diff --git a/apps/admin-web/src/pages/AdminEditorAssetQueryPage.test.tsx b/apps/admin-web/src/pages/AdminEditorAssetQueryPage.test.tsx
index c0cca000c..ddbc7a502 100644
--- a/apps/admin-web/src/pages/AdminEditorAssetQueryPage.test.tsx
+++ b/apps/admin-web/src/pages/AdminEditorAssetQueryPage.test.tsx
@@ -245,6 +245,40 @@ test('后台素材查询不展示分类筛选和分类列', async () => {
   expect(screen.queryByRole('columnheader', { name: '分类' })).toBeNull();
 });
 
+test('后台素材查询将中间产物标记为不单独计费', async () => {
+  vi.mocked(listAdminEditorAssets).mockResolvedValueOnce({
+    entries: [
+      {
+        ...generatedAsset,
+        assetId: 'asset-source',
+        label: '角色形象原图',
+        assetKind: 'editor_green_screen_source',
+        generationCostMudPoints: 0,
+      },
+      generatedAsset,
+    ],
+    nextCursor: null,
+  });
+
+  render(
+    <AdminEditorAssetQueryPage token="admin-token" onUnauthorized={vi.fn()} />,
+  );
+
+  const sourceRow = (await screen.findByText('角色形象原图')).closest('tr');
+  const finalRow = screen.getByText('角色形象 1').closest('tr');
+  expect(sourceRow).not.toBeNull();
+  expect(finalRow).not.toBeNull();
+  expect(within(sourceRow!).getByText('不单独计费')).toBeTruthy();
+  expect(within(finalRow!).getByText('12 泥点')).toBeTruthy();
+
+  fireEvent.click(within(sourceRow!).getByRole('button', { name: '详情' }));
+  expect(
+    within(await screen.findByRole('dialog', { name: '素材详情' })).getByText(
+      '不单独计费',
+    ),
+  ).toBeTruthy();
+});
+
 test('后台素材查询缩略图使用 objectKey 换签后展示', async () => {
   render(
     <AdminEditorAssetQueryPage token="admin-token" onUnauthorized={vi.fn()} />,
diff --git a/apps/admin-web/src/pages/AdminEditorAssetQueryPage.tsx b/apps/admin-web/src/pages/AdminEditorAssetQueryPage.tsx
index 5bb5d25fa..d8276fe8f 100644
--- a/apps/admin-web/src/pages/AdminEditorAssetQueryPage.tsx
+++ b/apps/admin-web/src/pages/AdminEditorAssetQueryPage.tsx
@@ -24,6 +24,10 @@ const ADMIN_ASSET_READ_EXPIRE_SECONDS = 300;
 const ADMIN_ASSET_READ_DISPATCH_SPACING_MS = 40;
 const ADMIN_ASSET_READ_RETRY_DELAYS_MS = [400, 1_200, 3_000] as const;
 const ADMIN_ASSET_THUMBNAIL_ROOT_MARGIN = '240px 0px';
+const NON_BILLED_INTERMEDIATE_ASSET_KINDS = new Set([
+  'editor_green_screen_source',
+  'editor_image_edit_source',
+]);
 const AUDIO_ASSET_COVER_SRC = `${import.meta.env.DEV ? import.meta.env.BASE_URL : '/'}creation-home/audio-asset-cover.png`;
 let adminAssetReadDispatchTail = Promise.resolve();
 
@@ -217,7 +221,7 @@ export function AdminEditorAssetQueryPage({
                         {promptText}
                       </button>
                     </td>
-                    <td>{entry.generationCostMudPoints} 泥点</td>
+                    <td>{generationCostLabel(entry)}</td>
                     <td>
                       <button
                         className="admin-secondary-button"
@@ -516,7 +520,7 @@ function AdminAssetDetailDialog({
               {entry.width} x {entry.height}
             </AdminInfoItem>
             <AdminInfoItem label="成本">
-              {entry.generationCostMudPoints} 泥点
+              {generationCostLabel(entry)}
             </AdminInfoItem>
             <AdminInfoItem label="模型">{entry.model || '-'}</AdminInfoItem>
             <AdminInfoItem label="Provider">
@@ -558,6 +562,13 @@ function AdminAssetDetailDialog({
   );
 }
 
+function generationCostLabel(entry: AdminEditorAssetPayload) {
+  return entry.assetKind &&
+    NON_BILLED_INTERMEDIATE_ASSET_KINDS.has(entry.assetKind)
+    ? '不单独计费'
+    : `${entry.generationCostMudPoints} 泥点`;
+}
+
 function AdminAssetPreviewDialog({
   entry,
   token,
diff --git a/docs/project-memory/shared-memory/decision-log.md b/docs/project-memory/shared-memory/decision-log.md
index 8ee050490..4df64e06f 100644
--- a/docs/project-memory/shared-memory/decision-log.md
+++ b/docs/project-memory/shared-memory/decision-log.md
@@ -28,15 +28,15 @@
 
 - 背景：图片画布的普通图片、规范、角色、图标图集、UI 设计、宣发素材、视频和音频默认使用“类型 + 数字”命名，用户只能在生成后单独重命名素材，画布图层、项目资源和素材库名称容易不一致。
 - 决策：主生成状态继续使用可选 `assetLabel`，名称最多 80 个字符并在提交时去除首尾空格；当前生成面板不展示“资源名称”标签和输入框，默认沿用现有自动编号名称，历史状态或内部调用若携带非空名称，仍必须让同一个名称贯穿 `assetLabel`、`canvasCompletion.title`、本地结果图层标题、项目资源和账号素材库，不允许各链路自行生成不同名称。移除名称输入后，角色、图标图集、UI 设计和角色动作等提示词输入恢复统一可见边框。
-- 派生产物：图标和角色动作后端契约补齐 `assetLabel`。带背景原图、图片修改 provider 原始输出等中间产物基于主名称追加“（原图）”或“（原始输出）”；图标切片继续按用户填写的图标描述命名，不继承图集名称覆盖独立素材语义。
+- 派生产物：图标和角色动作后端契约补齐 `assetLabel`。带背景原图、角色动作绿幕预览等具有独立复用价值的中间产物基于主名称追加“（原图）”等后缀；普通图片和图片修改的纯尺寸变换在内存完成后只上传一次，不生成“原始输出”副本。图标切片继续按用户填写的图标描述命名，不继承图集名称覆盖独立素材语义。
 - 影响范围：图片画布生成状态与面板、提交模型、图标和角色动作请求契约、项目资源 / 素材库持久化和相关编辑器文档。
 - 验证方式：覆盖生成面板不渲染资源名称输入、提示词边框、空白回退、内部自定义名与长度限制，以及图片 / 图标 / 视频 / 音频 / 角色动作的请求名称、完成快照标题和素材名称一致性；运行前端定向测试、Rust 契约与 API 定向测试、`npm run typecheck`、`npm run check:encoding`、`git diff --check`。
 
