ComfyGotchi

You already have a pipeline that turns prompts into images. This node turns the images back into a reason to care. ComfyGotchi is a Tamagotchi that lives on your ComfyUI canvas and eats whatever you pipe through it - the README calls it feeding on "AI slop," which is the most accurate description of a KSampler → VAE Decode chain anyone has ever written. It's the least productive custom node you'll ever install, and that's the whole point. It's a pet, not a tool.

What it actually is

Drop the node between VAE Decode and Save Image. Your image passes through untouched - the node's job is to log the feed to a tiny local state server, tick the creature's stats, and occasionally have it say something about your work. Hatch it with ten feeds, and what you've been generating during incubation decides what hatches: keyword detection on the image captions picks from dog, cat, bunny, dragon, robot, monster, alien, phantom, penguin, or the blob default for "you generated something unrecognizable." The aesthetic of those first ten images also stamps a personality tone - dark and moody slop hatches a snarky creature, bright stuff a cheerful one, horror a morbid one.

The mechanics are a nice piece of plumbing. The node talks to ComfyUI's own server over loopback HTTP, posting to /comfygotchi/event and reading /comfygotchi/state, and all the lifecycle logic - hunger, boredom, hygiene, sickness, the poop counter, evolution every 50 feeds, death and ~2-minute ghost reincarnation - lives in state.json in the pack folder. It survives restarts. There's also a real frontend: the node renders its own little Game Boy-style screen with PLAY, CLEAN, and MEDS buttons, and a listener hooks into ComfyUI's Partner API nodes (comfy_extras.nodes_partner and friends), so every time you run a Gemini or Kling or OpenAI node your creature's happiness goes up. It literally feels the love.

The inputs and outputs that matter

There's one required input, one optional pair, and honestly only two things you'll ever touch:

• ai_slop (IMAGE) - your generated image, the meal. Required.
• qwen_model - which Qwen-VL model to use for captions, scanned from models/LLM/Qwen-VL/. Default is none (rule-based), which works fine but gives you blander comments.
• keep_model_loaded (BOOLEAN, default on) - keep the VLM resident in VRAM between calls instead of unloading it.

The comment STRING output is what you wire to a ShowText/Display Text node to read your creature's remarks - the included example workflow does exactly that. The ai_slop IMAGE output is a passthrough, so you can keep chaining downstream.

How it's built, and the honest tradeoffs

With no Qwen model selected, the comment engine is rule-based - the whole pack has zero hard dependencies, which is refreshing. The fun version needs a Qwen3-VL checkpoint in models/LLM/Qwen-VL/ (it loads via HuggingFace transformers), and that's where the tradeoffs live: even a small Qwen-VL is heavy on VRAM next to your diffusion model. keep_model_loaded trades memory for snappy comments; turn it off on a small card and accept slower pings. There's a 25% chance per feed that it actually runs the VLM for a comment, so the creature feels alive without grinding your GPU on every single pass.

Installing

ComfyUI Manager's search should find ComfyGotchi; or clone it manually and restart:

cd ComfyUI/custom_nodes
git clone https://github.com/DenRakEiw/ComfyGotchi

Then restart ComfyUI. No pip install needed for the base behavior.

Troubleshooting

• Only rule-based comments - you haven't put a Qwen-VL model in models/LLM/Qwen-VL/ (a folder with config.json plus safetensors). That's the whole detection logic.
• Your creature died - that's the lifecycle, not a bug. Sickness or hunger hitting 100 turns it into a ghost; it reincarnates as a fresh egg after ~2 minutes.
• Start over - curl -X POST http://127.0.0.1:8188/comfygotchi/reset puts it back in the egg.
• Missing state - state.json lives inside the pack directory; a bad edit can be deleted to regenerate.

One honest note: this is an unknown-author pack with zero installs behind it, and it loads a VLM captioner - the exact shape of node that ecosystem has seen weaponized before. The source here is MIT, dependency-light, and clean on inspection, but ComfyUI custom nodes run arbitrary Python, so skim the repo first if you're cautious.

Is it worth it? Objectively no. You'll spend minutes keeping a digital pet alive that will almost certainly perish because you forgot to feed it. But it also turns a batch of failed gens into a running commentary from a creature that thinks your work is mid. Best $0 I've spent on this hobby.