feat: support MiniMax-H3-Image generation and editing

README.md

@@ -86,8 +86,19 @@ cat messages.json | mmx text chat --messages-file - --output json
 mmx image "A cat in a spacesuit"
 mmx image generate --prompt "A cat" --n 3 --aspect-ratio 16:9
 mmx image generate --prompt "Logo" --out-dir ./out/
+
+# MiniMax-H3-Image text-to-image
+mmx image generate --model MiniMax-H3-Image --prompt "An orange cat basking on a windowsill, warm afternoon light, film look" --aspect-ratio 16:9 --area 2K
+mmx image generate --model MiniMax-H3-Image --prompt "Minimal poster" --size 2720x1536 --output-format png --n 2
+
+# MiniMax-H3-Image image editing (1-16 reference images, numbered Picture 1, 2, ... in order)
+mmx image generate --model MiniMax-H3-Image --prompt "Replace the cat in Picture 1 with the Shiba Inu in Picture 2, keep the background" --image cat.jpg --image https://example.com/dog.png
 ```
 
+Without `--image`, MiniMax-H3-Image calls `/v1/images/generations` (text-to-image); with `--image` it calls `/v1/images/edits` (image editing). Choose one way to set the output size: `--size WIDTHxHEIGHT` (area 655,360–20,000,000 px, aspect ratio 1:3–3:1), or `--aspect-ratio` plus `--area 1K|2K|4K` (default 4K). `--no-prompt-enhance` turns off prompt enhancement for text-to-image and then requires `--size` or a concrete `--aspect-ratio`. A failed image (for example one that fails the content safety check) does not stop the others; files keep their original position in the name and failures are reported on stderr (even with `--quiet`). With `--response-format base64`, the total output area (`n` × width × height) is limited to 20,000,000 px for png and 80,000,000 px for jpg/webp; with `size auto` each image counts as its `--area` level (4K = 4096 × 4096). Requests over the limit are rejected locally.
+
+The H3-Image endpoints are not idempotent, so a retry after a timeout generates and bills again. The CLI never retries them automatically and raises the timeout for these requests to at least 600 seconds, matching the server. `--seed`, `--prompt-optimizer`, and `--subject-ref` apply only to image-01.
+
 ### `mmx video`
 
 ```bash

README_CN.md

@@ -81,8 +81,19 @@ cat messages.json | mmx text chat --messages-file - --output json
 mmx image "一只穿宇航服的猫"
 mmx image generate --prompt "科技感 Logo" --n 3 --aspect-ratio 16:9
 mmx image generate --prompt "山水画" --out-dir ./output/
+
+# MiniMax-H3-Image 文生图
+mmx image generate --model MiniMax-H3-Image --prompt "一只橘猫趴在窗台上晒太阳，午后暖光，胶片质感" --aspect-ratio 16:9 --area 2K
+mmx image generate --model MiniMax-H3-Image --prompt "极简海报" --size 2720x1536 --output-format png --n 2
+
+# MiniMax-H3-Image 图像编辑（1~16 张参考图，按顺序对应图1、图2…）
+mmx image generate --model MiniMax-H3-Image --prompt "把图1里的猫换成图2里的柴犬，保持背景不变" --image cat.jpg --image https://example.com/dog.png
 ```
 
+MiniMax-H3-Image 不传 `--image` 时调用 `/v1/images/generations`（文生图），传 `--image` 时调用 `/v1/images/edits`（图像编辑）。输出尺寸二选一：`--size 宽x高`（面积 655,360~20,000,000 像素、宽高比 1:3~3:1），或 `--aspect-ratio` + `--area 1K|2K|4K`（默认 4K）。`--no-prompt-enhance` 可关闭文生图的 Prompt 增强，此时须指定 `--size` 或具体的 `--aspect-ratio`。单张失败（如未通过内容安全检查）不会中断其它图片，文件按原始张序命名，失败原因输出到 stderr（`--quiet` 时也会输出）。使用 `--response-format base64` 时，总输出面积（`n` × 宽 × 高）上限为 png 20,000,000 像素、jpg/webp 80,000,000 像素；`size` 为 auto 时每张按 `--area` 计（4K 按 4096 × 4096），超出会在本地直接报错。
+
+H3-Image 接口不支持幂等，超时重试会重复生成并计费，因此 CLI 不会自动重试，并把该请求的超时下限设为 600 秒（与服务端一致）。`--seed`、`--prompt-optimizer`、`--subject-ref` 仅适用于 image-01。
+
 ### `mmx video`
 
