feat: adapt image tuning by engine

2026-05-29 14:32:02 +08:00
parent 4b21d2999c
commit e36f28a668
12 changed files with 138 additions and 10 deletions
--- a/app/api/generations/image/route.ts
+++ b/app/api/generations/image/route.ts
@@ -31,6 +31,7 @@ export async function POST(request: Request) {
      min_ratio?: number;
      max_ratio?: number;
      force_single?: boolean;
+      quality?: string;
    }>(request);
    const capability = body.capability || "image.generate";
    const assembled = body.promptAssembly
@@ -49,7 +50,8 @@ export async function POST(request: Request) {
      height: asNumber(body.height),
      min_ratio: asNumber(body.min_ratio),
      max_ratio: asNumber(body.max_ratio),
-      force_single: Boolean(body.force_single)
+      force_single: Boolean(body.force_single),
+      quality: typeof body.quality === "string" ? body.quality : undefined
    }, requestOrigin(request));
    return jsonOk({ job: await getGenerationJob(job.id) }, { status: 202 });
  } catch (error) {
--- a/app/api/v1/openapi.json/route.ts
+++ b/app/api/v1/openapi.json/route.ts
@@ -120,8 +120,9 @@ export async function GET(request: Request) {
            },
            width: { type: "integer", example: 1440 },
            height: { type: "integer", example: 2560 },
-            scale: { type: "number", minimum: 1, maximum: 100 },
+            scale: { type: "number", minimum: 1, maximum: 100, description: "Jimeng text influence for image.generate and detail strength for image.upscale." },
            force_single: { type: "boolean" },
+            quality: { type: "string", enum: ["low", "medium", "high"], description: "EvoLink image quality for image.generate and image.inpaint." },
            resolution: { type: "string", enum: ["4k", "8k"], description: "Upscale resolution for image.upscale." },
            seed: { type: "integer" },
            priority: { type: "integer", minimum: -100, maximum: 100 },
--- a/components/create-studio.tsx
+++ b/components/create-studio.tsx
@@ -13,12 +13,31 @@ import type { PromptMaterial } from "@/lib/prompt/assembler";
 type GenerateMode = "image" | "video";
 type StudioMode = GenerateMode | ImageEditMode;
 type MaterialKind = PromptMaterial["type"];
+type ImageGenerateEngine = "jimeng" | "evolink";

 type MentionState = {
  start: number;
  query: string;
 };

+type HealthCapability = {
+  id: string;
+  engine?: string;
+};
+
+const jimengInfluenceOptions = [
+  { id: "creative", label: "创意 35", scale: 35 },
+  { id: "balanced", label: "均衡 50", scale: 50 },
+  { id: "precise", label: "贴合 70", scale: 70 },
+  { id: "strict", label: "严格 85", scale: 85 }
+];
+
+const evolinkQualityOptions = [
+  { id: "low", label: "快速", quality: "low" },
+  { id: "medium", label: "标准", quality: "medium" },
+  { id: "high", label: "精细", quality: "high" }
+];
+
 const imageSizePresets = [
  { label: "1:1", width: 2048, height: 2048 },
  { label: "4:3", width: 2304, height: 1728 },
@@ -41,7 +60,9 @@ export function CreateStudio({ initialMode = "image" }: { initialMode?: StudioMo
  const [error, setError] = useState<string | null>(null);
  const [notice, setNotice] = useState<string | null>(null);
  const [imageSize, setImageSize] = useState(imageSizePresets[0]);
-  const [imageScale, setImageScale] = useState(50);
+  const [imageEngine, setImageEngine] = useState<ImageGenerateEngine>("jimeng");
+  const [jimengInfluence, setJimengInfluence] = useState(jimengInfluenceOptions[1].id);
+  const [evolinkQuality, setEvolinkQuality] = useState(evolinkQualityOptions[1].id);
  const [forceSingle, setForceSingle] = useState(true);
  const [videoRatio, setVideoRatio] = useState("9:16");
  const [videoDuration, setVideoDuration] = useState(VIDEO_DURATION_DEFAULT);
@@ -59,6 +80,8 @@ export function CreateStudio({ initialMode = "image" }: { initialMode?: StudioMo
  const isImageEditMode = mode === "inpaint" || mode === "upscale";
  const generateMode: GenerateMode = mode === "video" ? "video" : "image";
  const prompt = promptByMode[generateMode];
+  const selectedJimengInfluence = jimengInfluenceOptions.find((option) => option.id === jimengInfluence) || jimengInfluenceOptions[1];
+  const selectedEvolinkQuality = evolinkQualityOptions.find((option) => option.id === evolinkQuality) || evolinkQualityOptions[1];
  const visibleMaterials = pageItems(materials, materialPage, MATERIAL_PAGE_SIZE);
  const materialPageOffset = (clampPage(materialPage, materials.length, MATERIAL_PAGE_SIZE) - 1) * MATERIAL_PAGE_SIZE;
  const mentionSuggestions = useMemo(() => {
@@ -90,6 +113,20 @@ export function CreateStudio({ initialMode = "image" }: { initialMode?: StudioMo
    if (materialBoardRef.current) revealChildren(materialBoardRef.current, ".material-card");
  }, [materials.length, materialPage]);

