GPT-Image-2

GPT-Image-2 is OpenAI's image generation model launched in April 2026. It supports up to 2K resolution, strong in-image text rendering, and multi-image editing. Available through Clauddy's OpenAI-compatible /v1/images/generations endpoint.

Playground (Easiest Way)

No code needed — generate images directly in the Clauddy web UI:

1. Go to the Clauddy Playground (click 操练场 in the left sidebar)
2. Select gpt-image-2 from the model dropdown
3. Type your prompt in the input box at the bottom (e.g. "A cute cat sitting on the moon") and hit send

The generated image will appear directly in the chat area — right-click to save.

---

Quick CLI Test (No Client Install)

Fastest way to verify — one curl + Python base64 decode:

TOKEN="sk-you...oken"   # your Clauddy token

curl -sS https://clauddy.com/v1/images/generations \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "a cute cat sitting on the moon, digital illustration",
    "size": "1024x1024",
    "n": 1
  }' \
  | python3 -c "import sys,json,base64; open('output.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))" \
  && echo "Saved output.png"

You'll get output.png in the current directory. Example output:

About the response format

gpt-image-2 always returns base64 (b64_json field) — response_format=url is not supported. This differs from the DALL·E family; the client must decode it.

---

API Usage (Developers)

Model Info

• Model name: gpt-image-2
• Endpoint: POST https://clauddy.com/v1/images/generations (text-to-image)
• Edit endpoint: POST https://clauddy.com/v1/images/edits (image-to-image, multipart/form-data)
• Response: always b64_json
• Latency: ~30–60s for 1024×1024, up to 1–2 min for 2K. Set client timeouts to ≥ 300 seconds.

Key Parameters

Param	Values	Notes
model	"gpt-image-2"	Required
prompt	string, up to ~32 000 chars	Required, supports English and Chinese
size	"1024x1024" / "1536x1024" / "1024x1536" / "2048x2048" / "auto"	Default auto
quality	"low" / "medium" / "high" / "auto"	Default auto
n	1–10	Images per request
background	"transparent" / "opaque" / "auto"	Transparent requires output_format=png or webp
output_format	"png" / "jpeg" / "webp"	Default png

Python

import base64
from openai import OpenAI

client = OpenAI(
    api_key="sk-you...oken",
    base_url="https://clauddy.com/v1"
)

resp = client.images.generate(
    model="gpt-image-2",
    prompt="a cute cat sitting on the moon, digital illustration",
    size="1024x1024",
    quality="high",
    n=1,
)

with open("output.png", "wb") as f:
    f.write(base64.b64decode(resp.data[0].b64_json))

print("Saved output.png")

Image Edits (multipart)

The /v1/images/edits endpoint accepts one or more input images (up to 16) and edits, restyles, or composes them based on a prompt. Note: the request body is multipart/form-data, not JSON.

One-shot curl edit

TOKEN="sk-you...oken"

curl -sS https://clauddy.com/v1/images/edits \
  -H "Authorization: Bearer $TOKEN" \
  -F "model=gpt-image-2" \
  -F "image[][email protected]" \
  -F 'prompt=Change the cat collar to bright red and add tiny round sunglasses. Keep everything else identical.' \
  -F "size=1024x1024" \
  | python3 -c "import sys,json,base64; open('edited.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))" \
  && echo "Saved edited.png"

Before / after (input: the cat-on-the-moon from the previous section; prompt: add a red collar + tiny round sunglasses):

Before	After

Prompt tips

• Say "keep everything else identical" explicitly, or the model may redraw the whole scene
• To preserve composition, describe "same pose / angle / background"
• For style transfer, use "redraw this image in XXX style"

Python edit

import base64
from openai import OpenAI

client = OpenAI(
    api_key="sk-you...oken",
    base_url="https://clauddy.com/v1"
)

with open("input.png", "rb") as f:
    resp = client.images.edit(
        model="gpt-image-2",
        image=f,
        prompt="Change the cat collar to bright red and add tiny round sunglasses. Keep everything else identical.",
        size="1024x1024",
    )

with open("edited.png", "wb") as f:
    f.write(base64.b64decode(resp.data[0].b64_json))

print("Saved edited.png")

Multi-image composition / style transfer

Pass up to 16 images — useful for "redraw A in the style of B", "combine elements of A and B", etc.:

curl -sS https://clauddy.com/v1/images/edits \
  -H "Authorization: Bearer $TOKEN" \
  -F "model=gpt-image-2" \
  -F "image[][email protected]" \
  -F "image[]=@style_reference.png" \
  -F 'prompt=Redraw the subject from the first image in the watercolor style of the second image' \
  -F "size=1024x1024" \
  | python3 -c "import sys,json,base64; open('fused.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"

Inpainting (mask)

To edit only a specific region, pass a PNG mask (transparent area = region to modify):

curl -sS https://clauddy.com/v1/images/edits \
  -H "Authorization: Bearer $TOKEN" \
  -F "model=gpt-image-2" \
  -F "image[][email protected]" \
  -F "[email protected]" \
  -F 'prompt=Draw a flying parrot in the masked area' \
  -F "size=1024x1024" \
  | python3 -c "import sys,json,base64; open('inpaint.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"

Param	Notes
image[]=@file	Input image (up to 16, repeat the field)
mask=@file	Optional mask; transparent pixels are the editable region
input_fidelity	"high" / "low" — fidelity to source, edits endpoint only
prompt	What to change

---

Recommended Clients

Image-generation endpoint support varies a lot across clients. Ranked by ease:

🥇 Open WebUI — easiest setup

First-class image generation engine. Set:

ENABLE_IMAGE_GENERATION = true
IMAGE_GENERATION_ENGINE = openai
IMAGES_OPENAI_API_BASE_URL = https://clauddy.com/v1
IMAGES_OPENAI_API_KEY = sk-you...oken
IMAGE_GENERATION_MODEL = gpt-image-2
IMAGE_SIZE = 1024x1024

Click the image button in the chat UI to generate. Cleanest experience.

🥈 Cherry Studio — polished desktop UI

Has dedicated size / quality / n pickers. Gotcha: append a literal # to the API host (e.g. https://clauddy.com/v1/images/generations#) to stop Cherry Studio from rewriting the path to /chat/completions. Add the model name gpt-image-2 manually.

🥉 Chatbox — cross-platform, requires path override

Chatbox defaults to /chat/completions, so in a Custom Provider:

1. Choose Custom provider type
2. Set API Path to /v1/images/generations
3. Set timeout to ≥ 360 seconds
4. Set model name to gpt-image-2

Chatbox limitation

Because /v1/images/generations is stateless, you can't do follow-up edits like "now make the collar red". Use the Playground or Cherry Studio for iterative editing.

---

Which Image Model Should I Pick?

Use case	Recommendation
Realistic styles, posters, product shots, in-image text	GPT-Image-2
Creative scenes, artistic styles, character consistency, anime	Nano Banana Pro (Gemini)
Transparent background PNG/WebP	GPT-Image-2 (native background=transparent)
Multi-turn "edit this part" conversation	Nano Banana Pro (chat endpoint, naturally multi-turn)

Both have strengths — try the same prompt through both and compare.