 ## 2026-07-13 图片画布多产物生成任务必须保存全部可恢复产物
 
 - 背景：角色形象、图标 spritesheet 和 UI 素材提取会先得到带纯色背景的原图，再执行抠图或拆分；图片修改会先得到模型对齐尺寸的原始输出，角色动作会先得到绿幕预览视频，再抽帧和抠图。此前部分原始产物只登记到 OSS，或者要等后处理成功后才进入项目资源，用户无法在失败后找回已经生成成功的内容。
-- 决策：凡一次资产生成任务产生多个可恢复产物，后端必须先把上游已返回的中间产物写入 OSS、`asset_object`、项目资源和账号素材库，再执行抠图、尺寸恢复、抽帧或拆分；未指定素材文件夹时进入默认“项目”文件夹。角色形象、图标 spritesheet 和 UI 素材提取同时保留纯色背景原图与透明后处理结果；图片修改保留模型原始输出与尺寸恢复结果；角色动作把绿幕预览视频作为一个可复用素材保存，逐帧源图继续留在同一任务 OSS 路径，不把 32 至 48 帧逐张灌入素材库。普通图片、去背景、音频等没有独立上游中间产物的任务不制造重复副本。
-- 画布与降级：有项目上下文的图片多产物继续由同一次 `canvasCompletion` 写入权威画布快照，生成器 `generatedLayerId` 锚定主后处理结果。图标和 UI 图集自动拆分是 best-effort；识别或切片持久化失败仍完成整张透明图集，并在 inline、队列轮询和刷新后任务列表中提示非阻断 warning，不得借用失败错误字段。
+- 决策：凡一次资产生成任务产生多个具有独立复用价值的产物，后端必须把上游已返回的中间产物写入 OSS、`asset_object`、项目资源和账号素材库，再执行抠图、抽帧或拆分；未指定素材文件夹时进入默认“项目”文件夹。角色形象、图标 spritesheet 和 UI 素材提取同时保留纯色背景原图与透明后处理结果。普通图片和图片修改的纯尺寸变换不属于独立产物：provider 回图保留在内存，变换成功只上传变换结果，变换失败只上传 provider 原图，整个流程只写一次 OSS 并只创建一个素材，不能制造重复“原始输出”。`nanobanana2` 使用标量清晰度档位和独立比例，保留 provider 输出尺寸，不按 `WIDTHxHEIGHT` 解析。角色动作把绿幕预览视频作为一个可复用素材保存，逐帧源图继续留在同一任务 OSS 路径，不把 32 至 48 帧逐张灌入素材库。去背景、音频等没有独立上游中间产物的任务不制造重复副本。
+- 画布与降级：有项目上下文的图片多产物继续由同一次 `canvasCompletion` 写入权威画布快照，生成器 `generatedLayerId` 锚定主后处理结果。图标和 UI 图集自动拆分是 best-effort；识别或切片持久化失败仍完成整张透明图集，并在 inline、队列轮询和刷新后任务列表中提示非阻断 warning，不得借用失败错误字段。中间产物的 `generation_cost_mud_points = 0` 只表示不重复计费，后台必须显示“不单独计费”；任务成本仍只记在最终素材。扣费确认边界保持为 provider 成功，OSS、尺寸恢复和画布回填不延长退款保护。
 - 影响范围：`server-rs/crates/api-server/src/editor_project.rs`、`character_animation_assets.rs`、外部生成任务摘要、图片画布完成快照、账号素材库和前端生成提示。
 - 验证方式：覆盖中间产物登记先于后处理、默认素材文件夹、图集拆分降级、inline / queue warning 和主结果锚定的定向测试，并运行 `cargo check -p api-server --manifest-path server-rs/Cargo.toml`、`npm run check:spacetime-schema`、前端定向测试、`npm run check:encoding`、`git diff --check`。
 - 关联文档：`docs/【编辑器】生成类面板Lovart统一改造方案-2026-06-17.md`、`docs/technical/【前端架构】图片画布编辑器MVP接入方案-2026-06-11.md`。
diff --git a/docs/project-memory/shared-memory/pitfalls.md b/docs/project-memory/shared-memory/pitfalls.md
index 54b8b0b37..973477242 100644
--- a/docs/project-memory/shared-memory/pitfalls.md
+++ b/docs/project-memory/shared-memory/pitfalls.md
@@ -384,6 +384,14 @@
 - 验证：`npm run test -- src/components/image-editor/ImageCanvasGenerationDialogModel.test.ts src/components/image-editor/ImageCanvasQuickEditPanelView.test.tsx src/components/image-editor/useImageCanvasGenerationSubmissionWorkflow.test.tsx` 覆盖来源参数继承、模型参数切换、目标尺寸提交和图层回填；`cargo test -p api-server editor_image_edit --manifest-path server-rs/Cargo.toml` 覆盖图标类拒绝、provider 尺寸对齐和回图恢复。
 - 关联：`src/components/image-editor/ImageCanvasGenerationSubmissionModel.ts`、`src/components/image-editor/useImageCanvasGenerationSubmissionWorkflow.ts`、`src/components/image-editor/ImageCanvasGenerationLayerModel.ts`、`server-rs/crates/api-server/src/editor_project.rs`。
 