+  useEffect(() => {
+    let active = true;
+    void fetch("/api/health")
+      .then((response) => response.ok ? response.json() : null)
+      .then((payload: { capabilities?: HealthCapability[] } | null) => {
+        const engine = payload?.capabilities?.find((capability) => capability.id === "image.generate")?.engine;
+        if (active && (engine === "evolink" || engine === "jimeng")) setImageEngine(engine);
+      })
+      .catch(() => undefined);
+    return () => {
+      active = false;
+    };
+  }, []);
+
  function setPrompt(value: string) {
    setPromptByMode((items) => ({ ...items, [generateMode]: value }));
  }
@@ -229,7 +266,9 @@ export function CreateStudio({ initialMode = "image" }: { initialMode?: StudioMo
            materials: materials.filter((material) => material.type === "image"),
            width: imageSize.width,
            height: imageSize.height,
-            scale: imageScale,
+            ...(imageEngine === "evolink"
+              ? { quality: selectedEvolinkQuality.quality }
+              : { scale: selectedJimengInfluence.scale }),
            force_single: forceSingle
          }
        : {
@@ -432,9 +471,17 @@ export function CreateStudio({ initialMode = "image" }: { initialMode?: StudioMo
                  {imageSizePresets.map((preset) => <option key={preset.label}>{preset.label}</option>)}
                </select>
              </div>
-              <div className="field inline-field range-field">
-                <label htmlFor="imageScale">文本影响 {imageScale}</label>
-                <input id="imageScale" type="range" min="1" max="100" value={imageScale} onChange={(event) => setImageScale(Number(event.target.value))} />
+              <div className="field inline-field">
+                <label htmlFor="imageEngineTuning">{imageEngine === "evolink" ? "生成质量" : "文本影响"}</label>
+                {imageEngine === "evolink" ? (
+                  <select id="imageEngineTuning" value={evolinkQuality} onChange={(event) => setEvolinkQuality(event.target.value)}>
+                    {evolinkQualityOptions.map((option) => <option key={option.id} value={option.id}>{option.label}</option>)}
+                  </select>
+                ) : (
+                  <select id="imageEngineTuning" value={jimengInfluence} onChange={(event) => setJimengInfluence(event.target.value)}>
+                    {jimengInfluenceOptions.map((option) => <option key={option.id} value={option.id}>{option.label}</option>)}
+                  </select>
+                )}
              </div>
              <label className="toggle-line">
                <input type="checkbox" checked={forceSingle} onChange={(event) => setForceSingle(event.target.checked)} />
--- a/docs/API.md
+++ b/docs/API.md
@@ -174,10 +174,14 @@ curl -X POST https://你的域名/api/v1/assets \
  "imageUrls": ["https://example.com/reference.png"],
  "width": 1440,
  "height": 2560,
+  "scale": 50,
+  "quality": "medium",
  "force_single": true
 }
 ```

+图片生成参数会按当前引擎生效：即梦使用 `scale` 控制文本影响，EvoLink 使用 `quality` 控制生成质量。
+
 智能超清：

