PROMPT RECORD · 公开作品详情
name: flowers-upscale（花卉上采样）
公开 AI 图像创作记录，包含原图、提示词、使用模型、作者、标签与来源。
- 使用模型
- GPT-Image-2
- 创作者
- @igortr_
- 发布时间
- Mon Oct 12 12:49:46 +0000 2026
- 图片数量
- 3 张
完整提示词
源提示词按原始换行与标点完整保留--- name: flowers-upscale description: Photo to 8K; 2/3-pass contextual AI crop enhancement and registered blending. --- INPUT S=uploaded photo; missing→request. Unspecified→ask "2 passes (recommended) or 3? 8K default; another resolution?" Await choice. N=2|3; L=specified longest edge else8192; preserve aspect. Discover available AI image-editing+raster-processing capabilities; missing→report requirement. Use environment-native tools/paths; create helpers as needed. Preserve S; separate outputs. Execute end-to-end after choice. E(C): inspect; AI-edit inputs=[S=context,C=target], maximum practical supported resolution; retain actual output dimensions. Prompt=Q+only relevant material constraints. Concurrent independent calls where supported; persist outputs+source-coordinate boxes+prompts immediately. P1: S→3×3 cores; pad each side20% core dimension, clip to S→9 overlapping C→E→register/blend→M1; resize assembled M1 to L. P2: M1→4–8 padded salient ROIs(face/hair/hands/objects/textiles as present)→E→register/blend→M2. Exclude low-information regions. N=3: M2→4–8 tighter focal ROIs(individual eyes/lips/nose/fingers/material details as present)→E→register/blend→M3. Keep homologous features consistent. Every E receives S, never crop alone. Q="1=original: identity/anatomy/color/light authority. 2=target: return exact crop/framing/geometry. Reconstruct plausible photographic detail; preserve expression, pose, contours, clothing, genuine imperfections, focus falloff. Optical clarity; natural edge transitions; spatially varied texture. Repair inherited artifacts. Skin: irregular pores/fine creases, quiet intervals. Hair: cohesive groups, selective fibers, natural sparse fringe. Iris: source color, irregular branching fibers, coherent pupil, restrained reflections. Textiles: original weave direction/spacing/motifs. Avoid invented structures/marks, beautification, relighting, halos, ringing, embossed/crosshatched/repeated texture, synthetic grain, blanket sharpening." ASSEMBLE: map recorded boxes→master; global similarity registration first; bounded smooth local correction only with reliable correspondence, never distort rigid geometry/repeating patterns. Unreliable→retain prior pixels. Match broad exposure/color without transferring old fine texture. Normalized overlap/multiband blend; feather15–25% border→alpha0 at interior patch boundaries. Preserve native enhanced crops. Before next pass: inspect full image+100% focal/overlap crops; repair doubling/seams/drift/texture mismatch. ≤2 targeted retries/defect; rejected/failed→retain prior content+disclose. EXPORT: M_N→lossless PNG at L; verify dimensions+decode; deliver file+preview+prompts/settings. Describe detail as reconstructed; distinguish exported resolution from native generated resolution. Also, an important point: run it in Codex/Work rather than the standard chat, and bump the effort level up to at least "high." On the downside, it takes *forever* to think. The first image took me 10 minutes, and each subsequent one took even longer than the last 😐 You can also use any other image-generating agent instead of GPT - like Gemini + Nanobanana - though the result depends heavily on the agent's "brains." I got the best quality using the Opus 5.5 + Codex CLI image gen combo (a photo of a young Elon), but that was also the longest run by far 😁
提示词译文
译文与源提示词分开显示，不替代原文name: flowers-upscale（花卉上采样） description: 将照片提升至 8K；通过 2/3 轮上下文式 AI 增强与配准混合。 --- INPUT（输入）S=上传的照片；缺失→请求补发。未指定→询问"2 轮（推荐）还是 3 轮？默认 8K；需要其他分辨率吗？" 等待选择。N=2|3；L=指定的最长边，否则为 8192；保持原始比例。探明可用的 AI 图像编辑 + 光栅处理能力；不可用→报告需求。使用环境原生工具与路径；按需创建辅助程序。保留 S；输出与原图分离。选择完成后端到端执行。 E(C)：检查；AI 编辑输入=[S=context,C=target]，采用实际支持的最大分辨率；保留实际输出尺寸。Prompt=Q+仅与素材相关的约束。在支持的条件下并发独立调用；立即持久化输出 + 源坐标框 + prompt。 P1：S→3×3 核心；每侧填充核心尺寸的 20%，裁剪回 S→9 个重叠的 C→E→配准/混合→M1；将组装后的 M1 缩放到 L。 P2：M1→4–8 个带填充的显著 ROI（面部/头发/手部/物体/织物，按存在情况）→E→配准/混合→M2。排除低信息区域。 N=3：M2→4–8 个更紧凑的焦距 ROI（单眼/嘴唇/鼻子/手指/材质细节，按存在情况）→E→配准/混合→M3。保持同源特征一致。每次 E 都接收 S，绝不单独处理裁剪。 Q="1=原图：身份/解剖/色彩/光线的权威来源。2=目标：返回精确的裁剪/取景/几何。重构可信的摄影细节；保留表情、姿态、轮廓、服装、真实瑕疵、焦点衰减。光学清晰；自然边缘过渡；空间多变的纹理。修复继承性伪影。皮肤：不规则毛孔/细微皱纹，留白均匀。头发：成束的毛发，挑选性单根纤维，自然稀疏的边缘。虹膜：原图色彩，不规则放射状纤维，瞳孔一致，反光克制。织物：保留原始编织方向/间距/图案。避免凭空发明的结构/痕迹、美化、重新打光、光晕、振铃伪影、浮雕/交叉影线/重复纹理、人造颗粒、整体锐化。" ASSEMBLE（组装）：将记录的框映射→主图；先做全局相似性配准；仅在存在可靠对应时进行有界平滑的局部校正，绝不扭曲刚性几何或重复图案。不可靠→保留先前像素。仅匹配大范围曝光/色彩，不迁移旧有细微纹理。归一化重叠/多频段混合；在内部补丁边界做 15–25% 边缘羽化→alpha 归零。保留原生增强裁剪。下一轮之前：检查全图 + 100% 焦距/重叠裁剪；修复重影/接缝/漂移/纹理不匹配。每个缺陷≤2 次定向重试；被拒绝/失败→保留先前内容并披露。 EXPORT（导出）：M_N→以 L 尺寸输出无损 PNG；验证尺寸+解码；交付文件+预览+prompt/设置。将细节描述为重构所得；区分导出分辨率与原生生成分辨率。 另外一点重要说明：在 Codex/Work 中运行，而非标准聊天，并将 effort 级别至少提升至 "high"。 缺点是它思考起来*非常*慢。第一张图片花了 10 分钟，之后每张都比上一张更慢 😐 你也可以用其他任何图像生成代理代替 GPT——比如 Gemini + Nanobanana——但结果在很大程度上取决于该代理的"智商"。 我用 Opus 5.5 + Codex CLI 图像生成组合获得了最佳质量（一张年轻的 Elon 照片），但那也是耗时最长的一次 😁