+## 图片画布纯尺寸变换必须先在内存决策再单次上传
+
+- 现象：`nanobanana2` 已返回并成功解码图片，任务随后报“尺寸无效”；普通生图若把尺寸恢复作为硬失败，provider 已成功回图后仍可能没有素材进入素材库和画布。
+- 原因：通用尺寸恢复把 `nanobanana2` 的标量清晰度档位 `512 / 1024 / 2K` 当成 `WIDTHxHEIGHT` 解析；普通图片与快速编辑还把 OSS 持久化放在尺寸恢复之后。同步 `generateContent` 返回的是内联 base64，本地兜底 task id 不能用于向 provider 回查原图。
+- 处理：`nanobanana2` 保留 provider 输出几何尺寸；其它模型仍按显式像素目标尝试恢复。普通生图和快速编辑先把 provider 回图留在内存，尺寸恢复成功后只上传变换结果，恢复失败则降级为只上传 provider 原图；每个主结果只执行一次 OSS 持久化并只创建一个素材。角色、图标图集和 UI 提取的带背景原图仍按多产物语义单独保留。扣费确认仍以 provider 成功为界，不延长到 OSS、后处理或画布回填；后台把历史中间产物的零成本显示为“不单独计费”。
+- 验证：`cargo test -p api-server editor_project --manifest-path server-rs/Cargo.toml` 覆盖 nanobanana 标量尺寸、变换失败回落 provider 原图、变换先于单次持久化；`npm run test -- apps/admin-web/src/pages/AdminEditorAssetQueryPage.test.tsx` 覆盖中间产物成本文案。
+- 关联：`server-rs/crates/api-server/src/editor_project.rs`、`apps/admin-web/src/pages/AdminEditorAssetQueryPage.tsx`。
+
 ## 图片画布快速编辑元数据必须记录原图引用
 
