Make Replicate image requests schema-aware without overpromising multi-image output

Replicate image models no longer share one assumed Nano Banana input shape. This change routes Seedream and Wan requests through model-family-specific input builders, switches the default Replicate model to nano-banana-2, and adds local validation for the most common schema mismatches before the API returns a remote error. The implementation stays deliberately narrow. baoyu-imagine still writes one output file per Replicate request, so this commit blocks --n on the Replicate provider instead of exposing partial multi-image behavior that would silently drop outputs. Constraint: baoyu-imagine currently persists a single output path per request Constraint: Replicate model families use different request schemas and output defaults Rejected: Port the previous broader PR as-is | it enabled multi-image request fields without completing multi-image output handling Rejected: Add full multi-image output support here | larger behavioral surface than needed for a safe first merge Confidence: high Scope-risk: moderate Reversibility: clean Directive: Do not re-enable Replicate --n until the CLI can name, download, and report multiple output files end to end Tested: node --import tsx --test skills/baoyu-imagine/scripts/providers/replicate.test.ts Tested: node --import tsx --test skills/baoyu-imagine/scripts/main.test.ts Tested: npm test Not-tested: Live Replicate API calls against Seedream and Wan credentials Co-authored-by: justnodejs <justnodejs@gmail.com>
2026-04-11 21:16:39 -05:00 · 2026-04-11 21:16:39 -05:00 · 648f563378
parent ec5f4ffcc9
commit 648f563378
8 changed files with 383 additions and 21 deletions
--- a/README.md
+++ b/README.md
@ -806,7 +806,7 @@ AI SDK-based image generation using OpenAI, Azure OpenAI, Google, OpenRouter, Da
 | `GOOGLE_IMAGE_MODEL` | Google model | `gemini-3-pro-image-preview` |
 | `DASHSCOPE_IMAGE_MODEL` | DashScope model | `qwen-image-2.0-pro` |
 | `MINIMAX_IMAGE_MODEL` | MiniMax model | `image-01` |
-| `REPLICATE_IMAGE_MODEL` | Replicate model | `google/nano-banana-pro` |
+| `REPLICATE_IMAGE_MODEL` | Replicate model | `google/nano-banana-2` |
 | `JIMENG_IMAGE_MODEL` | Jimeng model | `jimeng_t2i_v40` |
 | `SEEDREAM_IMAGE_MODEL` | Seedream model | `doubao-seedream-5-0-260128` |
 | `OPENAI_BASE_URL` | Custom OpenAI endpoint | - |
@ -1146,7 +1146,7 @@ MINIMAX_IMAGE_MODEL=image-01

 # Replicate
 REPLICATE_API_TOKEN=r8_xxx
-REPLICATE_IMAGE_MODEL=google/nano-banana-pro
+REPLICATE_IMAGE_MODEL=google/nano-banana-2
 # REPLICATE_BASE_URL=https://api.replicate.com

 # Jimeng (即梦)
--- a/README.zh.md
+++ b/README.zh.md
@ -806,7 +806,7 @@ AI 驱动的生成后端。
 | `GOOGLE_IMAGE_MODEL` | Google 模型 | `gemini-3-pro-image-preview` |
 | `DASHSCOPE_IMAGE_MODEL` | DashScope 模型 | `qwen-image-2.0-pro` |
 | `MINIMAX_IMAGE_MODEL` | MiniMax 模型 | `image-01` |
-| `REPLICATE_IMAGE_MODEL` | Replicate 模型 | `google/nano-banana-pro` |
+| `REPLICATE_IMAGE_MODEL` | Replicate 模型 | `google/nano-banana-2` |
 | `JIMENG_IMAGE_MODEL` | 即梦模型 | `jimeng_t2i_v40` |
 | `SEEDREAM_IMAGE_MODEL` | 豆包模型 | `doubao-seedream-5-0-260128` |
 | `OPENAI_BASE_URL` | 自定义 OpenAI 端点 | - |
@ -1146,7 +1146,7 @@ MINIMAX_IMAGE_MODEL=image-01

 # Replicate
 REPLICATE_API_TOKEN=r8_xxx
-REPLICATE_IMAGE_MODEL=google/nano-banana-pro
+REPLICATE_IMAGE_MODEL=google/nano-banana-2
 # REPLICATE_BASE_URL=https://api.replicate.com

 # 即梦（Jimeng）
--- a/skills/baoyu-imagine/SKILL.md
+++ b/skills/baoyu-imagine/SKILL.md
@ -112,7 +112,7 @@ ${BUN_X} {baseDir}/scripts/main.ts --prompt "A girl stands by the library window
 # MiniMax with custom size (documented for image-01)
 ${BUN_X} {baseDir}/scripts/main.ts --prompt "A cinematic poster" --image out.jpg --provider minimax --model image-01 --size 1536x1024

