FetchFromRedis

FetchFromRedis is exactly what the name says: it goes to a Redis server, reads whatever JSON is stored under a key, and hands it to your graph as a JSON value. That's the whole node. There's no Redis included with it, no local server that spins up on install, no model to download - you point it at a Redis instance you already run (or one a teammate runs) and it reads.

Why would you want that? Redis is not a thing most people need inside ComfyUI, and this node comes from a pack that clearly drives a real production pipeline, not a hobby setup. You reach for it when state has to live outside the workflow: a second ComfyUI instance, a producer script writing JSON while the graph idles, a queue of detection results waiting to be picked up. The pack even ships the mirror image, SaveToRedis, so you can round-trip data between two workflows on the same server - write the params or results in one, read them in another.

How it works

It's three lines of Python dressed up as a node. It builds a connection from redis.Redis.from_url(REDIS_URL), calls get(key), and json.loads the result. Two details are worth knowing. First, REDIS_URL comes from nodes/config.py, which ships with every value blank - if you haven't edited it, the node fails before it does anything. Second, a missing key doesn't raise; you get an empty {} back. That's friendly on the surface, but it's a silent failure: any node downstream expecting real fields will happily choke on an empty dict.

Inputs and outputs

• key (STRING, required) - the only thing you set. The Redis key to fetch.
• Output: JSON - the parsed value stored under that key.

The node is marked as an output node (OUTPUT_NODE), but the JSON output is the point. Wire it into anything that consumes JSON - parameter holders, the pack's own mask nodes, a workflow loader. Pair it with SaveToRedis (same pack, database_ops category) to close the loop.

Installing it

The pack installs as comfyui-tensorop in ComfyUI Manager - search that title - or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/un-seen/comfyui-tensorops
cd comfyui-tensorops
pip install -r requirements.txt

Restart ComfyUI after. Fair warning: this pack's requirements.txt drags in a lot - transformers for the Florence-2 nodes, boto3, fal-client, replicate - and this node only needs redis. You're installing the whole buffet to eat one dish. The repo's README is empty, so the code is the only documentation.

Where people get burned

• The blank config. Open nodes/config.py and set REDIS_URL (e.g. redis://localhost:6379/0) before the node will connect at all.
• Is Redis even running? ConnectionError means no server is listening. Nothing in the pack starts one for you.
• Manager conflict warnings. This pack names some of its nodes identically to Kijai's ComfyUI-Florence2, so ComfyUI Manager may warn about a conflict when you install or load workflows. Kijai himself flagged it back in 2024 - if you use his Florence-2 nodes, install deliberately and double-check which pack actually owns the node you're dragging in.

If you don't already run Redis, this node is not a reason to start. It's pipeline plumbing for people with a queue to read.