 - 现象：快速编辑生成的新图可以替换画布，但打开图片信息时“生成输入”里看不到被修改的原图。
diff --git a/docs/technical/【前端架构】图片画布编辑器MVP接入方案-2026-06-11.md b/docs/technical/【前端架构】图片画布编辑器MVP接入方案-2026-06-11.md
index 21b7f9efc..67a4589fc 100644
--- a/docs/technical/【前端架构】图片画布编辑器MVP接入方案-2026-06-11.md
+++ b/docs/technical/【前端架构】图片画布编辑器MVP接入方案-2026-06-11.md
@@ -85,11 +85,11 @@
 - `POST /api/editor/assets`：批量或单个创建账号级素材，登录态上传必须写入 OSS / asset object 引用和 `/<objectKey>` 轻量路径，不允许把 Data URL / signed URL 写入素材库。
 - `PATCH /api/editor/assets/{assetId}`：重命名素材或移动素材到文件夹。
 - `DELETE /api/editor/assets/{assetId}`：删除素材。已放入画布的 project resource 不被级联删除，避免旧画布丢图。
-- `POST /api/editor/images/generations`：按提示词调用 VectorEngine 生成图片；角色生成可携带 `model`、`screenColor`、`segModel`、`aspectRatio`、`imageSize` 和 `referenceImageSrcs`，生成成功后 api-server 先保存带纯色背景源图，再调用 BgFilter 并传入 `screen_color=<screenColor>`、`seg_model=<segModel>` 生成透明 PNG。宣发素材携带 `kind: "publication-material"` 时固定归一为 `gpt-image-2`，不支持 `nanobanana2`。`nanobanana2` 参考图作为 `inline_data` 进入 `generateContent`，`gpt-image-2` 参考图进入 edits。普通重绘继续走该接口并把当前图层图片作为参考图；图片快速编辑不走该接口。请求可携带 `projectId`、`assetFolderId`、`assetKind`、`generationInputs` 和 `sourceResourceId`，后端生成成功后创建 project resource / 账号素材并在响应中返回 resource / asset 快照。
+- `POST /api/editor/images/generations`：按提示词调用 VectorEngine 生成图片；普通图片的 provider 回图先留在内存，尺寸变换成功后只上传变换结果，变换失败则只上传 provider 原图，主结果只写一次 OSS 且不额外创建“原始输出”。角色生成可携带 `model`、`screenColor`、`segModel`、`aspectRatio`、`imageSize` 和 `referenceImageSrcs`，生成成功后 api-server 先保存带纯色背景源图，再调用 BgFilter 并传入 `screen_color=<screenColor>`、`seg_model=<segModel>` 生成透明 PNG。宣发素材携带 `kind: "publication-material"` 时固定归一为 `gpt-image-2`，不支持 `nanobanana2`。`nanobanana2` 参考图作为 `inline_data` 进入 `generateContent`，`gpt-image-2` 参考图进入 edits；`nanobanana2` 的 `512 / 1024 / 2K` 是标量清晰度档位，后端保留 provider 输出几何尺寸，不按 `宽x高` 解析。普通重绘继续走该接口并把当前图层图片作为参考图；图片快速编辑不走该接口。请求可携带 `projectId`、`assetFolderId`、`assetKind`、`generationInputs` 和 `sourceResourceId`，后端生成成功后创建 project resource / 账号素材并在响应中返回 resource / asset 快照。
 - `POST /api/editor/images/background-removals`：接收当前图片源，校验登录态后由 api-server 解析为图片文件并转发到 BiRefNet 去背景服务；请求可携带 `projectId`、`targetLayerId`、`assetFolderId`、`assetLabel`、`sourceResourceId` 和 `canvasCompletion`，有 `canvasCompletion` 时完成后按生成占位写入结果图层，否则沿用旧的目标图层替换路径；响应返回 `imageSrc`、`objectKey`、`assetObjectId`、`width`、`height`、`taskId`、`elapsedMs`、`provider` 和可选 `project` 快照。服务地址由 `GENARRATIVE_EDITOR_BACKGROUND_REMOVAL_BASE_URL` 配置，令牌只在服务端通过 `GENARRATIVE_EDITOR_BACKGROUND_REMOVAL_TOKEN` 注入。
 - `POST /api/editor/icon-spritesheets/generations`：按图标规范图和素材描述数组生成 spritesheet，生成成功后 api-server 先保存带纯色背景 spritesheet 源图，再调用 BgFilter 生成透明 spritesheet。请求支持 `model`、`screenColor`、`segModel`、`aspectRatio`、`imageSize`、`priceMudPoints`、`projectId`、`assetFolderId` 和 `generationInputs`；`priceMudPoints` 必须来自编辑器生成计费配置中对应生图模型的尺寸档位（如 `nanobanana2` 的 `0.5K / 1K / 2K` 或 `gpt-image-2` 的 `1K / 2K`），后端用 `editor_generation_config` 校验后才调用上游；`nanobanana2` 走原生 `generateContent` 并写入 `generationConfig.imageConfig.aspectRatio/imageSize`，`0.5K` 传 `"512"`；`gpt-image-2` 走 `/v1/images/edits`。后端保存透明 spritesheet project resource / 账号素材，并随响应返回对应快照。
 - `POST /api/editor/ui-designs/assets/extractions`：前端把红色框选轮廓绘入本地临时图后，先将该图上传 OSS 并确认 asset object，再以返回的 `objectKey` 作为参考图入队；Data URL / Blob URL 只允许停留在上传前的浏览器临时态。接口固定 `gpt-image-2` 和自动决策纯色背景素材提取提示词生成素材 spritesheet，生成成功后 api-server 先保存带纯色背景 spritesheet 源图，再调用 BgFilter 生成透明 spritesheet，并按连通域自动拆分为 `素材 1..N`，返回结构复用图标 spritesheet 响应。请求必须携带 `screenColor`、`segModel`、`aspectRatio: "1:1"`、`imageSize: "1K" | "2K"` 和 `priceMudPoints`；框选数量不超过 6 个时前端按 `1:1·1K` 与 gpt-image-2 1K 价格提交，超过 6 个时按 `1:1·2K` 与 2K 价格提交。后端必须在调用上游前校验比例、尺寸和泥点价格，只允许 `1:1 / 1K / 2K`。请求可携带 `projectId`、`assetFolderId`、`generationInputs` 和 `spritesheetLabel`，后端保存 spritesheet / 拆分素材并返回对应 resource / asset 快照；前端必须把 spritesheet 原图与拆分素材都加入画布。
-- `POST /api/editor/images/edits`：按提示词和当前图片的已登记 `objectKey` / `resourceId` 调用 VectorEngine edits，返回新的生成图片元数据；图片快速编辑当前只提交 `sourceImageSrc`，不提交隐藏的 `referenceImageSrcs`。画布快速编辑必须把源图精确 `originalWidth x originalHeight` 作为业务目标 `size` 提交，不能重新映射为近似比例或 1K / 2K 预设；api-server 在 VectorEngine provider 边界把目标尺寸和所有 multipart 参考图临时补齐到 16 的倍数，回图后恢复到业务目标精确尺寸，再落 OSS、project resource、账号素材和画布快照。16 对齐尺寸不得泄漏到响应、持久化资源或图层 Resolution。本地红框标记图必须先上传再提交 objectKey；请求携带 project / asset 上下文时由后端创建新 resource / asset，前端只消费响应快照。
+- `POST /api/editor/images/edits`：按提示词和当前图片的已登记 `objectKey` / `resourceId` 调用 VectorEngine edits，返回新的生成图片元数据；图片快速编辑当前只提交 `sourceImageSrc`，不提交隐藏的 `referenceImageSrcs`。画布快速编辑必须把源图精确 `originalWidth x originalHeight` 作为业务目标 `size` 提交，不能重新映射为近似比例或 1K / 2K 预设；api-server 在 VectorEngine provider 边界把目标尺寸和所有 multipart 参考图临时补齐到 16 的倍数，回图后在内存恢复业务目标尺寸，成功时只上传恢复结果，失败时只上传 provider 原图。无论是否发生尺寸恢复都只创建一个 project resource / 账号素材，不显示重复“原始输出”。16 对齐尺寸不得泄漏到正常完成的最终响应、资源或图层 Resolution；变换失败降级时以实际 provider 原图尺寸为准。本地红框标记图必须先上传再提交 objectKey；请求携带 project / asset 上下文时由后端创建新 resource / asset，前端只消费响应快照。
 - `POST /api/editor/videos/generations`：按视频描述、模型、比例、时长、分辨率、模式、声音、默认联网搜索标记和泥点价格生成视频。前端可选模型为 `seedance2.0-fast`、`seedance2.0`、`kling3.0`、`kling3.0-omni`，默认 `seedance2.0-fast`；后端必须将 `seedance2.0-fast` 映射到 `doubao-seedance-2-0-fast-260128`，将 `seedance2.0` 映射到 `doubao-seedance-2-0-260128`，两者不得混用。后端允许 6 类比例、4 到 15 秒整数、`480p / 720p / 1080p`，并拒绝 `seedance2.0-fast + 1080p`；`sound=on/off` 映射 Ark `generate_audio=true/false`。后端复用 Ark / VectorEngine content generation task 轮询链路，下载最终视频并持久化到 OSS；请求携带 `projectId` / `assetFolderId` 时同步创建 project resource / 账号素材并返回 `project` / `asset` 快照，基础响应返回 `videoSrc`、尺寸、prompt、model、provider、taskId、durationSeconds、resolution 和 `priceMudPoints`。
 - `POST /api/editor/audios/sound-effects/generations` 与 `POST /api/editor/audios/background-music/generations`：按音效 / 背景音乐参数生成音频并持久化到 OSS；请求携带 `projectId` / `assetFolderId` 时同步创建 project resource / 账号素材并返回 `project` / `resource` / `asset` 快照，基础响应返回 `audioSrc`、prompt、model、provider、taskId、duration、歌词和 `priceMudPoints`。
 