-# Replicate (google/nano-banana-pro)
+# Replicate (google/nano-banana-2)
 ${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider replicate

 # Replicate with specific model
@ -136,7 +136,7 @@ ${BUN_X} {baseDir}/scripts/main.ts --batchfile batch.json --jobs 4 --json
      "promptFiles": ["prompts/hero.md"],
      "image": "out/hero.png",
      "provider": "replicate",
-      "model": "google/nano-banana-pro",
+      "model": "google/nano-banana-2",
      "ar": "16:9",
      "quality": "2k"
    },
@ -192,7 +192,7 @@ Paths in `promptFiles`, `image`, and `ref` are resolved relative to the batch fi
 | `GOOGLE_IMAGE_MODEL` | Google model override |
 | `DASHSCOPE_IMAGE_MODEL` | DashScope model override (default: `qwen-image-2.0-pro`) |
 | `MINIMAX_IMAGE_MODEL` | MiniMax model override (default: `image-01`) |
-| `REPLICATE_IMAGE_MODEL` | Replicate model override (default: google/nano-banana-pro) |
+| `REPLICATE_IMAGE_MODEL` | Replicate model override (default: google/nano-banana-2) |
 | `JIMENG_IMAGE_MODEL` | Jimeng model override (default: jimeng_t2i_v40) |
 | `SEEDREAM_IMAGE_MODEL` | Seedream model override (default: doubao-seedream-5-0-260128) |
 | `OPENAI_BASE_URL` | Custom OpenAI endpoint |
@ -324,7 +324,7 @@ Notes:

 Supported model formats:

- `owner/name` (recommended for official models), e.g. `google/nano-banana-pro`
+- `owner/name` (recommended for official models), e.g. `google/nano-banana-2`
 - `owner/name:version` (community models by version), e.g. `stability-ai/sdxl:<version>`

 Examples:
--- a/skills/baoyu-imagine/references/config/first-time-setup.md
+++ b/skills/baoyu-imagine/references/config/first-time-setup.md
@ -56,7 +56,7 @@ options:
  - label: "MiniMax"
    description: "MiniMax image generation with subject-reference character workflows"
  - label: "Replicate"
-    description: "Community models - nano-banana-pro, flexible model selection"
+    description: "Community models - nano-banana-2, flexible model selection"
 ```

 ### Question 2: Default Google Model
@ -263,7 +263,7 @@ Notes for DashScope setup:
 header: "Replicate Model"
 question: "Choose a default Replicate image generation model?"
 options:
-  - label: "google/nano-banana-pro (Recommended)"
+  - label: "google/nano-banana-2 (Recommended)"
    description: "Google's fast image model on Replicate"
  - label: "google/nano-banana"
    description: "Google's base image model on Replicate"
--- a/skills/baoyu-imagine/references/config/preferences-schema.md
+++ b/skills/baoyu-imagine/references/config/preferences-schema.md
@ -26,7 +26,7 @@ default_model:
  openrouter: null          # e.g., "google/gemini-3.1-flash-image-preview"
  dashscope: null           # e.g., "qwen-image-2.0-pro"
  minimax: null             # e.g., "image-01"
-  replicate: null           # e.g., "google/nano-banana-pro"
+  replicate: null           # e.g., "google/nano-banana-2"

 batch:
  max_workers: 10
@ -101,7 +101,7 @@ default_model:
  openrouter: "google/gemini-3.1-flash-image-preview"
  dashscope: "qwen-image-2.0-pro"
  minimax: "image-01"
-  replicate: "google/nano-banana-pro"
+  replicate: "google/nano-banana-2"
 batch:
  max_workers: 10
  provider_limits:
--- a/skills/baoyu-imagine/scripts/main.ts
+++ b/skills/baoyu-imagine/scripts/main.ts
@ -96,7 +96,7 @@ Batch file format:
        "promptFiles": ["prompts/hero.md"],
        "image": "out/hero.png",
        "provider": "replicate",
-        "model": "google/nano-banana-pro",
+        "model": "google/nano-banana-2",
        "ar": "16:9"
      }
    ]
@ -123,7 +123,7 @@ Environment variables:
  GOOGLE_IMAGE_MODEL        Default Google model (gemini-3-pro-image-preview)
  DASHSCOPE_IMAGE_MODEL     Default DashScope model (qwen-image-2.0-pro)