 ```bash

SDK.md

@@ -56,8 +56,31 @@ const result = await sdk.image.generate({
   height: 1024,
   n: 1,
 });
+
+// MiniMax-H3-Image text-to-image → POST /v1/images/generations
+const h3 = await sdk.image.generate({
+  model: 'MiniMax-H3-Image',
+  prompt: 'An orange cat basking on a windowsill',
+  aspect_ratio: '16:9',
+  area: '2K',
+  n: 2,
+});
+// Partial failures return 200; failed items only carry `error`.
+h3.data.forEach((item, i) => console.log(i, item.url ?? item.error?.message));
+
+// MiniMax-H3-Image image editing → POST /v1/images/edits (1-16 images)
+const edited = await sdk.image.generate({
+  model: 'MiniMax-H3-Image',
+  prompt: 'Replace the cat in Picture 1 with the dog in Picture 2',
+  images: [{ image_url: 'https://example.com/cat.jpg' }, { image_url: 'mm_file://123' }],
+});
+
+// Save successful images; the extension follows output_format.
+const paths = await sdk.image.save(edited, { outDir: './out' });
 ```
 
+H3-Image requests are checked locally against the documented limits (size, aspect ratio, `n`, reference images, `b64_json` total area, 64 MiB request body) before sending. Passing `images` (even an empty list) always targets the edits endpoint. All H3-Image requests use a timeout of at least 600 seconds. The endpoints are not idempotent, so the SDK does not retry them. Streaming (`stream: true`) is not supported by the SDK.
+
 ### Video
 
 ```typescript

skill/SKILL.md

@@ -84,7 +84,7 @@ cat conversation.json | mmx text chat --messages-file - --output json
 
 ### image generate
 
-Generate images. Model: `image-01`.
+Generate images. Default model: `image-01`. Use `--model MiniMax-H3-Image` for higher-resolution text-to-image (up to 20 MP) or for image editing with `--image`.
 
 ```bash
 mmx image generate --prompt <text> [flags]
@@ -104,6 +104,27 @@ mmx image generate --prompt <text> [flags]
 | `--response-format <format>` | string | `url` (default) or `base64`. Base64 bypasses CDN download |
 | `--out-dir <dir>` | string | Download images to directory |
 | `--out-prefix <prefix>` | string | Filename prefix (default: `image`) |
+| `--model <model>` | string | `image-01` (default), `image-01-live`, or `MiniMax-H3-Image`; other values go to the image-01 endpoint unchanged |
+
+**MiniMax-H3-Image only** (`--seed`, `--prompt-optimizer`, `--subject-ref` are rejected; `--width/--height` become `--size`):
+
+| Flag | Type | Description |
+|---|---|---|
+| `--image <path-or-url>` | string, repeatable | 1–16 reference images → image editing (`/v1/images/edits`). Order = Picture 1, 2, … |
+| `--size <WxH>` | string | `auto` or e.g. `2720x1536`; area 655,360–20,000,000 px, ratio 1:3–3:1. Not with `--aspect-ratio`/`--area` |
+| `--aspect-ratio <X:Y>` | string | `auto` (default) or X:Y (integers 1–300) between 1:3 and 3:1 |
+| `--area <level>` | string | `1K`, `2K`, `4K` (default) |
+| `--n <count>` | number | 1–10 |
+| `--output-format <fmt>` | string | `jpg` (default), `png`, `webp` |
+| `--output-compression <q>` | number | 0–100 (default 95), jpg/webp only |
+| `--no-prompt-enhance` | boolean | Text-to-image only; needs `--size` or a concrete `--aspect-ratio` |
+
+H3-Image is synchronous and **not idempotent** — do not re-run a command just because it was slow; each run is billed. Partial failures exit 0 with `failed_count`/`failures` in JSON output (failures are also printed to stderr, even with `--quiet`). With `--response-format base64`, the total area `n` × W × H is limited to 20 MP for png and 80 MP for jpg/webp (`size auto` counts the `--area` level, e.g. 4K = 4096×4096).
+
+```bash
+mmx image generate --model MiniMax-H3-Image --prompt "A cat on a windowsill" --aspect-ratio 16:9 --area 2K --output json --quiet
+mmx image generate --model MiniMax-H3-Image --prompt "Put the hat from Picture 2 on the person in Picture 1" --image person.jpg --image hat.png --out result.jpg --quiet
+```
 
 ```bash
 mmx image generate --prompt "A cat in a spacesuit" --output json --quiet

src/client/endpoints.ts

@@ -18,6 +18,14 @@ export function imageEndpoint(baseUrl: string): string {
   return `${baseUrl}/v1/image_generation`;
 }
 
+export function imageGenerationsEndpoint(baseUrl: string): string {
+  return `${baseUrl}/v1/images/generations`;
+}
+
+export function imageEditsEndpoint(baseUrl: string): string {
+  return `${baseUrl}/v1/images/edits`;
+}
+
 export function videoGenerateEndpoint(baseUrl: string): string {
   return `${baseUrl}/v1/video_generation`;
 }