🔶 Cache Reloader Latent-> SAVE

The write half of the latent cache pair, and the counterpart to chaosaiart_reloadLatent_Load. Drop your sampler's output latent into this node at the end of each pass, and it gets stashed so the Load node can hand it back at the start of the next run - that's the mechanism behind Chaosaiart's frame-by-frame animation trick, applied at the latent level instead of after a full VAE decode to pixels. Since ComfyUI graphs can't actually cycle back on themselves, persisting a value across separate workflow executions is how this pack fakes the loop.

It's an output node, so ComfyUI runs it every pass as a terminal step regardless of whether anything downstream reads its outputs - which matters here because the whole point is that the save has to happen every single time for the chain to keep working. Skip it once and the next run has nothing to load.

Two required fields: latent is what you're caching - wire in your sampler's output. reloader is the slot number (0–100) this gets filed under, and it needs to match whatever number is set on the chaosaiart_reloadLatent_Load node reading it back. If you've got more than one latent cache running in the same workflow, give each pair its own number - they all share the same slot space and a collision means one silently clobbers the other.

Outputs: LATENT, a straight passthrough so you can keep using the value further down the same run without rewiring back to your sampler, and Info, a text confirmation of what got saved - worth glancing at while debugging a loop that doesn't seem to be doing anything.

Install: search "Chaosaiart-Nodes" in ComfyUI Manager, or use "Install via Git URL" with https://github.com/chaosaiart/Chaosaiart-Nodes. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/chaosaiart/Chaosaiart-Nodes, then restart. Linux additionally needs pip install opencv-python and pip install tqdm inside your venv - the Windows install script handles that step for you, Linux doesn't.

The usual failure mode: a reloader number that doesn't match the paired Load node, which gets you a chain that technically runs without errors but never actually carries anything forward - every frame comes out as if it were the first one. It's worth deciding up front, before you build a big graph, whether you actually want the latent-level cache (this node) or the image-level one (chaosaiart_reloadIMG_Save) for a given chain - caching at latent avoids a decode/encode round trip between frames, which is cheaper and loses less detail over a long sequence, but it also means you can't casually inspect the cached frame as a viewable image mid-run the way you can with the IMG pair.