> ## Documentation Index > Fetch the complete documentation index at: https://docs.apiyi.com/llms.txt > Use this file to discover all available pages before exploring further. # GPT Image 2 Coze Plugin > Community-contributed Python plugin for the Coze platform that wraps GPT Image 2 calls, error classification, and an OSS upload pipeline via APIYI, so Coze workflows can do text-to-image, image-to-image, and direct result delivery. ## Overview This is a custom Python plugin for the Coze platform (`coze.cn`) that wraps OpenAI's GPT Image 2 model (`gpt-image-2`) through the **APIYI** gateway into a node that Coze workflows can call directly. The plugin ships with complete request construction, error-code classification, content-safety filtering detection, and an Alibaba Cloud OSS upload pipeline. **It returns a publicly accessible URL ready for display**, saving you from building another result-forwarding step in your Coze workflow. **Project Info** * ๐Ÿ“ฆ Distribution: shared as a code package (**not published on GitHub**) * ๐Ÿ‘ค Author: community contribution * ๐ŸŽฏ Target platform: Coze (China / global) custom plugins * ๐Ÿ”Œ Model called: `gpt-image-2` (APIYI, released April 21, 2026) * ๐ŸŒ Gateway: [APIYI](https://api.apiyi.com) โ€” direct access from mainland China, no VPN required * ๐Ÿ“ The full source code is provided in the 'Full Plugin Source Code' section below, ready to copy and use ## About the APIYI Gateway [APIYI](https://api.apiyi.com) is a gateway for GPT Image 2 with direct connectivity from mainland China, offering three routes that share a single API Key: | Domain | Description | | --------------- | ------------- | | `api.apiyi.com` | Default route | | `vip.apiyi.com` | VIP route | | `b.apiyi.com` | Backup route | APIYI offers three ways to access GPT Image 2: | Model ID | Channel | Billing | Generation speed | Highlights | | ----------------- | ----------------- | ----------------- | ---------------- | --------------------------------------------------------------------------- | | `gpt-image-2` | Official relay | Per-token billing | \~120s | Fully compatible with the official OpenAI API, supports quality/size/4K | | `gpt-image-2-all` | Reverse (ChatGPT) | \$0.03/image | 30-60s | Chinese-friendly, called via the Chat endpoint, returns image URLs directly | | `gpt-image-2-vip` | Reverse (Codex) | \$0.03/image | 90-150s | 30 locked size presets, including 4K | > This plugin uses **`gpt-image-2` (official relay)** by default, fully compatible with the official OpenAI API and supporting full parameter control. If you need faster generation, switch to `gpt-image-2-all` mode (see below). Request an API Key (starting with `sk-`) in the [APIYI console](https://api.apiyi.com/token). We recommend setting a daily quota limit (e.g. ยฅ20-50) to keep costs under control. ## Core Features Automatically switches between text-to-image (/v1/images/generations) and editing (/v1/images/edits) modes based on whether fileurls is empty โ€” no need to build two separate nodes in your Coze workflow All requests go through the APIYI gateway (api.apiyi.com) โ€” directly reachable from Chinese networks, with low latency and reliable stability Pass a list of image URLs and the plugin downloads them and injects them into the request as multipart/form-data file uploads โ€” up to 16 reference images (each โ‰ค 50MB), preserving original image details Distinguishes MODERATION\_BLOCKED, INVALID\_API\_KEY, RATE\_LIMIT, SERVER\_ERROR, TIMEOUT, NO\_DATA and other failure causes, making workflow branching easy Distinguishes input-stage moderation\_blocked (400) from output-stage content\_filter (200), and returns a clear rejection message when triggered, avoiding pointless retries The generated base64 image is uploaded straight to Alibaba Cloud OSS โ€” the workflow receives a URL ready to share externally or store Supports quality (low/medium/high/auto), moderation (auto/low), output\_format (png/jpeg/webp) and other parameters, so you can tune the generation strategy as needed ## Supported Models | Model | Model ID | Purpose | API Docs | | ---------------------------- | ----------------- | -------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | GPT Image 2 (official relay) | `gpt-image-2` | Text-to-image, image-to-image (editing), fully compatible with the official OpenAI API | [View docs](/en/api-capabilities/gpt-image-2/overview) | | GPT Image 2-All (reverse) | `gpt-image-2-all` | Text-to-image, image-to-image via the Chat endpoint, Chinese-friendly | [View docs](/en/api-capabilities/gpt-image-2-all/chat-completions) | | GPT Image 2-VIP (reverse) | `gpt-image-2-vip` | Fixed-size generation with 30 size presets including 4K | [View docs](/en/api-capabilities/gpt-image-2-vip/overview) | The plugin uses `gpt-image-2` (official relay) by default, with endpoints `https://api.apiyi.com/v1/images/generations` (text-to-image) and `https://api.apiyi.com/v1/images/edits` (image-to-image), and requires a valid APIYI API Key (starting with `sk-`). To switch routes, change `API_BASE` in the code to `https://vip.apiyi.com/v1` or `https://b.apiyi.com/v1`. ## GPT Image 2 Key Specs | Feature | Description | | ---------------------- | ----------------------------------------------------------------------- | | **Release date** | April 21, 2026 | | **Max resolution** | 3840ร—2160 (4K), total pixels โ‰ค 8,294,400 | | **Aspect ratios** | 1:1 / 16:9 / 9:16 / 4:3 / 3:2 / 3:1 / 1:3 | | **Quality levels** | low / medium / high / auto (default) | | **Output formats** | png (default) / jpeg / webp | | **Output compression** | 0-100 (jpeg / webp only) | | **Background modes** | auto / opaque (โŒ transparent not supported) | | **Moderation level** | auto (default) / low | | **Text rendering** | Accuracy > 99% | | **Images per request** | 1 (`n` only supports 1) | | **Response format** | b64\_json (raw base64, no data:image prefix) | | **input\_fidelity** | Locked to high, **must not be passed** (passing it returns a 400 error) | ## API Endpoints | Endpoint | Method | Content-Type | Purpose | | --------------------------------------------- | ------ | --------------------- | --------------------------------------------------------------------------- | | `https://api.apiyi.com/v1/images/generations` | POST | `application/json` | Text-to-image (generate from a text prompt) | | `https://api.apiyi.com/v1/images/edits` | POST | `multipart/form-data` | Image-to-image (upload reference images via `-F "image[]=@file"`, up to 16) | > To switch routes: `https://vip.apiyi.com/v1/...` or `https://b.apiyi.com/v1/...`. All routes are functionally identical. ## Plugin Architecture GPT Image 2 Coze plugin architecture diagram The plugin's core call chain: ```text theme={null} Coze workflow inputs (cleantext / fileurls / aspect_ratio / resolution / quality / apikey) โ†“ handler() entry point โ†“ Check whether reference images exist (fileurls) โ†“ โ†“ Text-to-image Image-to-image โ†“ โ†“ POST api.apiyi.com/v1/images/generations POST api.apiyi.com/v1/images/edits (application/json) (multipart/form-data) โ†“ โ†“ Parse response / classify error codes โ†“ upload_base64_to_oss() โ€” upload to Alibaba Cloud OSS โ†“ Return { analysis, url, error } ``` ## Resolution and Size Reference The plugin selects the size automatically from `aspect_ratio` and `resolution` (based on APIYI's official presets): | Aspect ratio | 1K (size / pixels) | 2K (size / pixels) | 4K (size / pixels) | | ------------ | ------------------ | ------------------ | ------------------ | | 1:1 | 1024ร—1024 โ‰ˆ 1.0M | 2048ร—2048 โ‰ˆ 4.2M | 3840ร—2160 โ‰ˆ 8.3M | | 16:9 | 1536ร—1024 โ‰ˆ 1.6M | 2048ร—1152 โ‰ˆ 2.4M | 3840ร—2160 โ‰ˆ 8.3M | | 9:16 | 1024ร—1536 โ‰ˆ 1.6M | 1152ร—2048 โ‰ˆ 2.4M | 2160ร—3840 โ‰ˆ 8.3M | | 4:3 | 1024ร—768 โ‰ˆ 0.8M | 2048ร—1536 โ‰ˆ 3.1M | 3264ร—2448 โ‰ˆ 8.0M | | 3:2 | 1536ร—1024 โ‰ˆ 1.6M | 2048ร—1360 โ‰ˆ 2.8M | 3456ร—2304 โ‰ˆ 8.0M | | 3:1 | 1536ร—512 โ‰ˆ 0.8M | 3072ร—1024 โ‰ˆ 3.1M | 3840ร—1280 โ‰ˆ 4.9M | | 1:3 | 512ร—1536 โ‰ˆ 0.8M | 1024ร—3072 โ‰ˆ 3.1M | 1280ร—3840 โ‰ˆ 4.9M | > **Constraints**: every dimension must be divisible by 16, aspect ratio โ‰ค 3:1, total pixels โ‰ค 8,294,400. > > **Note**: 1:1 at 4K outputs 3840ร—2160 (landscape 16:9), not a square โ€” this is an API limitation, and the effective aspect ratio becomes 16:9. Outputs above `2560ร—1440` are still experimental; for production, prefer the preset sizes. ## Input and Output Parameters ### Input (`Input`) | Parameter | Type | Required | Default | Description | | --------------- | --------- | -------- | ------- | -------------------------------------------------------------------------------------------------- | | `cleantext` | string | Yes | โ€” | User text prompt or edit instruction (up to 32,000 characters) | | `fileurls` | string\[] | No | โ€” | List of reference image URLs; leave empty for text-to-image | | `aspect_ratio` | string | Yes | โ€” | Aspect ratio, e.g. `1:1`, `16:9`, `9:16` | | `resolution` | string | Yes | โ€” | Resolution, must be uppercase: `1K` / `2K` / `4K` | | `quality` | string | No | `auto` | Quality level: `low` / `medium` / `high` / `auto` | | `moderation` | string | No | `auto` | Moderation level: `auto` / `low` (relaxed moderation) | | `output_format` | string | No | `png` | Output format: `png` / `jpeg` / `webp` | | `apikey` | string | Yes | โ€” | APIYI API Key (starts with `sk-`, request one in the [APIYI console](https://api.apiyi.com/token)) | ### Output (`Output`) | Field | Type | Description | | ---------- | -------------- | ---------------------------------------------------- | | `analysis` | string | Status text: `ๅ›พ็‰‡็”ŸๆˆๆˆๅŠŸ` (success) / `ๅ›พ็‰‡็”Ÿๆˆๅคฑ่ดฅ` (failure) | | `url` | string \| null | On success, the public OSS URL | | `error` | string \| null | On failure, a friendly error description | ## Deployment Steps * Request an API Key (starting with `sk-`) in the [APIYI console](https://api.apiyi.com/token); we recommend setting a daily quota limit (e.g. ยฅ20-50) * Create an OSS bucket on Alibaba Cloud and a RAM sub-account with `oss:PutObject` permission on that bucket * Record the `AccessKey ID`, `AccessKey Secret`, `Bucket name`, and `Endpoint` (e.g. `oss-cn-beijing.aliyuncs.com`) 1. Go to Coze workspace โ†’ Plugins โ†’ Plugin marketplace 2. Search for 'GPT Image 2' or 'APIYI' to find this plugin 3. Open the plugin card to review the details, then click 'Add' to install it into your workspace Coze plugin marketplace search Paste the full Python code from the 'Full Plugin Source Code' section below into the Coze IDE, and replace the Alibaba Cloud OSS configuration at the top with your own: ```python theme={null} # APIๆ˜“ ็บฟ่ทฏ้…็ฝฎ๏ผˆๅฏ้€‰๏ผ‰ API_BASE = "https://api.apiyi.com/v1" # ไนŸๅฏๅˆ‡ๆขไธบ: "https://vip.apiyi.com/v1" ๆˆ– "https://b.apiyi.com/v1" # ้˜ฟ้‡Œไบ‘ OSS ้…็ฝฎ ACCESS_KEY_ID = "ไฝ ็š„ AK" ACCESS_KEY_SECRET = "ไฝ ็š„ SK" BUCKET_NAME = "ไฝ ็š„ Bucket ๅ็งฐ" ENDPOINT = "oss-cn-beijing.aliyuncs.com" ``` Configure the Input / Output field types and required flags as shown below, matching the `args.input` fields in the code: Input parameter configuration: Coze plugin basic info Coze plugin input parameter configuration Output parameter configuration: Coze plugin output parameter configuration (part 1) Coze plugin output parameter configuration (part 2) * Fill in test parameters in the Coze IDE (start with `quality=low` + `resolution=1K` + a simple prompt to verify the APIYI pipeline) * Once tests pass, click 'Publish' and drag the plugin into any workflow ## Error Classification Strategy The plugin does not just report `success=True/False` โ€” it classifies the failure cause in the following priority order, so your Coze workflow can branch accordingly: | Priority | Error type | Trigger condition | Recommended handling | | -------- | ----------------------- | ------------------------------------------- | ----------------------------------------------------------------------------------------------------------- | | 1 | `MODERATION_BLOCKED` | HTTP 400 / 403, content-safety block | The prompt or image triggered moderation; rewrite and retry โ€” **do not retry with the original input** | | 2 | `INVALID_API_KEY` | HTTP 401 | Check whether the APIYI API Key is correct or has expired | | 3 | `RATE_LIMIT` | HTTP 429 | Request rate exceeded; try switching routes or lowering concurrency | | 4 | `SERVER_ERROR` | HTTP 500 / 502 / 503 | APIYI / OpenAI server-side failure; back off and retry 2-3 times | | 5 | `BAD_REQUEST` | HTTP 400 (non-moderation) | Check parameters: size validity, or whether input\_fidelity or background:transparent was passed by mistake | | 6 | `TIMEOUT` | Request exceeded the quality-based timeout | Lower quality or resolution and retry | | 7 | `NO_DATA` | `data` in the response is empty | Retry | | 8 | `NO_IMAGE_DATA` | `b64_json` field is empty | Retry | | 9 | `IMAGE_DOWNLOAD_FAILED` | Reference image URL could not be downloaded | Check URL accessibility | | 10 | `EDIT_FAILED` | Edit endpoint returned non-200 | Check reference image format, count (โ‰ค16), and per-image size (โ‰ค50MB) | ### Two-Stage Content Filtering GPT Image 2 uses **two-stage content-safety filtering**, unlike Nano Banana Pro: ```text theme={null} User request โ†“ [Stage 1: Input Filter] โ”œโ”€โ”€ Blocked โ†’ HTTP 400 / 403 (moderation_blocked) โ”‚ โ†‘ Rewriting the prompt fixes it (free, not billed) โ”œโ”€โ”€ Passed โ†“ [Model inference generates the image] (billed at this point) โ†“ [Stage 2: Output Filter] โ”œโ”€โ”€ Blocked โ†’ HTTP 200 but empty b64_json (content_filter, already billed) โ”œโ”€โ”€ Passed โ†“ Image returned (HTTP 200) ``` | Dimension | moderation\_blocked | content\_filter | | ------------- | -------------------------- | ------------------------------------- | | Trigger stage | Input stage | Output stage | | HTTP status | 400 | 200 | | Billed | No | **Yes** (inference already completed) | | How to fix | Rewrite the prompt wording | Redesign the entire scene | ### Common moderation\_blocked Triggers | # | Scenario | Notes | | - | --------------------------------------- | -------------------------------------------- | | 1 | Real-person portraits / celebrity names | Elon Musk, Taylor Swift, etc. | | 2 | Names of living artists | Hayao Miyazaki = blocked, Van Gogh = allowed | | 3 | Copyrighted characters / IP | Spider-Man, Pikachu, Mickey Mouse, etc. | | 4 | Violence / gore / weapon details | Blocked automatically | | 5 | Sexual innuendo / revealing clothing | Descriptors like bikini, tight-fitting, sexy | | 6 | Photorealistic