  MINIMAX_IMAGE_MODEL       Default MiniMax model (image-01)
-  REPLICATE_IMAGE_MODEL     Default Replicate model (google/nano-banana-pro)
+  REPLICATE_IMAGE_MODEL     Default Replicate model (google/nano-banana-2)
  JIMENG_IMAGE_MODEL        Default Jimeng model (jimeng_t2i_v40)
  SEEDREAM_IMAGE_MODEL      Default Seedream model (doubao-seedream-5-0-260128)
  OPENAI_BASE_URL           Custom OpenAI endpoint
--- a/skills/baoyu-imagine/scripts/providers/replicate.test.ts
+++ b/skills/baoyu-imagine/scripts/providers/replicate.test.ts
@ -5,7 +5,10 @@ import type { CliArgs } from "../types.ts";
 import {
  buildInput,
  extractOutputUrl,
+  generateImage,
+  getDefaultOutputExtension,
  parseModelId,
+  validateArgs,
 } from "./replicate.ts";

 function makeArgs(overrides: Partial<CliArgs> = {}): CliArgs {
@ -47,21 +50,20 @@ test("Replicate model parsing accepts official formats and rejects malformed one
  );
 });

-test("Replicate input builder maps aspect ratio, image count, quality, and refs", () => {
+test("Replicate input builder keeps nano-banana mapping for compatible models", () => {
  assert.deepEqual(
    buildInput(
      "A robot painter",
+      "google/nano-banana-2",
      makeArgs({
        aspectRatio: "16:9",
        quality: "2k",
-        n: 3,
      }),
      ["data:image/png;base64,AAAA"],
    ),
    {
      prompt: "A robot painter",
      aspect_ratio: "16:9",
-      number_of_images: 3,
      resolution: "2K",
      output_format: "png",
      image_input: ["data:image/png;base64,AAAA"],
@ -69,7 +71,7 @@ test("Replicate input builder maps aspect ratio, image count, quality, and refs"
  );

  assert.deepEqual(
-    buildInput("A robot painter", makeArgs({ quality: "normal" }), ["ref"]),
+    buildInput("A robot painter", "google/nano-banana-pro", makeArgs({ quality: "normal" }), ["ref"]),
    {
      prompt: "A robot painter",
      aspect_ratio: "match_input_image",
@ -80,6 +82,120 @@ test("Replicate input builder maps aspect ratio, image count, quality, and refs"
  );
 });

+test("Replicate input builder maps Seedream models to their native schema", () => {
+  assert.deepEqual(
+    buildInput(
+      "A robot painter",
+      "bytedance/seedream-4.5",
+      makeArgs({
+        size: "1536x1024",
+        aspectRatio: "16:9",
+      }),
+      ["data:image/png;base64,AAAA"],
+    ),
+    {
+      prompt: "A robot painter",
+      size: "custom",
+      width: 1536,
+      height: 1024,
+      image_input: ["data:image/png;base64,AAAA"],
+    },
+  );
+
+  assert.deepEqual(
+    buildInput(
+      "A robot painter",
+      "bytedance/seedream-5-lite",
+      makeArgs({
+        size: "3K",
+        aspectRatio: "4:3",
+      }),
+      [],
+    ),
+    {
+      prompt: "A robot painter",
+      size: "3K",
+      output_format: "png",
+      aspect_ratio: "4:3",
+    },
+  );
+});
+
+test("Replicate input builder maps Wan models to their native schema", () => {
+  assert.deepEqual(
+    buildInput(
+      "A robot painter",
+      "wan-video/wan-2.7-image-pro",
+      makeArgs({
+        quality: "2k",
+      }),
+      ["data:image/png;base64,AAAA"],
+    ),
+    {
+      prompt: "A robot painter",
+      size: "2K",
+      images: ["data:image/png;base64,AAAA"],
+      thinking_mode: false,
+    },
+  );
+
+  assert.deepEqual(
+    buildInput(
+      "A robot painter",
+      "wan-video/wan-2.7-image",
+      makeArgs({
+        size: "1536x1024",
+      }),
+      [],
+    ),
+    {
+      prompt: "A robot painter",
+      size: "1536*1024",
+      thinking_mode: true,
+    },
+  );
+});
+
+test("Replicate validation rejects unsupported schema combinations before the API call", () => {
+  assert.throws(
+    () => validateArgs("google/nano-banana-2", makeArgs({ n: 2 })),