 ```json
--- a/findings.md
+++ b/findings.md
@@ -136,3 +136,8 @@
 - Uploaded and generated assets created through public API flows are tagged as `api-client:<clientId>` so integrations can query and download their own results later.
 - OpenAPI is generated dynamically from the current deployment origin at `/api/v1/openapi.json`.
 - Operations handoff docs live in `docs/DEPLOYMENT.md`; partner API docs live in `docs/API.md`.
+
+## 2026-05-29 Image Tuning Findings
+- Jimeng image generation supports the current `scale` parameter, so UI presets should submit numeric text-influence values for that engine.
+- EvoLink image generation does not use Jimeng `scale`; the per-request engine-aware control should submit EvoLink `quality` instead.
+- Current local `/api/health` reports `image.generate` using EvoLink, so `/create` should show `生成质量` options rather than `文本影响`.
--- a/lib/evolink/image-client.ts
+++ b/lib/evolink/image-client.ts
@@ -87,7 +87,8 @@ export function buildEvolinkImagePayload(
  };

  if (!payload.prompt) throw new Error("Prompt is required for image generation.");
-  if (settings.quality) payload.quality = settings.quality;
+  const quality = cleanOptional(typeof input.quality === "string" ? input.quality : undefined) || settings.quality;
+  if (quality) payload.quality = quality;
  if (settings.resolution) payload.resolution = settings.resolution;
  assignSize(payload, input);

--- a/lib/server/generation-service.ts
+++ b/lib/server/generation-service.ts
@@ -42,6 +42,7 @@ export type SubmitImageJobInput = {
  max_ratio?: number;
  force_single?: boolean;
  resolution?: "4k" | "8k";
+  quality?: string;
  seed?: number;
  retryOf?: string;
  idempotencyKey?: string;
--- a/lib/server/public-api-jobs.ts
+++ b/lib/server/public-api-jobs.ts
@@ -35,6 +35,7 @@ export type PublicJobCreateBody = {
  max_ratio?: number;
  force_single?: boolean;
  resolution?: "4k" | "8k";
+  quality?: string;
  seed?: number;
 };

@@ -101,6 +102,7 @@ export async function createPublicGenerationJob(input: {
    max_ratio: asNumber(input.body.max_ratio),
    force_single: Boolean(input.body.force_single),
    resolution: input.body.resolution,
+    quality: normalizeQuality(input.body.quality),
    seed: asNumber(input.body.seed)
  } satisfies SubmitImageJobInput, input.origin);
  return { job, reused: false };
@@ -132,6 +134,12 @@ function asNumber(value: unknown): number | undefined {
  return Number.isFinite(parsed) ? parsed : undefined;
 }

+function normalizeQuality(value: unknown): string | undefined {
+  if (typeof value !== "string") return undefined;
+  const normalized = value.trim().toLowerCase();
+  return ["low", "medium", "high"].includes(normalized) ? normalized : undefined;
+}
+
 function fingerprintBody(body: PublicJobCreateBody): string {
  const { idempotencyKey: _idempotencyKey, ...fingerprintSource } = body;
  return createHash("sha256").update(stableStringify(fingerprintSource)).digest("hex");
--- a/progress.md
+++ b/progress.md
@@ -295,6 +295,27 @@
  - Local HTTP smoke test uploaded a PNG through `/api/v1/assets`, fetched `/api/v1/assets/:id`, and downloaded matching binary bytes from `/api/v1/assets/:id/download`.
  - Mock-provider task flow created a queued job through `/api/v1/jobs`, Worker processed it to `succeeded`, generated output asset carried the API client tag, and download returned an attachment response.

+## Session: 2026-05-29 - Engine-Aware Image Tuning
+
+### Implementation
+- **Status:** complete
+- Actions taken:
+  - Replaced the image generation text-influence range slider with select options.
+  - Added create-page engine detection from `/api/health`.
+  - For Jimeng image generation, the UI now shows `文本影响` options: `创意 35`, `均衡 50`, `贴合 70`, `严格 85`, and submits `scale`.
+  - For EvoLink image generation, the UI now shows `生成质量` options: `快速`, `标准`, `精细`, and submits `quality`.
+  - Added per-request EvoLink `quality` support in the payload builder.
+  - Updated public API request typing, OpenAPI schema, API docs, and focused tests.
+
+### Verification
+- **Status:** complete
+- Results:
+  - `npm test`: 7 files / 22 tests passed.
+  - `npm run build`: production build succeeded.
+  - Agent-browser snapshot on `/create` showed current EvoLink mode rendering `生成质量` with `快速 / 标准 / 精细`.
+  - Desktop and mobile screenshots showed the create-page controls fitting cleanly.
+  - Browser page errors list was empty.
+
 ## Error Log - Server Deployment Support
 | Timestamp | Error | Attempt | Resolution |
 |-----------|-------|---------|------------|
--- a/task_plan.md
+++ b/task_plan.md
@@ -86,6 +86,14 @@ Complete - latest update: Task management and public API v1
 - [x] Verify real HTTP API calls, Worker task processing, tests, and production build
 - **Status:** complete

