Josia图像编码

If you've built an img2img graph in ComfyUI, you know the ritual: Load Image → Image Scale → VAE Encode → sampler. Three nodes, two sizing decisions that must agree, and one wire you always forget. JosiaImageEncode (displayed as Josia图像编码) collapses that into a single box that hands you back the image, a mask, and the latent, already at the size you asked for.

What it is, and when you'd reach for it

It's the "turn a picture into something a sampler can eat" node - a drop-in replacement for Load Image on img2img, inpainting prep, or reference-latent workflows. The reason to switch isn't features, it's the graph: instead of a loader, a scaler and an encoder each holding a piece of the sizing, you get one node with one size decision and a latent output you can unplug.

It ships inside ComfyUI_JosiaNodes, a Chinese-language utility pack from the Bilibili creator 吉客AI (JeekAI). Be warned: every widget label and tooltip here is Chinese - 缩放类型, 对齐倍数, 通道切换 - which is the whole reason this pack has essentially zero English-language footprint.

How it works

Loading follows the native Load Image convention: 选择图像 lists files in ComfyUI/input/ with the standard upload button, or you wire an upstream IMAGE instead - the port wins and the dropdown greys out. If the source PNG has alpha, that becomes a mask, and the node separately re-reads the raw alpha off disk so it can put real transparency back into the fourth channel later.

Resizing happens on the numpy/PIL side rather than via torch: the image uses BOX (fastest, and you're about to encode it anyway), the mask uses NEAREST so edges don't smear. Then 对齐倍数 floors width and height to a multiple of 2/4/8/16/32/64 - default 16, which is what almost every VAE latent wants. Read the fine print: alignment only crops or shrinks, never upscales, and 缩放类型 = 关 skips scaling and alignment entirely.

The interesting bit is the latent gate. ComfyUI hands the node the whole prompt graph through hidden inputs, and the node checks whether the Latent output (third slot) is wired to anything. If it isn't, no VAE is loaded and no encode happens - you just get a resized image passthrough. So you can leave it bolted into a workflow permanently: connect Latent for img2img, disconnect it for text-to-image, no node surgery.

The controls that actually matter

缩放类型 is the top switch and the only one you touch most days: 关 (no scaling), 按系数缩放 (by factor), 按长边缩放 (long edge), 按短边缩放 (short edge), 按像素缩放 (megapixels). Whichever you pick, the single parameter row underneath morphs into the matching field - 缩放系数 (default 1.0), 长边尺寸 (1024), 短边尺寸 (768), or 百万像素 (1.0) - so there's no way to set a contradictory pair. The rest:

• 通道切换 - 自动 passes through whatever channel count came in; RGB forces three channels and routes transparency out the 遮罩 port instead; RGBA forces four and fills alpha from the mask when the mask output is connected.
• VAE模型 - defaults to 请选择VAE模型 (don't encode). 使用Josia模型加载VAE reuses the VAE already loaded by the pack's checkpoint node; otherwise pick one from models/vae. Wiring the VAE port greys this out.
• 选择图像 - pick or upload; greys out when the image port is wired.

Outputs are 图像 (IMAGE), 遮罩 (MASK) and Latent (LATENT). In practice you wire Latent into the sampler's latent_image (it replaces VAE Encode) and 遮罩 into whatever your inpainting setup wants. Only the first three channels ever go to the VAE - alpha isn't encoded.

Install

There are no extra dependencies. The pack's requirements.txt literally says so, and pyproject.toml declares an empty dependency list - everything is ComfyUI's bundled torch/numpy/Pillow. Nothing to download either, beyond the VAE you're expected to already have.

cd ComfyUI/custom_nodes
git clone https://github.com/Josia-doit/ComfyUI_JosiaNodes
# restart ComfyUI

Success looks like [JosiaNodes] ✅ JosiaNodes 加载成功，注册节点数：18 in the console. ComfyUI Manager will find it by pack title, but the README is upfront that the pack ships mainly via git and hasn't guaranteed passing the registry's automated safety scan (older versions were flagged over a folder-picker routine in a different node). If Manager hides it, clone manually. And when you update, delete the pack folder rather than unzipping over it - stale filenames are a known source of duplicate frontend extensions.

Where people get burned

A black frame with no error. If neither the port nor the dropdown has an image, the node doesn't complain - it outputs a black image at the target size plus an all-opaque mask. An empty file slot looks like a result. Check the node preview before you debug anything else.

Transparent PNGs look black. Preview only draws RGB, and in PNGs the RGB under a fully transparent pixel is usually 0 - so it renders as a black square. This node paints those pixels white for the preview only; the real alpha still rides in channel four and on the 遮罩 port (bright = transparent, matching native Merge Image Alpha). Your pixels are fine.

"Latent 输出已连接，但没有可用的 VAE". You wired the latent output but left VAE模型 on the placeholder and the VAE port empty. Pick a VAE, wire one, or let the pack's checkpoint node supply it - and make sure it's the autoencoder your checkpoint was trained against. A mismatched VAE gives noise or flat colour, not a subtle shift.

The size isn't what you typed. Alignment rounds down. 缩放类型 = 关 disables alignment too, so if you want your input pixels back untouched, that's the setting.