以圖生圖
參考圖片
以圖生圖，幫你的照片換一個方向
從一張參考照片開始，寫下你希望它變成什麼樣子。用 PixJelly 以圖生圖換掉風格、場景或整體的氛圍。
從參考圖到全新的詮釋
- 1
加入參考圖
上傳一張照片、拖進工作區，或是從剪貼簿貼上。
- 2
寫下方向
說清楚哪些要變、哪些要留，再挑圖片模型和畫面比例。
- 3
生成並存下來
確認新的圖片，下載下來用在你的下一個專案。
同樣的構圖，換一種氛圍
拿一張照片當起點，往新的視覺方向走。換材質、換光線、換風格都可以，你在意的元素依然留得住。
從照片走進更柔軟的世界
這張水彩插畫是用同一張參考照片，加上一句簡短的風格指示生成的。換上你自己的想法，看看你的照片可以變成什麼。
一個安靜的地方，慢慢試
參考圖你帶來，方向你來定。
一張參考圖，很多種想法
從現成的圖片出發，試一個新的場景或一種新的風格。
模型隨你挑
可以用 PixJelly V2.0，也可以用其他開放選擇的圖片模型。
馬上就能下載
在工作區直接把成果存下來。
What else is in the workspace
Transforming a photo is one job. These are the others.
Text to Image
No reference at hand? Write a prompt and generate an image from words alone.
Annotate & Edit
Keep the photo as it is and change one part of it. Draw an arrow at the spot, write the instruction there.
Remove Background
Lift the subject out as a transparent PNG. That tool runs on your own device rather than on a server.
繼續做下去
你下一個專案的起點
創意方向
- 試一種插畫風格
- 換一種光線氛圍看看
商品照
- 描述一個新的背景
- 探索另一種畫面調性
What is image to image AI?
A picture and a written direction go in together. It suits anyone holding a photo that is close to right, but wrong in style, setting or mood. Four things decide what comes back.
Your reference
One PNG, JPEG or WebP up to 25 MB. The workspace prepares a working copy at up to 1,800 pixels on the long edge and sends that copy to the model, so a much larger file buys you nothing here. What the reference does carry is everything a paragraph struggles to describe: the pose, the spacing, the way the light falls.
Your direction
Write what should change and what should stay. Naming the things worth keeping (a pose, a layout, a product shape) usually matters more than describing what is already visible in the frame. The prompt box holds up to 5,000 characters, so there is no need to compress the brief.
Your model
PixJelly V2.0 is our curated workflow, currently running on Nano Banana Pro. Nano Banana 2, GPT Image 2 and Seedream 5 Pro are available too, and each one reads the same reference differently. Trying one brief on two models costs little and tells you a lot.
Your result
The output is a newly generated image at the ratio you picked, not an edit applied on top of your file. Your reference stays untouched, and the result panel holds one image at a time, so download anything you want to keep before the next run.
A new image, not a retouched one
Image to image AI regenerates the whole frame. Everything in the result was drawn by the model, including the parts that look untouched, so lettering, hands, logos and small background objects can come back changed. Compare the result with your reference before you use it anywhere.
That trade decides which tool fits. An image to image AI transformation is the right move when style, setting and mood should shift together. When a single region is wrong and the rest of the photograph is already correct, pointing at that region and describing the change preserves far more of what you started with. When you only need the subject lifted away from its background, an on-device cutout leaves the original pixels alone entirely.
Starting from a picture rather than a sentence is worth it for exactly the reason it is risky. The reference hands the model information no paragraph can carry, and it also hands over the parts you were not thinking about. So choose a reference that is genuinely close to the composition you want, and read the result for drift rather than for style alone.
以圖生圖常見問題
一定要有參考圖嗎？
要。以圖生圖是從你上傳的圖片開始的。如果只想靠文字生出一張新圖，請用文字生成圖片。
原圖的每個細節都會保留嗎？
不會。生成式的轉換有可能改動細節，使用前請先確認結果。想針對某一處精準修改，請改用標註修圖。
這是在我的裝置上處理的嗎？
以圖生圖會把參考圖和提示詞送到你選的圖片服務商。只有去背是在你自己的裝置上完成。