(deforum) Load Cached String

Same trick as the image cache, one dimension simpler. Deforum animations recompute and shuffle values every frame, and sometimes you need to stash a string - a prompt, a path, a resolved schedule value - and read it back a few frames later. Cache String writes it into a numbered slot; this node, Load Cached String, reads it back out.

It's a dead-simple utility: one input, one output, no models, no files. If you've already met Load Cached Image this is the same idea with the pixels swapped for text.

Input and output

• cache_index - the slot number, 0 to 16. Must match the index you passed to the Cache String write node, obviously.
• STRING - whatever string was stored in that slot, or None if the slot is empty.

That's genuinely the whole node. The cache lives in the pack's shared in-memory store, keyed globally by index rather than by workflow.

How it behaves - and the catch

The storage is global (gs.deforum_cache["string"]), so the value you get back isn't scoped to your workflow. That's fine when a run is linear, and it's the cause of the classic complaint: you change something, re-queue, and an old string pops back out of a slot you thought was fresh. Unlike its image sibling, this read node doesn't bail out on the global reset flag - it just returns whatever is in the slot, stale or not. If you see a prompt or path you didn't write this run, that's why.

Because strings are cheap and the slots are limited, the usual pattern is: reserve a slot per purpose ("slot 0 = current prompt," "slot 1 = output path") and never reuse it for something else. That discipline makes stale data obvious instead of subtle.

Installation

It ships in the Deforum Nodes pack:

cd ComfyUI/custom_nodes
git clone https://github.com/XmYx/deforum-comfy-nodes.git

Restart ComfyUI, or find "Deforum Nodes" by XmYx in ComfyUI Manager. No special dependencies beyond the pack itself - this node is pure plumbing, though in practice you'll be using it inside the full animation workflow with the deforum[comfy] backend installed (Python 3.10 required, per the README).

When you'd actually reach for it

Honestly? Less often than you'd think - most people only need it when a workflow handoff gets awkward, like a node that computes a value once at the start and the loop needs it frame after frame. If you find yourself copying text through the graph, this is the clean way to do it.