+### Phase 12: Engine-Aware Image Tuning
+- [x] Replace the free text-influence slider with user-facing option presets
+- [x] Detect the active image generation engine from `/api/health`
+- [x] Send Jimeng `scale` only when the active engine is Jimeng
+- [x] Send EvoLink `quality` only when the active engine is EvoLink
+- [x] Verify tests, production build, and desktop/mobile create-page layout
+- **Status:** complete
+
 ## Key Questions
 1. How should the selected image engine be stored and exposed in settings?
 2. Which current capabilities should EvoLink handle first?
--- a/tests/evolink-image-client.test.ts
+++ b/tests/evolink-image-client.test.ts
@@ -34,15 +34,18 @@ describe("EvoLink image client helpers", () => {
  it("maps inpainting original and mask URLs", () => {
    const payload = buildEvolinkImagePayload("image.inpaint", {
      prompt: "移除背景杂物",
+      quality: "high",
      imageUrls: ["https://example.com/original.png", "https://example.com/mask.png"]
    }, {
      baseUrl: "https://api.evolink.ai",
-      model: "gpt-image-2"
+      model: "gpt-image-2",
+      quality: "medium"
    });

    expect(payload).toMatchObject({
      image_urls: ["https://example.com/original.png"],
-      mask_url: "https://example.com/mask.png"
+      mask_url: "https://example.com/mask.png",
+      quality: "high"
    });
  });

--- a/tests/task-management.test.ts
+++ b/tests/task-management.test.ts
@@ -21,6 +21,9 @@ const envNames = [
  "SUPABASE_SERVICE_ROLE_KEY",
  "ZHINIAN_API_KEYS",
  "JIMENG_VISUAL_MOCK",
+  "IMAGE_GENERATE_ENGINE",
+  "EVOLINK_MOCK",
+  "EVOLINK_API_KEY",
  "VOLCENGINE_ACCESS_KEY_ID",
  "VOLCENGINE_SECRET_ACCESS_KEY",
  "ZHINIAN_WEBHOOK_SECRET"
@@ -33,6 +36,9 @@ describe("task management and public API helpers", () => {
    process.env.ZHINIAN_RUNTIME_DIR = runtimeDir;
    process.env.ZHINIAN_API_KEYS = "agent-a:secret-a,agent-b:secret-b";
    process.env.JIMENG_VISUAL_MOCK = "true";
+    delete process.env.IMAGE_GENERATE_ENGINE;
+    delete process.env.EVOLINK_MOCK;
+    delete process.env.EVOLINK_API_KEY;
    delete process.env.NEXT_PUBLIC_SUPABASE_URL;
    delete process.env.SUPABASE_SERVICE_ROLE_KEY;
    delete process.env.VOLCENGINE_ACCESS_KEY_ID;
@@ -91,6 +97,27 @@ describe("task management and public API helpers", () => {
    })).rejects.toBeInstanceOf(PublicApiConflictError);
  });

+  it("passes EvoLink quality through public image jobs", async () => {
+    process.env.IMAGE_GENERATE_ENGINE = "evolink";
+    process.env.EVOLINK_MOCK = "true";
+    process.env.EVOLINK_API_KEY = "test-key";
+    const result = await createPublicGenerationJob({
+      client: { id: "agent-a", key: "secret-a" },
+      request: new Request("http://local.test/api/v1/jobs"),
+      origin: "http://local.test",
+      body: {
+        capability: "image.generate",
+        prompt: "商品海报",
+        quality: "high"
+      }
+    });
+    expect(result.job.provider).toBe("mock");
+    expect(result.job.requestPayload.providerPayload).toMatchObject({
+      model: "gpt-image-2",
+      quality: "high"
+    });
+  });
+
  it("claims local jobs without duplicate ownership", async () => {
    await Promise.all(Array.from({ length: 6 }, (_, index) => createGenerationJob({
      ownerId: DEFAULT_OWNER_ID,