diff --git a/docs/【后端架构】server-rs与SpacetimeDB数据契约-2026-05-15.md b/docs/【后端架构】server-rs与SpacetimeDB数据契约-2026-05-15.md
index cf34190b6..2f4e8d87a 100644
--- a/docs/【后端架构】server-rs与SpacetimeDB数据契约-2026-05-15.md
+++ b/docs/【后端架构】server-rs与SpacetimeDB数据契约-2026-05-15.md
@@ -62,7 +62,7 @@ npm run check:server-rs-ddd
 - 外部 OpenAPI：`/api/external/v1/openapi.json`、`/api/external/v1/assets/direct-upload-tickets`、`/api/external/v1/assets/objects/confirm`、`/api/external/v1/assets/read-url`、`/api/external/v1/editor/*`，使用 Bearer API Key 鉴权；API Key 管理仍在登录态 `/api/profile/api-keys`，不进入外部 OpenAPI JSON。主站和 External 的 asset object confirm 都必须从已认证主体派生 owner，不能信任请求体 owner；同 bucket / key 已登记后不得改变 owner。
 - 创作 / 游玩支撑能力：`/api/creation-entry/config`、`/api/ai/tasks*`、`/api/runtime/frontend-config`、`/api/runtime/chat/*`、`/api/runtime/settings`、`/api/runtime/save/snapshot`、`/api/profile/browse-history`、`/api/profile/save-archives*`、`/api/profile/play-stats`、`/api/assets/history`、`/api/assets/character-visual/*`、`/api/assets/character-animation/*`、`/api/assets/character-workflow-cache*`、`/api/assets/hyper3d/*`、`/api/runtime/custom-world/asset-studio/*`、`/api/editor/projects*`、`/api/editor/projects/{projectId}/agent-conversations`、`/api/editor/agent-conversations/{conversationId}*`。`/api/runtime/frontend-config` 由 `api-server` 从运行时环境变量下发非敏感 UI 开关；画板右侧 Agent 入口由 `GENARRATIVE_ENABLE_IMAGE_EDITOR_AGENT_SIDEBAR` 控制，默认关闭，前端不再读取 `VITE_*` 构建期变量决定生产显示。`/api/runtime/custom-world/asset-studio/*` 解析默认角色形象 / 动作提示词时可以在 OSS 缓存不可用或未配置时按无缓存返回默认提示；保存 workflow 缓存和真实素材读写仍必须要求 OSS 正常可用。
 - 后台入口配置：`/admin/api/creation-entry/config`、`/admin/api/creation-entry/config/banners` 和 `/admin/api/creation-entry/config/interactions`。
-- 后台素材查询：`GET /admin/api/editor-assets` 通过 `admin_list_editor_assets_and_return` 后台只读 procedure 读取私有账号级 `editor_asset` 中 `source_type = 'generated'` 的素材，支持 `ownerUserId`、`keyword`、`createdAfter`、`createdBefore`、`cursor` 和 `limit`；返回缩略图 / Object Key、作者展示名、陶泥号、提示词、生成输入和生成成本，只用于查询，不提供分类筛选，也不执行精选审核、返还或展示状态修改，不通过后台 SQL 直查私有表。图片放大、视频和音频预览统一调用仅后台管理员会话可访问的 `GET /admin/api/assets/read-url`；该入口仅为预览允许后台跨 owner 签名，不改变主站 `/api/assets/read-*` 或 External API 的 owner 边界。成功换签后必须写入 `event_key = admin_asset_read_url` 的 `tracking_event`，记录管理员 subject、Object Key / legacy path 和有效期，不记录 signed URL。
+- 后台素材查询：`GET /admin/api/editor-assets` 通过 `admin_list_editor_assets_and_return` 后台只读 procedure 读取私有账号级 `editor_asset` 中 `source_type = 'generated'` 的素材，支持 `ownerUserId`、`keyword`、`createdAfter`、`createdBefore`、`cursor` 和 `limit`；返回缩略图 / Object Key、作者展示名、陶泥号、提示词、生成输入和生成成本，只用于查询，不提供分类筛选，也不执行精选审核、返还或展示状态修改，不通过后台 SQL 直查私有表。`editor_green_screen_source`、历史 `editor_image_edit_source` 是同一付费任务的可恢复中间产物，`generation_cost_mud_points = 0` 表示不重复计费，列表和详情统一显示“不单独计费”，不能显示成误导性的“0 泥点”。图片放大、视频和音频预览统一调用仅后台管理员会话可访问的 `GET /admin/api/assets/read-url`；该入口仅为预览允许后台跨 owner 签名，不改变主站 `/api/assets/read-*` 或 External API 的 owner 边界。成功换签后必须写入 `event_key = admin_asset_read_url` 的 `tracking_event`，记录管理员 subject、Object Key / legacy path 和有效期，不记录 signed URL。
 - 自定义世界 / RPG：`/api/runtime/custom-world*`、`/api/story/*`、`/api/runtime/chat/*`。
 - 拼图：`/api/runtime/puzzle/*`。
 - 抓大鹅 Match3D：`/api/creation/match3d/*`、`/api/runtime/match3d/*`。
diff --git a/server-rs/crates/api-server/src/editor_project.rs b/server-rs/crates/api-server/src/editor_project.rs
index e13a581e6..23885150e 100644
--- a/server-rs/crates/api-server/src/editor_project.rs
+++ b/server-rs/crates/api-server/src/editor_project.rs
@@ -116,7 +116,6 @@ const EDITOR_SPEC_IMAGE_ASSET_KIND: &str = "editor_spec_image";
 const EDITOR_QUICK_EDIT_IMAGE_ASSET_KIND: &str = "editor_quick_edit_image";
 const EDITOR_UI_DESIGN_IMAGE_ASSET_KIND: &str = "editor_ui_design_image";
 const EDITOR_IMAGE_EDIT_ASSET_KIND: &str = "editor_image_edit";
-const EDITOR_IMAGE_EDIT_SOURCE_ASSET_KIND: &str = "editor_image_edit_source";
 const EDITOR_ICON_SPRITESHEET_ASSET_KIND: &str = "editor_icon_spritesheet";
 const EDITOR_ICON_SPRITESHEET_SLICE_ASSET_KIND: &str = "editor_icon_spritesheet_slice";
 const EDITOR_ICON_SPRITESHEET_SLICE_WARNING_COMPONENTS: &str = "insufficient-connected-components";
@@ -1604,10 +1603,16 @@ pub(crate) async fn generate_editor_image_for_owner(
             "message": "VectorEngine 未返回图片",
         }))
     })?;
+    let (provider_width, provider_height) = image::load_from_memory(image.bytes.as_slice())
+        .map(|image| (image.width(), image.height()))
+        .map_err(|error| {
+            AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
+                "provider": "vector-engine",
+                "message": format!("素材生成结果不是有效图片：{error}"),
+            }))
+        })?;
+    let storage_profile = editor_generated_image_storage_profile(normalized_kind);
     let source_record = if is_character_generation {
-        let (source_width, source_height) = image::load_from_memory(image.bytes.as_slice())
-            .map(|image| (image.width(), image.height()))
-            .unwrap_or((1024, 1024));
         let source_persisted = persist_editor_green_screen_source_image(
             state,
             caller.owner_user_id.as_str(),
@@ -1630,8 +1635,8 @@ pub(crate) async fn generate_editor_image_for_owner(
                     .clone()
                     .or_else(|| Some(EDITOR_ASSET_DEFAULT_FOLDER_ID.to_string())),
                 label: editor_generated_asset_variant_label(asset_label.as_str(), "原图"),
-                width: source_width,
-                height: source_height,
+                width: provider_width,
+                height: provider_height,
                 prompt: role_setting.clone(),
                 actual_prompt: generated.actual_prompt.clone(),
                 model: generation_options.model.to_string(),
@@ -1664,11 +1669,23 @@ pub(crate) async fn generate_editor_image_for_owner(
 
