🔶 Cache Reloader Latent-> LOAD

This is the latent-space sibling of chaosaiart_reloadIMG_Load, and which one you reach for depends on where in your pipeline you want the frame-to-frame handoff to happen. If you want the previous frame's latent (pre-VAE-decode) carried forward instead of a decoded image, this is the node - useful for chained denoise setups where you don't want the lossy round-trip through pixel space between every single frame, which matters more the more frames you're stacking.

Same underlying mechanism as every Cache Reloader in this pack: a matching chaosaiart_reloadLatent_Save node elsewhere in the graph writes a latent tagged with a reloader slot number at the end of a run, and this node reads it back at the start of the next one - because ComfyUI graphs can't literally loop, so persisting a value across separate executions is how the pack fakes a feedback loop. The restart input, wired from chaosaiart_restarter_advanced, is what tells this node whether there's actually a previous run to pull from.

On the first run there's nothing cached, so you need a starting point. preSave_Image_Mode (widescreen 16:9, portrait 9:16, or square) and preSave_Image_Size (360p / 480p / HD / Full HD) define the blank empty-latent size used on that first pass - pick whatever matches your target output resolution and aspect ratio. If you'd rather start from a specific latent instead of an empty one, the optional preSave_Latent_override does exactly that. reloader is the slot number and has to match the paired Save node's number, same as elsewhere in the pack.

Output: a single LATENT, either the previous frame's cached latent or your empty/override starting latent on the first pass - wire it straight into your sampler's latent input.

Install steps are shared across the whole pack: search "Chaosaiart-Nodes" in ComfyUI Manager, or use "Install via Git URL" pointed at https://github.com/chaosaiart/Chaosaiart-Nodes. Manual route: cd ComfyUI/custom_nodes && git clone https://github.com/chaosaiart/Chaosaiart-Nodes, then restart ComfyUI. On Linux, also run pip install opencv-python and pip install tqdm inside your venv - the pack's Windows installer does this for you automatically, the Linux path doesn't.

Where people get tripped up: pairing this with the wrong reloader number gets you the same silent failure every Cache Reloader node has - you'll just keep getting your fallback empty latent every single run, no error, no animation. The other thing worth knowing: latent sizes are tied to your VAE's downscale factor, so if preSave_Image_Mode/Size doesn't line up with what your checkpoint's VAE expects, you can end up with a latent shape mismatch against your sampler's model - match the size preset to your checkpoint's native resolution, not an arbitrary pick.