展: 镜子里的真实色 / 窗外的不同时间 / book pages 内的另一世界

  其他 PASSED:
    · Editorial portrait (high contrast · shallow DoF · 编辑摄影感)
    · Layered grunge wall (multiple overlap posters + 不同 paper texture · graffiti)
    · Multi-prop (headphones + glasses + 2 frames + tee + watch + bracelet)
    · 35-40s wall · cost ¥0.5 · production-friendly

  Production use:
    · streetwear / punk / band marketing
    · 唱片店 / livehouse 内容 mock
    · "music tribute" 系列 (替换乐队 + 海报组)
    · selective-color 风格 photoshoot mock

  GOTCHA: Selective color 边界 model 自决 (不只 frame 内 · 也涉 brand items) · 严格控制需 prompt 加 "color ONLY in
  frames"