     // TODO the image size passed to api is already normalized, should remove those normalize
     // and let resize here to get the proper size
-    image = restore_editor_generated_image_output_dimensions(image, image_size.as_ref())?;
+    let (restored_image, dimension_restore_error) =
+        restore_editor_generated_image_output_dimensions_or_original(
+            image,
+            generation_options.model,
+            image_size.as_ref(),
+        );
+    if let Some(error) = dimension_restore_error {
+        tracing::warn!(
+            task_id = %generated.task_id,
+            error = ?error,
+            "画板生成图片尺寸恢复失败，保留并上传 provider 原图"
+        );
+    }
+    image = restored_image;
     let (width, height) = image::load_from_memory(image.bytes.as_slice())
         .map(|image| (image.width(), image.height()))
         .unwrap_or((1024, 1024));
-    let storage_profile = editor_generated_image_storage_profile(normalized_kind);
     let persisted = persist_editor_generated_image(
         state,
         caller.owner_user_id.as_str(),
@@ -1681,8 +1698,8 @@ pub(crate) async fn generate_editor_image_for_owner(
         "image",
         storage_profile.slot,
         "vector-engine",
-    );
-    let persisted = persisted.await?;
+    )
+    .await?;
     let image_src = editor_media_src_from_object_key(persisted.object_key.as_str());
     let generated_asset = persist_editor_generated_asset(
         state,
@@ -2188,9 +2205,9 @@ fn prepare_editor_image_edit_references(
 }
 
 fn restore_editor_image_edit_output_dimensions(
-    output: DownloadedOpenAiImage,
+    output: &DownloadedOpenAiImage,
     dimensions: &EditorImageEditDimensions,
-) -> Result<DownloadedOpenAiImage, AppError> {
+) -> Result<Option<DownloadedOpenAiImage>, AppError> {
     let decoded = image::load_from_memory(output.bytes.as_slice()).map_err(|error| {
         AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
             "provider": "vector-engine",
@@ -2198,7 +2215,7 @@ fn restore_editor_image_edit_output_dimensions(
         }))
     })?;
     if decoded.width() == dimensions.target_width && decoded.height() == dimensions.target_height {
-        return Ok(output);
+        return Ok(None);
     }
 
     let restored = if decoded.width() == dimensions.provider_width
@@ -2212,7 +2229,7 @@ fn restore_editor_image_edit_output_dimensions(
             image::imageops::FilterType::Lanczos3,
         )
     };
-    Ok(DownloadedOpenAiImage {
+    Ok(Some(DownloadedOpenAiImage {
         bytes: encode_editor_image_edit_png(
             restored,
             StatusCode::BAD_GATEWAY,
@@ -2220,13 +2237,42 @@ fn restore_editor_image_edit_output_dimensions(
         )?,
         mime_type: "image/png".to_string(),
         extension: "png".to_string(),
-    })
+    }))
 }
-fn restore_editor_generated_image_output_dimensions(
+
+fn restore_editor_image_edit_output_dimensions_or_original(
     output: DownloadedOpenAiImage,
-    // TODO primitive obsession
+    dimensions: &EditorImageEditDimensions,
+) -> (DownloadedOpenAiImage, Option<AppError>) {
+    match restore_editor_image_edit_output_dimensions(&output, dimensions) {
+        Ok(Some(restored)) => (restored, None),
+        Ok(None) => (output, None),
+        Err(error) => (output, Some(error)),
+    }
+}
+
+fn restore_editor_generated_image_output_dimensions_or_original(
+    output: DownloadedOpenAiImage,
+    model: &str,
     target_size: &str,
-) -> Result<DownloadedOpenAiImage, AppError> {
+) -> (DownloadedOpenAiImage, Option<AppError>) {
+    match restore_editor_generated_image_output_dimensions(&output, model, target_size) {
+        Ok(Some(restored)) => (restored, None),
+        Ok(None) => (output, None),
+        Err(error) => (output, Some(error)),
+    }
+}
+
+fn restore_editor_generated_image_output_dimensions(
+    output: &DownloadedOpenAiImage,
+    model: &str,
+    target_size: &str,
+) -> Result<Option<DownloadedOpenAiImage>, AppError> {
+    // nanobanana 使用标量清晰度档位和独立比例；provider 回图已经是目标几何尺寸，
+    // 不能再把 512 / 1024 / 2K 当作 WIDTHxHEIGHT 解析。
+    if model == EDITOR_IMAGE_MODEL_NANOBANANA2 {
+        return Ok(None);
+    }
     let (target_width, target_height) = target_size.split_once('x').ok_or_else(|| {
         AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
             "provider": "vector-engine",
@@ -2252,7 +2298,7 @@ fn restore_editor_generated_image_output_dimensions(
         }))
     })?;
     if decoded.width() == target_width && decoded.height() == target_height {
-        return Ok(output);
+        return Ok(None);
     }
 