images of children | Near-zero tolerance | | 7 | Hate symbols / political extremism | Blocked automatically | ### APIYI-Specific Errors | Error | Cause | Fix | | ----------------------- | ---------------------------------------------- | ------------------------------------------ | | 401 + `invalid_api_key` | Key is missing the `sk-` prefix or has expired | Copy the full Key from the APIYI console | | 404 Not Found | base\_url is missing the `/v1` suffix | Make sure it is `https://api.apiyi.com/v1` | | 429 + frequent triggers | APIYI rate limit hit | Switch routes and retry | | Connection timeout | DNS / network fluctuation | Try another route and retry | ## Expected Latency by Resolution/Quality | Resolution | Quality | Expected latency | Plugin timeout | | -------------- | ------- | ---------------- | -------------- | | 1K (1024ร—1024) | low | 3-8 s | 180 s | | 1K (1024ร—1024) | medium | 20-40 s | 360 s | | 1K (1024ร—1024) | high | 145-280 s | 900 s | | 2K (2048ร—2048) | medium | 80-120 s | 360 s | | 2K (2048ร—2048) | high | 200-250 s | 900 s | | 4K (3840ร—2160) | medium | 150-200 s | 360 s | | 4K (3840ร—2160) | high | 300-600 s | 900 s | > Recommendation: use `resolution=1K + quality=medium` for day-to-day work (20-40 s per image), and `resolution=4K + quality=high` for final deliverables. > > With `quality=auto` (omitted or set to auto), the plugin applies a uniform 360-second timeout, and the API decides the actual quality level. ## Full Plugin Source Code Below is the complete code of `coze-gptimage2.py`, ready to paste straight into the Coze IDE. **Only the OSS configuration at the top needs to be changed** before use. The plugin source is kept verbatim as contributed; comments and user-facing error strings are in Chinese and can be freely localized for your workflow. ```python coze-gptimage2.py theme={null} from runtime import Args from typings.gptimage2.gptimage2 import Input, Output import requests import base64 import io import oss2 import uuid import re # โ•”โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•— # โ•‘ APIๆ˜“ ็บฟ่ทฏ้…็ฝฎ๏ผˆๆŒ‰้œ€ๅˆ‡ๆข๏ผ‰ โ•‘ # โ•šโ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ• API_BASE = "https://api.apiyi.com/v1" # ไนŸๅฏๅˆ‡ๆขไธบ: "https://vip.apiyi.com/v1" ๆˆ– "https://b.apiyi.com/v1" # โ•”โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•— # โ•‘ ้˜ฟ้‡Œไบ‘ OSS ้…็ฝฎ๏ผˆ่ฏทไฟฎๆ”นไธบไฝ ็š„ๅ€ผ๏ผ‰ โ•‘ # โ•šโ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ• ACCESS_KEY_ID = "" # ๅกซๅ…ฅไฝ ็š„้˜ฟ้‡Œไบ‘ Access Key ID ACCESS_KEY_SECRET = "" # ๅกซๅ…ฅไฝ ็š„้˜ฟ้‡Œไบ‘ Access Key Secret BUCKET_NAME = "" # ๅกซๅ…ฅไฝ ็š„้˜ฟ้‡Œไบ‘ OSS Bucket ๅ็งฐ ENDPOINT = "oss-cn-beijing.aliyuncs.com" # ๅกซๅ…ฅไฝ ็š„ OSS Endpoint # โ•”โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•— # โ•‘ ่ดจ้‡่ถ…ๆ—ถ้…็ฝฎ๏ผˆGPT Image 2 ๅŸบไบŽ quality ๅˆ†็บง๏ผ‰ โ•‘ # โ•šโ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ• TIMEOUT = { "low": 180, # ไฝŽ่ดจ้‡ๅฟซ้€Ÿๅ‡บๅ›พ๏ผˆ3-8 ็ง’ๅฎž้™…่€—ๆ—ถ๏ผ‰ "medium": 360, # ไธญ็ญ‰่ดจ้‡๏ผˆ20-40 ็ง’ๅฎž้™…่€—ๆ—ถ๏ผŒๆŽจ่๏ผ‰ "high": 900, # ้ซ˜่ดจ้‡็ฒพ็ป†ๆธฒๆŸ“๏ผˆ145-280 ็ง’ๅฎž้™…่€—ๆ—ถ๏ผ‰ } # โ•”โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•— # โ•‘ ๅˆ†่พจ็Ž‡ โ†’ ๅฐบๅฏธๆ˜ ๅฐ„่กจ๏ผˆๅฎฝ้ซ˜ๆฏ” ร— ๅˆ†่พจ็Ž‡ โ†’ Wร—H๏ผ‰ โ•‘ # โ•šโ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ• RESOLUTION_SIZES = { "1:1": {"1K": "1024x1024", "2K": "2048x2048", "4K": "3840x2160"}, "16:9": {"1K": "1536x1024", "2K": "2048x1152", "4K": "3840x2160"}, "9:16": {"1K": "1024x1536", "2K": "1152x2048", "4K": "2160x3840"}, "4:3": {"1K": "1024x768", "2K": "2048x1536", "4K": "3264x2448"}, "3:2": {"1K": "1536x1024", "2K": "2048x1360", "4K": "3456x2304"}, "3:1": {"1K": "1536x512", "2K": "3072x1024", "4K": "3840x1280"}, "1:3": {"1K": "512x1536", "2K": "1024x3072", "4K": "1280x3840"}, } # ============================== # OSS ไธŠไผ ๅทฅๅ…ท # ============================== def upload_base64_to_oss(image_base64: str) -> str: """ ๅฐ† base64 ๅ›พ็‰‡ไธŠไผ ๅˆฐ้˜ฟ้‡Œไบ‘ OSS ๅนถ่ฟ”ๅ›žๅ…ฌ็ฝ‘ URL ๆ”ฏๆŒๅธฆ data:image/...;base64, ๅ‰็ผ€ๅ’Œ็บฏ base64 ไธค็งๆƒ…ๅ†ต """ base64_str = re.sub(r"^data:image/[^;]+;base64,", "", image_base64) image_data = base64.b64decode(base64_str) image_io = io.BytesIO(image_data) auth = oss2.Auth(ACCESS_KEY_ID, ACCESS_KEY_SECRET) bucket = oss2.Bucket(auth, ENDPOINT, BUCKET_NAME) object_name = f"coze/gptimage2_{uuid.uuid4().hex}.png" bucket.put_object(object_name, image_io) return f"https://{BUCKET_NAME}.