+    /Replicate --n is not supported yet/,
+  );
+
+  assert.throws(
+    () => validateArgs("bytedance/seedream-4.5", makeArgs({ size: "8x8" })),
+    /must keep width and height between 1024 and 4096/,
+  );
+
+  assert.throws(
+    () => validateArgs("bytedance/seedream-5-lite", makeArgs({ size: "4K" })),
+    /requires --size to be 2K or 3K/,
+  );
+
+  assert.throws(
+    () => validateArgs("wan-video/wan-2.7-image-pro", makeArgs({ aspectRatio: "16:9" })),
+    /do not accept --ar/,
+  );
+
+  assert.throws(
+    () => validateArgs("wan-video/wan-2.7-image", makeArgs({ referenceImages: Array.from({ length: 10 }, () => "ref.png") })),
+    /at most 9 reference images/,
+  );
+
+  assert.throws(
+    () => validateArgs("wan-video/wan-2.7-image-pro", makeArgs({ size: "4K", referenceImages: ["ref.png"] })),
+    /only supports 4K for text-to-image requests/,
+  );
+
+  assert.doesNotThrow(
+    () => validateArgs("bytedance/seedream-4.5", makeArgs({ size: "1536x1024" })),
+  );
+
+  assert.doesNotThrow(
+    () => validateArgs("wan-video/wan-2.7-image", makeArgs({ size: "1920x1080" })),
+  );
+});
+
 test("Replicate output extraction supports string, array, and object URLs", () => {
  assert.equal(
    extractOutputUrl({ output: "https://example.com/a.png" } as never),
@ -99,3 +215,31 @@ test("Replicate output extraction supports string, array, and object URLs", () =
    /Unexpected Replicate output format/,
  );
 });
+
+test("Replicate default output extension matches model family behavior", () => {
+  assert.equal(getDefaultOutputExtension("bytedance/seedream-4.5"), ".jpg");
+  assert.equal(getDefaultOutputExtension("bytedance/seedream-5-lite"), ".png");
+  assert.equal(getDefaultOutputExtension("google/nano-banana-2"), ".png");
+});
+
+test("Replicate generateImage validates arguments before making API requests", async () => {
+  const previousToken = process.env.REPLICATE_API_TOKEN;
+  process.env.REPLICATE_API_TOKEN = "test-token";
+
+  try {
+    await assert.rejects(
+      generateImage(
+        "A robot painter",
+        "wan-video/wan-2.7-image-pro",
+        makeArgs({ aspectRatio: "16:9" }),
+      ),
+      /do not accept --ar/,
+    );
+  } finally {
+    if (previousToken === undefined) {
+      delete process.env.REPLICATE_API_TOKEN;
+    } else {
+      process.env.REPLICATE_API_TOKEN = previousToken;
+    }
+  }
+});
--- a/skills/baoyu-imagine/scripts/providers/replicate.ts
+++ b/skills/baoyu-imagine/scripts/providers/replicate.ts
@ -2,10 +2,16 @@ import path from "node:path";
 import { readFile } from "node:fs/promises";
 import type { CliArgs } from "../types";

-const DEFAULT_MODEL = "google/nano-banana-pro";
+const DEFAULT_MODEL = "google/nano-banana-2";
 const SYNC_WAIT_SECONDS = 60;
 const POLL_INTERVAL_MS = 2000;
 const MAX_POLL_MS = 300_000;
+const SEEDREAM_45_SIZES = new Set(["2K", "4K"]);
+const SEEDREAM_5_LITE_SIZES = new Set(["2K", "3K"]);
+const WAN_SIZES = new Set(["1K", "2K"]);
+const WAN_PRO_SIZES = new Set(["1K", "2K", "4K"]);
+const SEEDREAM_45_CUSTOM_MIN = 1024;
+const SEEDREAM_45_CUSTOM_MAX = 4096;

 export function getDefaultModel(): string {
  return process.env.REPLICATE_IMAGE_MODEL || DEFAULT_MODEL;
@ -31,7 +37,63 @@ export function parseModelId(model: string): { owner: string; name: string; vers
  return { owner: parts[0], name: parts[1], version: version || null };
 }

-export function buildInput(prompt: string, args: CliArgs, referenceImages: string[]): Record<string, unknown> {
+function isSeedream45Model(model: string): boolean {
+  return model.startsWith("bytedance/seedream-4.5");
+}
+
+function isSeedream5LiteModel(model: string): boolean {
+  return model.startsWith("bytedance/seedream-5-lite");
+}
+
+function isSeedreamModel(model: string): boolean {