     let restored = decoded.resize_to_fill(
@@ -2260,7 +2306,7 @@ fn restore_editor_generated_image_output_dimensions(
         target_height,
         image::imageops::FilterType::Lanczos3,
     );
-    Ok(DownloadedOpenAiImage {
+    Ok(Some(DownloadedOpenAiImage {
         bytes: encode_editor_image_edit_png(
             restored,
             StatusCode::BAD_GATEWAY,
@@ -2268,7 +2314,7 @@ fn restore_editor_generated_image_output_dimensions(
         )?,
         mime_type: "image/png".to_string(),
         extension: "png".to_string(),
-    })
+    }))
 }
 
 pub async fn edit_editor_image(
@@ -2393,57 +2439,27 @@ pub(crate) async fn edit_editor_image_for_owner(
     let (provider_width, provider_height) =
         image::load_from_memory(provider_image.bytes.as_slice())
             .map(|image| (image.width(), image.height()))
-            .unwrap_or((
-                edit_dimensions.provider_width,
-                edit_dimensions.provider_height,
-            ));
-    let source_persisted = persist_editor_generated_image(
-        state,
-        caller.owner_user_id.as_str(),
-        generated.task_id.as_str(),
-        &provider_image,
-        prompt.as_str(),
-        generated.actual_prompt.as_deref(),
-        EDITOR_IMAGE_EDIT_SOURCE_ASSET_KIND,
-        "edit-images",
-        "image-provider-source",
-        "provider_source",
-        "vector-engine",
-    )
-    .await?;
+            .map_err(|error| {
+                AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
+                    "provider": "vector-engine",
+                    "message": format!("图片改造结果不是有效图片：{error}"),
+                }))
+            })?;
     let asset_label =
         resolve_editor_generated_asset_label(payload.asset_label.clone(), "图片修改结果");
-    let source_label =
-        editor_generated_asset_variant_label(asset_label.as_str(), "原始输出");
-    let _source_record = persist_editor_generated_asset(
-        state,
-        PersistEditorGeneratedAssetInput {
-            project_id: payload.project_id.clone(),
-            owner_user_id: caller.owner_user_id.clone(),
-            folder_id: payload
-                .asset_folder_id
-                .clone()
-                .or_else(|| Some(EDITOR_ASSET_DEFAULT_FOLDER_ID.to_string())),
-            label: source_label,
-            image_src: editor_media_src_from_object_key(source_persisted.object_key.as_str()),
-            object_key: Some(source_persisted.object_key),
-            asset_object_id: Some(source_persisted.asset_object_id),
-            width: provider_width,
-            height: provider_height,
-            prompt: prompt.clone(),
-            actual_prompt: generated.actual_prompt.clone(),
-            model: generation_options.model.to_string(),
-            provider: "VectorEngine".to_string(),
-            task_id: generated.task_id.clone(),
-            source_resource_id: payload.source_resource_id.clone(),
-            asset_kind: Some(EDITOR_IMAGE_EDIT_SOURCE_ASSET_KIND.to_string()),
-            generation_inputs: payload.generation_inputs.clone(),
-            thumbnail_src: None,
-            generation_cost_mud_points: 0,
-        },
-    )
-    .await?;
-    let image = restore_editor_image_edit_output_dimensions(provider_image, &edit_dimensions)?;
+    let (image, dimension_restore_error) =
+        restore_editor_image_edit_output_dimensions_or_original(provider_image, &edit_dimensions);
+    if let Some(error) = dimension_restore_error {
+        tracing::warn!(
+            task_id = %generated.task_id,
+            provider_width,
+            provider_height,
+            target_width = edit_dimensions.target_width,
+            target_height = edit_dimensions.target_height,
+            error = ?error,
+            "画板快速编辑尺寸恢复失败，保留并上传 provider 原图"
+        );
+    }
     let (width, height) = image::load_from_memory(image.bytes.as_slice())
         .map(|image| (image.width(), image.height()))
         .unwrap_or((1024, 1024));
@@ -7530,15 +7546,14 @@ mod tests {
         let extra = image::load_from_memory(references[1].bytes.as_slice()).unwrap();
         assert_eq!((extra.width(), extra.height()), (512, 336));
 
-        let restored = restore_editor_image_edit_output_dimensions(
-            DownloadedOpenAiImage {
-                bytes: encode_png(1552, 1040),
-                mime_type: "image/png".to_string(),
-                extension: "png".to_string(),
-            },
-            &dimensions,
-        )
-        .expect("provider output should restore source dimensions");
+        let provider_output = DownloadedOpenAiImage {
+            bytes: encode_png(1552, 1040),
+            mime_type: "image/png".to_string(),
+            extension: "png".to_string(),
+        };
+        let restored = restore_editor_image_edit_output_dimensions(&provider_output, &dimensions)
+            .expect("provider output should restore source dimensions")
+            .expect("aligned provider output should require a transformation");
         let restored_image = image::load_from_memory(restored.bytes.as_slice()).unwrap();
         assert_eq!(
             (restored_image.width(), restored_image.height()),
@@ -7556,22 +7571,83 @@ mod tests {
             .write_to(&mut bytes, image::ImageFormat::Png)
             .expect("test image should encode");
 