{ENDPOINT}/{object_name}" # ============================== # ๅทฅๅ…ทๅ‡ฝๆ•ฐ # ============================== def get_size(aspect_ratio: str, resolution: str) -> str: """ๆ นๆฎๅฎฝ้ซ˜ๆฏ”ๅ’Œๅˆ†่พจ็Ž‡่Žทๅ–ๆŽจ่ๅฐบๅฏธ""" ratio_map = RESOLUTION_SIZES.get(aspect_ratio, RESOLUTION_SIZES["1:1"]) return ratio_map.get(resolution, ratio_map.get("1K", "1024x1024")) def guess_mime_from_url(url: str) -> str: """ๆ นๆฎ URL ๅŽ็ผ€็Œœๆต‹ MIME ็ฑปๅž‹""" url_lower = url.lower() if url_lower.endswith(".png"): return "image/png" if url_lower.endswith(".jpg") or url_lower.endswith(".jpeg"): return "image/jpeg" if url_lower.endswith(".webp"): return "image/webp" if url_lower.endswith(".gif"): return "image/gif" return "image/png" # ============================== # ๆ ธๅฟƒ๏ผšGPT Image 2 ็”Ÿๅ›พ / ็ผ–่พ‘ # ============================== def generate_image(prompt: str, aspect_ratio: str, resolution: str, quality: str, apikey: str, output_format: str = "png", moderation: str = "auto", image_urls=None): """ GPT Image 2 ๆ–‡็”Ÿๅ›พ / ๅ›พ็”Ÿๅ›พๆ ธๅฟƒๅ‡ฝๆ•ฐ - image_urls ไธบ็ฉบ๏ผš็บฏๆ–‡็”Ÿๅ›พ โ†’ APIๆ˜“ /v1/images/generations๏ผˆJSON๏ผ‰ - image_urls ไธไธบ็ฉบ๏ผšๅ‚่€ƒๅ›พ็ผ–่พ‘ โ†’ APIๆ˜“ /v1/images/edits๏ผˆmultipart/form-data๏ผ‰ """ size = get_size(aspect_ratio, resolution) headers = { "Authorization": f"Bearer {apikey}", "Content-Type": "application/json" } # โ”€โ”€ ๅˆ†ๆ”ฏ 1๏ผšๆœ‰ๅ‚่€ƒๅ›พ โ†’ ๅ›พ็”Ÿๅ›พ๏ผˆ็ผ–่พ‘๏ผ‰ โ”€โ”€ if image_urls: return _generate_edit(prompt, size, quality, apikey, output_format, moderation, image_urls, headers) # โ”€โ”€ ๅˆ†ๆ”ฏ 2๏ผšๆ— ๅ‚่€ƒๅ›พ โ†’ ๆ–‡็”Ÿๅ›พ๏ผˆ/v1/images/generations๏ผŒJSON๏ผ‰โ”€โ”€ payload = { "model": "gpt-image-2", "prompt": prompt, "size": size, } if quality and quality != "auto": payload["quality"] = quality if output_format and output_format != "png": payload["output_format"] = output_format if moderation and moderation != "auto": payload["moderation"] = moderation timeout_seconds = TIMEOUT.get(quality, 360) api_url = f"{API_BASE}/images/generations" try: response = requests.post( api_url, headers=headers, json=payload, timeout=timeout_seconds ) # โ”€โ”€ HTTP ้”™่ฏฏๅˆ†ๅ‘ โ”€โ”€ if response.status_code in (400, 403): try: err_body = response.json() err = err_body.get("error", {}) err_msg = err.get("message", "") except Exception: err_msg = response.text if "moderation" in err_msg.lower() or response.status_code == 403: return { "success": False, "errorType": "MODERATION_BLOCKED", "error": "โŒ ๅ†…ๅฎนๅฎ‰ๅ…จๅฎกๆ ธไธ้€š่ฟ‡\n" "ๆ‚จ็š„ๆ็คบ่ฏ่งฆๅ‘ไบ†ๅ†…ๅฎนๅฎ‰ๅ…จ็ญ–็•ฅ๏ผŒ" "่ฏทไฟฎๆ”นๆ็คบ่ฏๅŽ้‡่ฏ•๏ผˆไธ่ฆ็”จๅŽŸๆ็คบ่ฏ้‡่ฏ•๏ผ‰", } return { "success": False, "errorType": "BAD_REQUEST", "error": f"โŒ ่ฏทๆฑ‚ๅ‚ๆ•ฐ้”™่ฏฏ: {err_msg[:500]}\n" "ๅธธ่งๅŽŸๅ› ๏ผš่ฏฏไผ ไบ† input_fidelity ๆˆ– background:transparent", } if response.status_code == 401: return { "success": False, "errorType": "INVALID_API_KEY", "error": "โŒ API Key ๆ— ๆ•ˆ\n่ฏทๆฃ€ๆŸฅๆ‚จ็š„ APIๆ˜“ API ๅฏ†้’ฅๆ˜ฏๅฆๆญฃ็กฎ๏ผŒ" "ๆˆ–ๆ˜ฏๅฆๅทฒ่ฟ‡ๆœŸใ€‚ๅฏๅœจ https://api.apiyi.com/token ๆŸฅ็œ‹", } if response.status_code == 429: return { "success": False, "errorType": "RATE_LIMIT", "error": "โŒ ่ฏทๆฑ‚้ข‘็Ž‡่ถ…้™\nAPI ่ฐƒ็”จ่ฟ‡ไบŽ้ข‘็น๏ผŒ" "ๅฏๅฐ่ฏ•ๅˆ‡ๆข็บฟ่ทฏๆˆ–้™ไฝŽๅนถๅ‘", } if response.status_code in (500, 502, 503): return { "success": False, "errorType": "SERVER_ERROR", "error": f"โŒ ๆœๅŠก็ซฏๆ•…้šœ๏ผˆHTTP {response.status_code}๏ผ‰๏ผŒ" "่ฏท็จๅŽ้‡่ฏ•ๆˆ–ๅฐ่ฏ•ๅˆ‡ๆข็บฟ่ทฏ", } if response.status_code != 200: return { "success": False, "errorType": "HTTP_ERROR", "error": f"HTTP {response.status_code}: " f"{(response.text or '')[:500]}", } # โ”€โ”€ JSON ่งฃๆž โ”€โ”€ try: data = response.json() except ValueError: return { "success": False, "errorType": "INVALID_JSON", "error": "ๅ“ๅบ”ไธๆ˜ฏๆœ‰ๆ•ˆ JSON", } images = data.get("data", []) if not isinstance(images, list) or len(images) == 0: return { "success": False, "errorType": "NO_DATA", "error": "็”Ÿๆˆๅคฑ่ดฅ๏ผšๆœช่ฟ”ๅ›žๅ›พ็‰‡ๆ•ฐๆฎ๏ผˆๅฏ่ƒฝ่งฆๅ‘ไบ† output filter๏ผ‰", "response": data, } # โ”€โ”€ ๆๅ– b64_json๏ผˆAPIๆ˜“ ่ฟ”ๅ›ž็บฏ base64๏ผŒๆ—  data:image ๅ‰็ผ€๏ผ‰โ”€โ”€ image_b64 = images[0].get("b64_json", "") if not image_b64: return { "success": False, "errorType": "NO_IMAGE_DATA", "error": "็”Ÿๆˆๅคฑ่ดฅ๏ผšb64_json ไธบ็ฉบ๏ผˆๅฏ่ƒฝ่ขซ content_filter ่ฟ‡ๆปค๏ผ‰", "response": data, } return {"success": True, "image_data": image_b64} except requests.exceptions.Timeout: return { "success": False, "errorType": "TIMEOUT", "error": f"ๅ›พ็‰‡็”Ÿๆˆ่ฏทๆฑ‚่ถ…ๆ—ถ" f"๏ผˆ่ถ…่ฟ‡ {timeout_seconds} ็ง’๏ผŒ" f"ๅฝ“ๅ‰ quality={quality}๏ผ‰\n" f"ๅปบ่ฎฎ้™ไฝŽ quality ๆˆ– resolution ้‡่ฏ•", } except Exception as e: return { "success": False, "errorType": "EXCEPTION", "error": f"ๅ›พ็‰‡็”Ÿๆˆ่ฏทๆฑ‚ๅคฑ่ดฅ: {str(e)}", } def _generate_edit(prompt: str, size: str, quality: str, apikey: str, output_format: str, moderation: str, image_urls: list, headers: dict): """ GPT Image 2 ๅ›พ็”Ÿๅ›พ๏ผˆ็ผ–่พ‘๏ผ‰ๅญๅ‡ฝๆ•ฐ ่ฐƒ็”จ APIๆ˜“ /v1/images/edits ็ซฏ็‚น๏ผˆmultipart/form-data ๆ–นๅผไธŠไผ ๅ‚่€ƒๅ›พ๏ผ‰ ๆณจๆ„๏ผšAPIๆ˜“ ็š„ /v1/images/edits ่ฆๆฑ‚ Content-Type: multipart/form-data๏ผŒ ้€š่ฟ‡ -F "image[]=@file" ๆ–นๅผไผ ๅ›พ๏ผŒไธๆ”ฏๆŒ JSON base64 data URIใ€‚ ๅ‚่€ƒๅ›พๆ•ฐ้‡ๆœ€ๅคš 16 ๅผ ๏ผŒๅ•ๅผ  โ‰ค 50MB๏ผˆๅปบ่ฎฎๅŽ‹ๅˆฐ 1.5MB ไปฅๅ†…๏ผ‰ใ€‚ """ # โ”€โ”€ ไธ‹่ฝฝๅ‚่€ƒๅ›พๅˆฐๅ†…ๅญ˜ โ”€โ”€ image_files = [] for i, url in enumerate(image_urls): try: resp = requests.get(url, timeout=180) if resp.status_code != 200: return { "success": False, "errorType": "IMAGE_DOWNLOAD_FAILED", "error": f"ๅ›พ็‰‡่Žทๅ–ๅคฑ่ดฅ๏ผˆ{url}๏ผ‰HTTP {resp.status_code}", } mime = guess_mime_from_url(url) ext = mime.split("/")[-1] # png / jpeg / webp if ext == "jpeg": ext = "jpg" image_files.append( ("image[]", (f"image{i}.