+  return isSeedream45Model(model) || isSeedream5LiteModel(model);
+}
+
+function isWanProModel(model: string): boolean {
+  return model.startsWith("wan-video/wan-2.7-image-pro");
+}
+
+function isWanModel(model: string): boolean {
+  return model.startsWith("wan-video/wan-2.7-image");
+}
+
+function parsePixelSize(size: string): { width: number; height: number } | null {
+  const match = size.trim().match(/^(\d+)\s*[xX*]\s*(\d+)$/);
+  if (!match) return null;
+
+  const width = Number.parseInt(match[1]!, 10);
+  const height = Number.parseInt(match[2]!, 10);
+  if (!Number.isFinite(width) || !Number.isFinite(height) || width <= 0 || height <= 0) {
+    return null;
+  }
+
+  return { width, height };
+}
+
+function normalizePixelSize(size: string): string {
+  const parsed = parsePixelSize(size);
+  if (!parsed) return size;
+  return `${parsed.width}*${parsed.height}`;
+}
+
+function normalizePresetSize(size: string): string {
+  return size.trim().toUpperCase();
+}
+
+function getSeedreamSize(model: string, args: CliArgs): string | null {
+  if (args.size) return args.size;
+  if (isSeedream45Model(model) || isSeedream5LiteModel(model)) return "2K";
+  return null;
+}
+
+function getWanSize(args: CliArgs): string | null {
+  if (args.size) return args.size;
+  if (args.quality === "normal") return "1K";
+  if (args.quality === "2k") return "2K";
+  return null;
+}
+
+function buildNanoBananaInput(prompt: string, args: CliArgs, referenceImages: string[]): Record<string, unknown> {
  const input: Record<string, unknown> = { prompt };

  if (args.aspectRatio) {
@ -59,11 +121,165 @@ export function buildInput(prompt: string, args: CliArgs, referenceImages: strin
  return input;
 }

+function buildSeedreamInput(
+  prompt: string,
+  model: string,
+  args: CliArgs,
+  referenceImages: string[],
+): Record<string, unknown> {
+  const input: Record<string, unknown> = { prompt };
+  const requestedSize = getSeedreamSize(model, args);
+
+  if (requestedSize) {
+    if (isSeedream45Model(model)) {
+      const customSize = parsePixelSize(requestedSize);
+      if (customSize) {
+        input.size = "custom";
+        input.width = customSize.width;
+        input.height = customSize.height;
+      } else {
+        input.size = normalizePresetSize(requestedSize);
+      }
+    } else {
+      input.size = normalizePresetSize(requestedSize);
+      input.output_format = "png";
+    }
+  }
+
+  if (args.aspectRatio && input.size !== "custom") {
+    input.aspect_ratio = args.aspectRatio;
+  } else if (!args.aspectRatio && referenceImages.length > 0 && input.size !== "custom") {
+    input.aspect_ratio = "match_input_image";
+  }
+
+  if (referenceImages.length > 0) {
+    input.image_input = referenceImages;
+  }
+
+  return input;
+}
+
+function buildWanInput(prompt: string, args: CliArgs, referenceImages: string[]): Record<string, unknown> {
+  const input: Record<string, unknown> = { prompt };
+  const requestedSize = getWanSize(args);
+
+  if (requestedSize) {
+    input.size = parsePixelSize(requestedSize)
+      ? normalizePixelSize(requestedSize)
+      : normalizePresetSize(requestedSize);
+  }
+
+  if (referenceImages.length > 0) {
+    input.images = referenceImages;
+    input.thinking_mode = false;
+  } else {
+    input.thinking_mode = true;
+  }
+
+  return input;
+}
+
+export function getDefaultOutputExtension(model: string): ".png" | ".jpg" {
+  if (isSeedream45Model(model)) return ".jpg";
+  return ".png";
+}
+
+export function validateArgs(model: string, args: CliArgs): void {
+  if (args.n > 1) {
+    throw new Error(
+      "Replicate --n is not supported yet in baoyu-imagine because this provider currently writes a single output file per request."
