高清放大

Upscaling has three very different jobs - interpolate pixels, restore detail, or re-render at higher res - and the local community has built a whole zoo of models and workflows for them (ESRGAN variants, tiled diffusion, ControlNet-Tile, you name it). This node skips all of it: it's a cloud upscaler that takes your image and returns it up to 6x larger, with detail added, not just stretched. The README's claim - "preserves detail while improving sharpness" - is the signature of a generative upscaler rather than a bicubic interpolation, and it's the right tool for a product hero image that needs to survive print or a 4K catalog page.

How it works

Standard wrapper: your image_input is base64'd and POSTed to the gateway under the model id image_upscale, with the multiple factor. The server runs its restoration upscaler and returns the enlarged image, which the node converts back to an IMAGE tensor. The batch handling is identical to the pack's other tool nodes - loop over every image in the input, one call each. If you've ever juggled Real-ESRGAN weights and tile sizes on a 6GB card, the appeal here is obvious: no model files, no VRAM math, one dial.

The inputs that matter

• image_input (IMAGE) - your source, batch-capable.
• multiple (x2 / x4 / x6) - the whole control surface. x2 is the safe default; x6 is where detail starts to get invented rather than recovered.
• seed (INT) - reproducibility.

One output, output (IMAGE).

Installing it

Part of comfyui-MJAPI-party. ComfyUI Manager → "mojieapi_party" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MoJIeAIGC/comfyui-MJAPI-party.git

Register at mojieaigc.com, set your key in config.ini (leave BASE_URL = https://www.mojieaigc.com/v1/completions untouched), restart:

[API]
KEY = your_api_key
BASE_URL = https://www.mojieaigc.com/v1/completions

No downloads. This is one of the few nodes in the pack that doesn't even need a decent GPU - the upscale happens on someone else's machine.

Common issues

• 6x on a small source fabricates detail. Generative upscalers add plausible-looking texture, and past ~4x that texture is increasingly the model's imagination. If the client needs a faithful enlargement, stay at x2–x4.
• Cost per image. It's metered per call, and unlike a local upscaler it's not "free after the electricity." For a handful of hero shots that's a rounding error; for batch-upscaling a whole catalog you'd feel it - that's the case where a local ESRGAN stays the right tool.
• Sharpening artifacts on already-clean images. If the source is already crisp, an upscaler that's tuned to add detail can over-sharpen. Feed it the original, not something you've already sharpened in post.
• Error image with text - the pack's standard failure display (key, balance, timeout on big jobs).