+        let provider_output = DownloadedOpenAiImage {
+            bytes: bytes.into_inner(),
+            mime_type: "image/png".to_string(),
+            extension: "png".to_string(),
+        };
         let restored = restore_editor_generated_image_output_dimensions(
-            DownloadedOpenAiImage {
-                bytes: bytes.into_inner(),
-                mime_type: "image/png".to_string(),
-                extension: "png".to_string(),
-            },
+            &provider_output,
+            GPT_IMAGE_2_MODEL,
             "720x540",
         )
-        .expect("provider output should restore delivery dimensions");
+        .expect("provider output should restore delivery dimensions")
+        .expect("mismatched provider output should require a transformation");
         let restored_image = image::load_from_memory(restored.bytes.as_slice()).unwrap();
 
-        assert_eq!((restored_image.width(), restored_image.height()), (720, 540));
+        assert_eq!(
+            (restored_image.width(), restored_image.height()),
+            (720, 540)
+        );
         assert_eq!(restored.mime_type, "image/png");
         assert_eq!(restored.extension, "png");
     }
 
+    #[test]
+    fn nanobanana_generation_keeps_provider_dimensions_for_scalar_size_tiers() {
+        let image = image::DynamicImage::new_rgba8(896, 896);
+        let mut bytes = Cursor::new(Vec::new());
+        image
+            .write_to(&mut bytes, image::ImageFormat::Png)
+            .expect("test image should encode");
+
+        let provider_output = DownloadedOpenAiImage {
+            bytes: bytes.into_inner(),
+            mime_type: "image/png".to_string(),
+            extension: "png".to_string(),
+        };
+        let (restored, restore_error) =
+            restore_editor_generated_image_output_dimensions_or_original(
+                provider_output,
+                EDITOR_IMAGE_MODEL_NANOBANANA2,
+                "1024",
+            );
+        assert!(restore_error.is_none());
+        let restored_image = image::load_from_memory(restored.bytes.as_slice()).unwrap();
+
+        assert_eq!(
+            (restored_image.width(), restored_image.height()),
+            (896, 896)
+        );
+    }
+
+    #[test]
+    fn failed_dimension_restore_keeps_provider_output_for_single_upload() {
+        let image = image::DynamicImage::new_rgba8(896, 896);
+        let mut bytes = Cursor::new(Vec::new());
+        image
+            .write_to(&mut bytes, image::ImageFormat::Png)
+            .expect("test image should encode");
+        let provider_output = DownloadedOpenAiImage {
+            bytes: bytes.into_inner(),
+            mime_type: "image/png".to_string(),
+            extension: "png".to_string(),
+        };
+        let expected_bytes = provider_output.bytes.clone();
+
+        let (delivery_output, restore_error) =
+            restore_editor_generated_image_output_dimensions_or_original(
+                provider_output,
+                GPT_IMAGE_2_MODEL,
+                "invalid-size",
+            );
+
+        assert!(restore_error.is_some());
+        assert_eq!(delivery_output.bytes, expected_bytes);
+        assert_eq!(delivery_output.mime_type, "image/png");
+        assert_eq!(delivery_output.extension, "png");
+    }
+
     #[test]
     fn editor_generation_dimensions_follow_model_options() {
         let default_generation = normalize_editor_generation_options(None, Some("1:1"), Some("1K"));
@@ -9278,14 +9354,37 @@ mod tests {
         assert_function_contains_in_order(
             source,
             "pub(crate) async fn edit_editor_image_for_owner",
-            "pub async fn generate_editor_icon_spritesheet",
+            "pub async fn remove_editor_image_background",
             &[
                 "let provider_image",
-                "EDITOR_IMAGE_EDIT_SOURCE_ASSET_KIND",
-                "persist_editor_generated_asset",
-                "restore_editor_image_edit_output_dimensions",
+                "restore_editor_image_edit_output_dimensions_or_original",
+                "persist_editor_generated_image",
             ],
         );
+        assert_function_contains_in_order(
+            source,
+            "pub(crate) async fn generate_editor_image_for_owner",
+            "fn normalize_editor_image_generation_size",
+            &[
+                "let mut image",
+                "restore_editor_generated_image_output_dimensions_or_original",
+                "persist_editor_generated_image",
+            ],
+        );
+        assert_function_occurrence_count(
+            source,
+            "pub(crate) async fn edit_editor_image_for_owner",
+            "pub async fn remove_editor_image_background",
+            "persist_editor_generated_image(",
+            1,
+        );
+        assert_function_occurrence_count(
+            source,
+            "pub(crate) async fn generate_editor_image_for_owner",
+            "fn normalize_editor_image_generation_size",
+            "persist_editor_generated_image(",
+            1,
+        );
         assert_function_contains_in_order(
             source,
             "pub(crate) async fn extract_editor_ui_design_assets_for_owner",
@@ -9388,4 +9487,26 @@ mod tests {
             );
         }
     }
+
+    fn assert_function_occurrence_count(
+        source: &str,
+        start: &str,
+        end: &str,
+        snippet: &str,
+        expected_count: usize,
+    ) {
+        let start_index = source
+            .find(start)
+            .unwrap_or_else(|| panic!("missing function start marker: {start}"));
+        let tail = &source[start_index..];
+        let end_index = tail
+            .find(end)
+            .unwrap_or_else(|| panic!("missing function end marker: {end}"));
+        let body = &tail[..end_index];
+        assert_eq!(
+            body.matches(snippet).count(),
+            expected_count,
+            "function body starting with {start} should contain {snippet} exactly {expected_count} times"
+        );
+    }
 }