+    );
+  }
+
+  if (isSeedream45Model(model)) {
+    const requestedSize = getSeedreamSize(model, args);
+    if (requestedSize) {
+      const customSize = parsePixelSize(requestedSize);
+      if (customSize) {
+        if (
+          customSize.width < SEEDREAM_45_CUSTOM_MIN ||
+          customSize.width > SEEDREAM_45_CUSTOM_MAX ||
+          customSize.height < SEEDREAM_45_CUSTOM_MIN ||
+          customSize.height > SEEDREAM_45_CUSTOM_MAX
+        ) {
+          throw new Error("Seedream 4.5 on Replicate custom --size must keep width and height between 1024 and 4096.");
+        }
+      } else {
+        const normalizedSize = normalizePresetSize(requestedSize);
+        if (!SEEDREAM_45_SIZES.has(normalizedSize)) {
+          throw new Error("Seedream 4.5 on Replicate requires --size to be 2K, 4K, or custom dimensions like 1536x1024.");
+        }
+      }
+    }
+
+    if (args.referenceImages.length > 14) {
+      throw new Error("Seedream 4.5 on Replicate supports at most 14 reference images per request.");
+    }
+  }
+
+  if (isSeedream5LiteModel(model)) {
+    const requestedSize = getSeedreamSize(model, args);
+    if (requestedSize && !SEEDREAM_5_LITE_SIZES.has(normalizePresetSize(requestedSize))) {
+      throw new Error("Seedream 5 lite on Replicate requires --size to be 2K or 3K.");
+    }
+
+    if (args.referenceImages.length > 14) {
+      throw new Error("Seedream 5 lite on Replicate supports at most 14 reference images per request.");
+    }
+  }
+
+  if (isWanModel(model)) {
+    if (args.aspectRatio) {
+      throw new Error("Wan image models on Replicate do not accept --ar. Use --size with a preset like 2K or explicit dimensions like 1920x1080.");
+    }
+
+    if (args.referenceImages.length > 9) {
+      throw new Error("Wan image models on Replicate support at most 9 reference images per request.");
+    }
+
+    const requestedSize = getWanSize(args);
+    if (requestedSize) {
+      const customSize = parsePixelSize(requestedSize);
+      if (!customSize) {
+        const normalizedSize = normalizePresetSize(requestedSize);
+        const allowedSizes = isWanProModel(model) ? WAN_PRO_SIZES : WAN_SIZES;
+        if (!allowedSizes.has(normalizedSize)) {
+          throw new Error(
+            `Wan image models on Replicate require --size to be one of ${Array.from(allowedSizes).join(", ")} or custom dimensions like 1920x1080.`
+          );
+        }
+      }
+    }
+
+    if (args.referenceImages.length > 0 && requestedSize && normalizePresetSize(requestedSize) === "4K") {
+      throw new Error("Wan 2.7 Image Pro on Replicate only supports 4K for text-to-image requests without input images.");
+    }
+  }
+}
+
+export function buildInput(
+  prompt: string,
+  model: string,
+  args: CliArgs,
+  referenceImages: string[],
+): Record<string, unknown> {
+  if (isSeedreamModel(model)) {
+    return buildSeedreamInput(prompt, model, args, referenceImages);
+  }
+
+  if (isWanModel(model)) {
+    return buildWanInput(prompt, args, referenceImages);
+  }
+
+  return buildNanoBananaInput(prompt, args, referenceImages);
+}
+
 async function readImageAsDataUrl(p: string): Promise<string> {
  const buf = await readFile(p);
  const ext = path.extname(p).toLowerCase();
  let mimeType = "image/png";
  if (ext === ".jpg" || ext === ".jpeg") mimeType = "image/jpeg";
+  else if (ext === ".bmp") mimeType = "image/bmp";
  else if (ext === ".gif") mimeType = "image/gif";
  else if (ext === ".webp") mimeType = "image/webp";
  return `data:${mimeType};base64,${buf.toString("base64")}`;
@ -177,6 +393,8 @@ export async function generateImage(
  const apiToken = getApiToken();
  if (!apiToken) throw new Error("REPLICATE_API_TOKEN is required. Get one at https://replicate.com/account/api-tokens");

+  validateArgs(model, args);
+
  const parsedModel = parseModelId(model);

  const refDataUrls: string[] = [];
@ -184,7 +402,7 @@ export async function generateImage(
    refDataUrls.push(await readImageAsDataUrl(refPath));
  }

-  const input = buildInput(prompt, args, refDataUrls);
+  const input = buildInput(prompt, model, args, refDataUrls);

  console.log(`Generating image with Replicate (${model})...`);