{ext}", io.BytesIO(resp.content), mime)) ) except Exception as e: return { "success": False, "errorType": "IMAGE_DOWNLOAD_FAILED", "error": f"ๅ›พ็‰‡่Žทๅ–ๅคฑ่ดฅ๏ผˆ{url}๏ผ‰: {e}", } # โ”€โ”€ ๆž„้€  multipart/form-data ่ฏทๆฑ‚๏ผˆ-F ๆ–นๅผ๏ผ‰โ”€โ”€ form_data = { "model": "gpt-image-2", "prompt": prompt, } if size: form_data["size"] = size if quality and quality != "auto": form_data["quality"] = quality if output_format and output_format != "png": form_data["output_format"] = output_format if moderation and moderation != "auto": form_data["moderation"] = moderation # multipart/form-data ไธไผ  Content-Type๏ผˆ่ฎฉ requests ่‡ชๅŠจ็”Ÿๆˆ boundary๏ผ‰ auth_headers = { "Authorization": headers["Authorization"], } timeout_seconds = TIMEOUT.get(quality, 360) api_url = f"{API_BASE}/images/edits" try: response = requests.post( api_url, headers=auth_headers, data=form_data, files=image_files, timeout=timeout_seconds ) # โ”€โ”€ ้”™่ฏฏๅค„็† โ”€โ”€ if response.status_code in (400, 403): try: err = response.json().get("error", {}) err_msg = err.get("message", "") except Exception: err_msg = response.text[:500] if "moderation" in err_msg.lower() or response.status_code == 403: return { "success": False, "errorType": "MODERATION_BLOCKED", "error": "โŒ ๅ†…ๅฎนๅฎ‰ๅ…จๅฎกๆ ธไธ้€š่ฟ‡", } return { "success": False, "errorType": "EDIT_FAILED", "error": f"ๅ›พ็‰‡็ผ–่พ‘่ฏทๆฑ‚ๅ‚ๆ•ฐ้”™่ฏฏ: {err_msg[:500]}\n" "ๅธธ่งๅŽŸๅ› ๏ผš่ฏฏไผ ไบ† input_fidelityใ€background:transparentใ€" "่ถ…่ฟ‡ 16 ๅผ ๅ‚่€ƒๅ›พๆˆ–ๅ•ๅผ ่ถ…่ฟ‡ 50MB", } if response.status_code == 401: return { "success": False, "errorType": "INVALID_API_KEY", "error": "โŒ API Key ๆ— ๆ•ˆ", } if response.status_code == 429: return { "success": False, "errorType": "RATE_LIMIT", "error": "โŒ ่ฏทๆฑ‚้ข‘็Ž‡่ถ…้™๏ผŒๅฏๅฐ่ฏ•ๅˆ‡ๆข็บฟ่ทฏ้‡่ฏ•", } if response.status_code in (500, 502, 503): return { "success": False, "errorType": "SERVER_ERROR", "error": f"โŒ ๆœๅŠก็ซฏๆ•…้šœ๏ผˆHTTP {response.status_code}๏ผ‰", } if response.status_code != 200: try: err = response.json().get("error", {}).get("message", "") except Exception: err = response.text[:500] return { "success": False, "errorType": "EDIT_FAILED", "error": f"ๅ›พ็‰‡็ผ–่พ‘ๅคฑ่ดฅ๏ผˆHTTP {response.status_code}๏ผ‰: {err}", } # โ”€โ”€ ๆๅ–ๅ›พ็‰‡๏ผˆb64_json ๆ˜ฏ็บฏ base64๏ผŒๆ— ๅ‰็ผ€๏ผ‰โ”€โ”€ data = response.json() images = data.get("data", []) if not images: return { "success": False, "errorType": "NO_DATA", "error": "็ผ–่พ‘็ป“ๆžœไธบ็ฉบ", } image_b64 = images[0].get("b64_json", "") if not image_b64: return { "success": False, "errorType": "NO_IMAGE_DATA", "error": "็ผ–่พ‘็ป“ๆžœๅ›พ็‰‡ๆ•ฐๆฎไธบ็ฉบ", } return {"success": True, "image_data": image_b64} except requests.exceptions.Timeout: return { "success": False, "errorType": "TIMEOUT", "error": f"ๅ›พ็‰‡็ผ–่พ‘่ฏทๆฑ‚่ถ…ๆ—ถ๏ผˆ่ถ…่ฟ‡ {timeout_seconds} ็ง’๏ผ‰", } except Exception as e: return { "success": False, "errorType": "EXCEPTION", "error": f"ๅ›พ็‰‡็ผ–่พ‘่ฏทๆฑ‚ๅคฑ่ดฅ: {str(e)}", } # ============================== # Coze Node ๅ…ฅๅฃ # ============================== def handler(args: Args[Input]) -> Output: """ Coze / GPT Image 2๏ผˆAPIๆ˜“ ไปฃ็†๏ผ‰่Š‚็‚นๅ…ฅๅฃ - args.input.cleantext: ็”จๆˆทๆ–‡ๅญ—ๆ็คบ่ฏ - args.input.fileurls: ๅ‚่€ƒๅ›พ URL ๅˆ—่กจ๏ผˆ็”จไบŽๅ›พ็”Ÿๅ›พ๏ผ‰ - args.input.aspect_ratio: ๅฎฝ้ซ˜ๆฏ”๏ผŒๅฆ‚ "1:1" / "16:9" / "9:16" - args.input.resolution: ๅˆ†่พจ็Ž‡๏ผŒๅฆ‚ "1K" / "2K" / "4K" - args.input.quality: ่ดจ้‡็ญ‰็บง๏ผŒๅฆ‚ "low" / "medium" / "high"๏ผˆ้ป˜่ฎค auto๏ผ‰ - args.input.moderation: ๅฎกๆ ธๅผบๅบฆ๏ผŒๅฆ‚ "auto" / "low"๏ผˆ้ป˜่ฎค auto๏ผ‰ - args.input.output_format: ่พ“ๅ‡บๆ ผๅผ๏ผŒๅฆ‚ "png" / "jpeg" / "webp"๏ผˆ้ป˜่ฎค png๏ผ‰ - args.input.apikey: APIๆ˜“ API Key๏ผˆsk-ๅผ€ๅคด๏ผ‰ """ API_KEY = args.input.apikey cleantext = args.input.cleantext or "" fileurls = args.input.fileurls or [] aspect_ratio = args.input.aspect_ratio or "1:1" resolution = args.input.resolution or "1K" quality = getattr(args.input, 'quality', None) or "auto" output_format = getattr(args.input, 'output_format', None) or "png" moderation = getattr(args.input, 'moderation', None) or "auto" prompt = cleantext.strip() if not prompt: prompt = "ๆ นๆฎๅ‚่€ƒๅ›พ็‰‡่ฟ›่กŒๅˆ็†็š„็ผ–่พ‘ไธŽไผ˜ๅŒ–ใ€‚" # ่ฐƒ็”จ GPT Image 2 ็”Ÿๅ›พ / ็ผ–่พ‘ result = generate_image( prompt=prompt, aspect_ratio=aspect_ratio, resolution=resolution, quality=quality, apikey=API_KEY, output_format=output_format, moderation=moderation, image_urls=fileurls if fileurls else None ) if result["success"]: image_base64 = result["image_data"] oss_url = upload_base64_to_oss(image_base64) return { "analysis": "ๅ›พ็‰‡็”ŸๆˆๆˆๅŠŸ", "url": oss_url, "error": None, } else: return { "analysis": "ๅ›พ็‰‡็”Ÿๆˆๅคฑ่ดฅ", "url": None, "error": result.get("error", "ๆœช็Ÿฅ้”™่ฏฏ"), } ``` ## Optional: gpt-image-2-all Fast Mode If you need **faster generation (30-60s) and do not care about size parameter control**, you can switch the plugin to APIYI's `gpt-image-2-all` (reverse edition), called via the Chat Completions endpoint. This mode costs \$0.03/image and returns image URLs directly, with no base64 parsing needed. The core change (just replace the `generate_image` function): ```python theme={null} def generate_image_chat(prompt: str, apikey: str, image_urls=None): """ gpt-image-2-all ๅฟซ้€Ÿๆจกๅผ๏ผˆ้€š่ฟ‡ APIๆ˜“ Chat Completions ็ซฏ็‚น๏ผ‰ ไปทๆ ผ $0.03/ๅผ ๏ผŒๅ‡บๅ›พ 30-60s๏ผŒๅฐบๅฏธ็”ฑ prompt ๆ่ฟฐ้ฉฑๅŠจ """ api_url = f"{API_BASE}/chat/completions" headers = { "Authorization": f"Bearer {apikey}", "Content-Type": "application/json" } # ๆž„้€ ๆถˆๆฏ if image_urls: # ๅ›พ็”Ÿๅ›พ๏ผšๅคšๆจกๆ€ message content = [{"type": "text", "text": prompt}] for url in image_urls: content.append({ "type": "image_url", "image_url": {"url": url} }) else: # ๆ–‡็”Ÿๅ›พ๏ผš็บฏๆ–‡ๆœฌ message content = prompt payload = { "model": "gpt-image-2-all", "messages": [{"role": "user", "content": content}], } try: response = requests.post(api_url, headers=headers, json=payload, timeout=300) if response.status_code != 200: err = response.json().get("error", {}).get("message", response.text) return {"success": False, "errorType": "API_ERROR", "error": str(err)[:500]} data = response.json() content_text = data["choices"][0]["message"]["content"] # ไปŽ Markdown ![image](url) ไธญๆๅ–ๅ›พ็‰‡ URL match = re.search(r'!\[[^\]]*\]\((.*?)\)', content_text) if not match: return {"success": False, "errorType": "NO_URL", "error": "ๆœชไปŽๅ“ๅบ”ไธญๆๅ–ๅˆฐๅ›พ็‰‡ URL"} image_url = match.group(1) # ๅฆ‚ๆžœๆ˜ฏ base64 data URL๏ผŒ็›ดๆŽฅไฝฟ็”จ if image_url.startswith("data:image/"): return {"success": True, "image_data": image_url.split(",", 1)[1]} # ๅฆ‚ๆžœๆ˜ฏ HTTP URL๏ผŒไธ‹่ฝฝๅ›พ็‰‡ โ†’ ่ฝฌ base64 img_resp = requests.get(image_url, timeout=60) if img_resp.status_code != 200: return {"success": False, "errorType": "DOWNLOAD_FAILED", "error": f"ไธ‹่ฝฝๅ›พ็‰‡ๅคฑ่ดฅ: HTTP {img_resp.status_code}"} image_b64 = base64.b64encode(img_resp.content).decode("utf-8") return {"success": True, "image_data": image_b64} except requests.exceptions.Timeout: return {"success": False, "errorType": "TIMEOUT", "error": "่ฏทๆฑ‚่ถ…ๆ—ถ"} except Exception as e: return {"success": False, "errorType": "EXCEPTION", "error": str(e)} ``` > How to switch: replace `generate_image(...)` in `handler()` with `generate_image_chat(...)`. The only inputs needed are `prompt`, `apikey`, and optionally `fileurls`. ## Using It in a Coze Workflow After publishing the plugin, drag the plugin node into the Coze workflow editor and wire it up as follows: ```text theme={null} Start node (user enters prompt + images) โ†“ Prompt/image splitter (code node that splits the user message into cleantext and fileurls) โ†“ Per-user apikey dispatch (dictionary lookup that maps each user to their APIYI API Key) โ†“ gptimage2 plugin node (this plugin) โ†“ Success / failure branches โ†“ End node (outputs url or error) ``` We recommend pairing this with the [Feishu Base AI image generation solution](/en/scenarios/ecosystem/feishu-bitable-image-shortcut). The combined setup lets ops/design teammates **batch-generate images just by filling in prompts in a Feishu table**, without opening any code. Simply swap the Nano Banana Pro plugin in that solution for this one. ## Comparison with Nano Banana Pro | Dimension | Nano Banana Pro (APIYI) | GPT Image 2 (APIYI) | | ----------------------- | ------------------------------------------------------- | -------------------------------------------- | | Model | `gemini-3-pro-image-preview` | `gpt-image-2` | | Gateway | APIYI (same platform, same Key) | APIYI (same platform, same Key) | | Text-to-image endpoint | Gemini `generateContent` (JSON) | `/v1/images/generations` (JSON) | | Image-to-image endpoint | Same endpoint + inline\_data (JSON) | `/v1/images/edits` (**multipart/form-data**) | | Resolution scheme | 1K / 2K / 4K (fixed) | Flexible resolution (up to 3840ร—2160) | | Timeout strategy | By resolution (360s / 600s / 1200s) | By quality (180s / 360s / 900s) | | Content filtering | Single-stage (ZERO\_CANDIDATES\_TOKEN + TEXT\_RESPONSE) | Two-stage (HTTP 400/403 + content\_filter) | | Reference image count | Unlimited (inline\_data) | Up to 16 | | Transparent background | โœ… Supported | โŒ Not supported | | Text rendering | Good | Excellent (>99%) | | Output compression | Not supported | โœ… jpeg/webp compression | | Moderation control | Not supported | โœ… moderation parameter (auto/low) | ## FAQ Yes. The 'Full Plugin Source Code' section of this page contains the complete `coze-gptimage2.py`. **Just change the OSS configuration and API\_BASE at the top** and paste it straight into the Coze IDE โ€” nothing else to request. If you also need: * The Feishu field shortcut code โ†’ see the 'Full Feishu Field Shortcut Source Code' section in the [Feishu Base AI image generation solution](/en/scenarios/ecosystem/feishu-bitable-image-shortcut) * The Nano Banana Pro plugin โ†’ see the [Nano Banana Pro Coze plugin](/en/scenarios/ecosystem/coze-nanobanana-plugin) So each user can be handed a different APIYI API Key. In the Coze workflow, put a 'per-user apikey dispatch' dictionary node in front that matches the caller's name to their API Key โ€” convenient for usage accounting and access control. APIYI is a gateway with direct connectivity from mainland China. Its API Keys also start with `sk-`, but: * Direct access from mainland China, no VPN required * Requested and managed in the [APIYI console](https://api.apiyi.com/token) * Supports daily/monthly quota limits, making cost control easy * One Key works for both Nano Banana Pro and GPT Image 2 All three routes are functionally identical โ€” use any of them with the same API Key: | Domain | Description | | --------------- | ------------- | | `api.apiyi.com` | Default route | | `vip.apiyi.com` | VIP route | | `b.apiyi.com` | Backup route | Just change the `API_BASE` variable in the code to switch. Downstream Coze workflow nodes (especially Feishu field shortcuts) mostly require an **accessible URL** to convert the result into an image attachment. Returning base64 directly makes the data shuttle back and forth through the workflow โ€” poor performance, and the Feishu side cannot render it directly. An OSS link is also convenient for long-term archiving and external sharing. It means the input prompt or reference image triggered content-safety filtering. This error **does not need a retry** โ€” retrying yields the same result. Suggestions: 1. Rewrite the prompt wording 2. Avoid real-person names, copyrighted character names, and names of living artists 3. Avoid sexual innuendo, violence, gore, and other sensitive descriptions This usually means the model completed inference (already billed) but the output was blocked by the content-safety filter (`content_filter`). Suggestions: 1. Redesign the entire visual scene rather than tweaking wording 2. Try a completely different prompt direction 3. Lowering quality can sometimes get past the stricter output filter GPT Image 2's high quality takes 145-280 seconds even at 1K, and 4K can exceed 600 seconds. The plugin already sets a 900-second timeout for high quality. If it still times out: 1. Debug your prompt with `quality=medium` first 2. Check the APIYI console for rate limiting 3. Try switching routes and retrying 4. Reduce the number of concurrent calls 5. Consider the `gpt-image-2-all` mode (30-60s per image) **GPT Image 2 does not support it.** If you need transparent backgrounds, use the [Nano Banana Pro plugin](/en/scenarios/ecosystem/coze-nanobanana-plugin) instead. **No.** The parameter list of APIYI's official-relay gpt-image-2 is not fully identical to OpenAI's; `thinking` is not among the parameters APIYI supports. For fine control over output quality, use the `quality` parameter (low / medium / high / auto) instead. Other unsupported parameters include: * `response_format` โ€” the response always returns `b64_json` * `n` โ€” fixed at 1 * `background: "transparent"` โ€” transparent backgrounds are not supported * `input_fidelity` โ€” locked to high; **passing it returns a 400 error** Both plugins use the **same APIYI platform**, and one API Key works for both. Recommendations: | Scenario | Recommended | | ------------------------------------------------- | ------------------------------------- | | High text-rendering demands (posters, covers, UI) | GPT Image 2 (text accuracy > 99%) | | 4K ultra-high resolution | GPT Image 2 (up to 3840ร—2160) | | Editing with multiple reference images | GPT Image 2 (up to 16) | | Output compression (smaller file sizes) | GPT Image 2 (jpeg / webp compression) | | Transparent backgrounds | Nano Banana Pro | | Budget-sensitive | GPT Image 2-All (\$0.03/image) | | Chinese-friendly prompts | GPT Image 2-All (reverse edition) | ## Related Resources The perfect companion to this plugin: connect the whole Coze workflow to Feishu Base so ops teammates can batch-generate images by filling in a table Another Coze image-generation solution, built on Gemini 3 Pro Image, sharing the same APIYI Key with GPT Image 2 Full documentation, parameter reference and code examples for APIYI's official-relay GPT Image 2 APIYI's reverse-edition Chat Completions endpoint docs (\$0.03/image, 30-60s per image) Manage API keys, check usage and balance, set quota limits moderation\_blocked 400 error diagnosis and mitigation strategies