ComfyGotchi *DRE

ComfyGotchi is exactly what the tooltip says: you feed it AI-generated images and it devours them and grows. It's a virtual pet that lives inside ComfyUI, and the food is your output. Feed it a "slop" image and it captions what it sees, reacts with a comment, and slowly evolves from an egg into a blob, a cat, a dragon, a robot - ten variants, each with its own personality. This is not a serious productivity node and it never pretends to be; it's the pack author's sense of humor shipped as a graph node, and honestly it's kind of a good reward loop for people who generate a hundred images a day anyway.

What it actually does

The node takes one IMAGE input (ai_slop), passes it straight through, and produces a second output: a comment string - a snarky one-liner about what you just fed it. Behind the scenes it does more: every image you feed gets captioned, the caption is recorded in the pet's state file (state_dre.json, kept next to the pack), and the pet's mood, hunger, love and hygiene stats tick over the minutes like any self-respecting virtual pet. Let the egg sit hungry long enough and the state machine has a ghost stage for it. It is a Tamagotchi with diffusion as the food source.

How it works

Captioning is the interesting part. By default it's rule-based - a local heuristic that reads basic image properties and picks from a tone table (snarky, cheerful, robotic, morbid…) mapped from caption keywords. But if you drop a Qwen-VL model into ComfyUI/models/LLM/Qwen-VL/, the node will pick it up automatically (the qwen_model dropdown is populated by scanning that folder) and use real vision-language captioning instead. That's the same VLM-in-the-graph pattern the modidex knowledge base covers under LLM captioning: it's a separate model you bolt onto ComfyUI, not the text encoder inside your checkpoint.

The pet itself is rendered by a bundled web frontend (web/comfygotchi_dre.js) that talks to the pack's own little server, which exposes /comfygotchi_dre/event and /comfygotchi_dre/state endpoints on your ComfyUI port. Don't worry about any of that - it's automatic. If the web part fails to init, the node still runs and just logs warnings.

Inputs that matter

• ai_slop (IMAGE, required) - the only input you have to connect. Wire any sampler output (or VAE-decoded image) into it.
• qwen_model - none (rule-based) by default; select a Qwen-VL folder if you installed one.
• keep_model_loaded - keep the Qwen model resident in VRAM between calls. Turn this off if your card is struggling and the gotchi is hogging it.

Outputs: ai_slop (the image you fed it, unchanged) and comment (STRING - wire it to a text display node or just read it).

Installing it

This ships inside DenRakEiw_Nodes, so you install the pack, not the node. In ComfyUI Manager search "DenRakEiw Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/DenRakEiw/DenRakEiw_Nodes
cd DenRakEiw_Nodes
pip install -r requirements.txt

then restart ComfyUI. No extra models required for the default rule-based mode; Qwen-VL is optional flavor.

Gotchas

The rule-based mode is a toy and can't really see your image - captions will be generic. If you want the gotchi to actually react to what you made, you need the Qwen model. Also, the pet's stats decay in real time, so if you only open ComfyUI once a week your gotchi may have dramatic news for you. And don't expect support threads: this is a small one-author pack with basically zero community footprint, so treat the GitHub issues page as the only help desk.