PixelPanda: Edit with Prompt

PixelPanda: Edit with Prompt is the instruction-editing node in the ComfyUI-PixelPanda pack, and the one that actually earns the cloud cost. Where its siblings do one deterministic-feeling thing each, this one lets you type "change the shirt to a leather jacket, keep everything else" and get back a new image where that happened. No mask, no ControlNet spaghetti, no inpainting model to download. It runs on hosted FLUX Kontext Pro and costs 2 credits a shot - the most expensive node in the pack, for the most capable job.

What it is

This is not a local diffusion node. The name is a lie in the good way: it calls the PixelPanda API (/edit under https://pixelpanda.ai/api/v2), sends your image plus your prompt, and waits up to five minutes for the result to come back as a URL, which it downloads and drops into your graph as a normal IMAGE. No weights, no extra VRAM, no sampler settings. The whole class is arbitrary Python that phones home with a key by design, so read the repo before first run - this one is 196 lines of clean MIT code, and the only thing it does is talk to pixelpanda.ai over HTTPS.

If you've been in ComfyUI a while, you know instruction editors took over most of what masked inpainting used to do - change the garment, swap the background, delete an object, all from a sentence. The KB's inpainting doc is blunt about it: the discourse around masks halved because models like Flux Kontext and Qwen-Image-Edit are better at "fix what's wrong" for most cases. This node is that trend, as a hosted one-liner.

The inputs that matter

Only three, and two are boring:

• image - what gets edited.
• api_key - a pk_live_… key, or leave blank to use the PIXELPANDA_API_KEY environment variable.
• prompt - the instruction, plain text. This is the whole craft. "Make it night time" is a full-frame restyle; "remove the coffee cup on the table" is targeted. It reads like any Kontext prompt.
• strength - a float from 0 to 1, default 0.75, step 0.05. This is your denoise dial in disguise: lower keeps the source closer, higher lets the edit really take. Start at the default and nudge down if unrelated pixels start drifting.

Output is a single IMAGE - wire it into whatever comes next, same as any sampler.

How to install it

cd ComfyUI/custom_nodes
git clone https://github.com/RyanKramer/ComfyUI-PixelPanda

Restart ComfyUI and it shows up under the image/PixelPanda menu. Or use ComfyUI Manager and search "PixelPanda". There are no extra requirements - the pack's pyproject.toml declares zero dependencies and it uses only torch, numpy and Pillow, all already in ComfyUI. Grab a key at https://pixelpanda.ai/profile. Credits never expire and a $1.99 starter pack needs no subscription; this node burns 2 per edit.

Where it loses

Everything the inpainting KB says about whole-frame edit models applies. The output is a new image, not a surgical patch: unmasked regions come back close but not identical, and drift compounds if you chain edits. If you need bit-identical pixels outside the edit - say, fixing one sign in a photo you can't regenerate - masked inpainting with crop-and-stitch still owns that, and running Kontext locally on your own GPU is free once you have it. Where this wins is the machine that can't run Kontext at all, and the workflow where one node beats a 40-node graph.

Gotchas

The node processes only the first frame of any batch you feed it (image[0]) - don't drop a video or a batch in expecting N edits. It looks frozen while the API works; that's the 300-second timeout, and the console line [PixelPanda] /edit ok - credits used: … is how you know it finished. No credits (402) or a bad key (401) both fail loudly with a readable error, so the fix is never mysterious.