Prepare Image for AI JNK

Vision models bill by the pixel, and this node exists to stop you overpaying. "Prepare Image for AI JNK" sits in front of any multimodal API node - the README pairs it with the pack's own Ask Google Gemini - and compresses whatever you feed it down to something a token budget can afford. If you've ever sent a 4096×4096 render to an LLM and watched the request crawl, this is the thing that fixes that.

The name undersells the one genuinely useful behavior: it's not just a resize, it's a re-encode. The node takes your image tensor, and if its longest edge exceeds max_size (default 512), scales it down proportionally with Lanczos. Then it round-trips the image through a WebP encode/decode at your chosen webp_quality (default 80), which is where the real size savings come from. WebP at quality 80 on a flat diffusion render can be a fraction of the PNG bytes, and for an API that charges per image, per token, or per byte, that's the difference between a cheap batch run and a bill you notice. Images with alpha get flattened onto white first - the node doesn't assume the API can see through transparency.

Here's the part a lot of people miss: the bytes output isn't the pixel dimensions, it's a formatted string of the encoded file size ("34.2 KB"). So you can wire width, height, quality and bytes into a text node or a log and see exactly what you'd have sent before you actually send it. The returned image is the re-encoded, re-decoded result - same node pipeline as the API would receive, which keeps your preview honest. show_preview just toggles the temp-file preview in the UI; the node re-runs every execution (IS_CHANGED is forced to nan), so the byte count stays current even on cached runs.

Two settings matter, and they're the only ones a beginner needs to touch:

• max_size - the long edge target. 512 is a sane default for captioning and Q&A; 1024 if you actually need fine detail for the model to read it.
• webp_quality - 80 is a good balance for photos, but drop it to 50–60 for simple renders and you'll barely notice.

Install: ComfyUI Manager → search "JNK" → Install, then restart. Or:

cd ComfyUI/custom_nodes/
git clone https://github.com/Aljnk/ComfyUI-JNK-Tiny-Nodes.git

Then restart ComfyUI. The pack's requirements.txt lists opencv-python, pygame, and google-genai, but this node only needs Pillow, which ComfyUI already ships. If you installed via Manager and it pulled the extra deps, fine - they're harmless here, they're there for the Gemini and sound nodes.

Gotchas: the output image is RGB with a white background baked in, so don't use this node if you need to preserve transparency downstream. And the bytes value is informational - there's no validation gate; it will happily return a 5 MB image if you set max_size to 4096 and webp_quality to 100. If your goal is genuinely "make it small," leave the defaults alone and resist the urge to crank them.

It's a tiny utility in a grab-bag pack, but it's the one I'd keep if I had to pick. Any workflow that talks to a paid vision API should route its input through this node first - it's the difference between a token bill you ignore and one you screenshot in disbelief.