Aabao 图像编辑

AabaoImageEdit is the edit half of the pack's Sora-image pair: give it an image and a text instruction, and the Aabao reseller's sora-image model re-renders it per your prompt. The default prompt is telling - "将这张图片变成油画风格" (turn this image into an oil painting). It's a style-and-content rewrite, not a pixel-level retouch: you're asking a generative model to interpret "keep this, change that," which is a different animal from running a local inpaint or a Photoshop pass. Think "consistent redraw" rather than "edit layer."

How it works

The node base64-encodes your input IMAGE, bundles it with the prompt and model into a request to the provider's image editing endpoint, and downloads the result into your output folder. The image and the requested edit both go to the provider's servers - worth remembering if you're feeding it anything sensitive, because that's true of every node in this API-wrapper pack. The model list is the same sora-image family as AabaoImageGen (sora-image, -landscape, -portrait), and size is likewise capped at 360x360, 540x360, 360x540.

Inputs and outputs

• image - required IMAGE input; from Load Image or any upstream generator (including AabaoImageGen).
• prompt - the instruction. This is the whole game. "Turn into oil painting" works; "make it better" doesn't. The model follows style directives best - painting styles, art movements, mood, lighting changes.
• model - pick the variant matching your source image's orientation.
• size - output resolution, again only the three small options.

Only two outputs: image (IMAGE) and status (STRING). Wire image into the next node or straight to a preview. Output node by default, but the IMAGE output plugs into anything.

Install

The whole pack installs the same way - ComfyUI Manager → search "ComfyUI_Sora", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_Sora
cd ComfyUI_Sora && pip install -r requirements.txt

Restart, put aabao_api_key in config.json (or the node field), and it's ready.

Common issues

• "Edit did nothing" - the model faithfully follows the image and treats your prompt as a loose suggestion. For structural changes ("remove the background") you're fighting the model; for style transfer it's usually on point.
• Low-res output - 360/540px is the ceiling here. Use it for fast iteration and storyboard stills; take the result elsewhere if you need print quality.
• Edit changed too much - small resolutions leave the model little room to preserve detail, so strong edits drift. One edit at a time beats one big edit.
• status shows an error - decode the JSON; it's normally key/balance or an unsupported model/size combo.

Reach for this node when you want to restyle a frame in the same ecosystem you're already generating video in. For anything surgical, keep a local inpaint node around - this one is a style brush, not a scalpel.