le-gpt-image v1.3：完整引入 GPT-Image2-Skill Reference Gallery（163 条精调 prompt / 31 类 / craft + 官方提示词指南，MIT，附 NOTICE 与原版 LICENSE，仅删预览图外链）；SKILL.md 工作流升级为 Gallery 优先（查库选案例做底改写，勿从零现写）；对照实测同主题海报 Gallery 框架质感显著优于即兴 prompt

2026-10-06 22:49:19 +08:00
parent d995f725d0
commit 3b1ee9f997
46 changed files with 4150 additions and 8 deletions
@@ -0,0 +1,43 @@
+# GPT Image 2.5: migration and evaluation
+
+Read only for model migration or an authorized comparison, not routine image
+requests. Sources checked **2026-09-09**; locally organized official-source notes,
+not a new MIT Cookbook snapshot or locally verified image results.
+
+## What changes from Image 2
+
+- OpenAI positions [Flare](https://developers.openai.com/api/docs/models/gpt-image-2.5-flare)
+  around speed and [Sunburst](https://developers.openai.com/api/docs/models/gpt-image-2.5-sunburst)
+  around demanding quality. Both generate and edit; Sunburst is not edit-only.
+- If Image 2 already passes your quality bar, evaluate Flare first. Otherwise,
+  evaluate Sunburst. Compare the same inputs and explicit settings before
+  rewriting prompts; keep the `quality` setting fixed in the first comparison.
+  Establish acceptable output quality before optimizing latency.
+  [Official migration guidance](https://developers.openai.com/api/docs/guides/image-prompting?model=gpt-image-2.5#migrate-an-existing-workflow)
+- `xhigh` and `max` are options, not automatic upgrades. Higher quality may not
+  improve a particular task. Test against acceptance criteria and budget.
+  [Official quality guidance](https://developers.openai.com/api/docs/guides/image-prompting?model=gpt-image-2.5#model-parameters)
+
+Quoted text, layout constraints and reference invariants are shared techniques
+already covered in [craft.md](craft.md), not newly invented 2.5 prompt syntax.
+Keep parameter details in [models.md](models.md), not in task prompts.
+
+## Local evaluation checklist
+
+Only run comparisons or extra variants when authorized. Before a trial, write
+down what would make the output unusable; use the selected task slice's checks.
+Record model, full prompt, attached inputs, dimensions, quality, format, elapsed
+time, reported usage, output path and pass/fail reason. Keep failed attempts in
+the cost accounting. A documentation example is not our measured result.
+
+## Source boundaries
+
+- [Current official prompting guide](https://developers.openai.com/api/docs/guides/image-prompting?model=gpt-image-2.5):
+  illustrated 2.5 workflows. Its complete runnable example still selects Image 2;
+  do not copy that model or batch count over the user's confirmed choices.
+- [Historical local Cookbook](openai-cookbook.md): captured 2026-04-23, with its
+  existing source/license retained. The official notebook checked on 2026-09-09
+  [still targets Image 2](https://github.com/openai/openai-cookbook/blob/a78f3f37bd23637aac2b3f1e8b1251cf5bb9e1a7/examples/multimodal/image-gen-models-prompting-guide.ipynb).
+  Its recommendations and executable defaults are not current 2.5 settings.
+- [Community templates](templates-gpt-image-2.5.md): attributed adaptations with
+  author-reported source models, not OpenAI recommendations or verified outputs.