🅛🅣🅧 Low VRAM Checkpoint Loader

Low VRAM Checkpoint Loader is the loader you use when a normal "Load Checkpoint" makes LTX blow past your VRAM. LTX-2 is a heavy beast - a 19B (22B on 2.3) diffusion transformer plus a 22GB Gemma text encoder - and if everything tries to land in memory at once, cards under 32GB (and even 32GB cards) fall over. This loader is built to control when things load so peak usage stays under the ceiling.

It's part of ComfyUI-LTXVideo, Lightricks' official pack, and it comes straight out of the pack's low-VRAM loader set the README points you at: the nodes "ensure the correct order of execution and perform the model offloading such that generation fits in 32 GB VRAM." That ordering is the whole trick.

How it works

A standard loader grabs the model and hands back MODEL, CLIP, and VAE. This one does the same but adds a dependencies socket you chain from another loader's output. That connection forces sequential loading - this loader waits for the one before it to finish (and offload) before it claims memory - so the peak never stacks the checkpoint and the text encoder on top of each other. It's the classic offloading pattern, made explicit through a graph edge instead of left to ComfyUI's memory manager to guess.

The inputs and outputs that matter

There's really just one control and one wiring trick:

• ckpt_name - pick your LTX checkpoint from models/checkpoints. If the dropdown is empty, the checkpoint isn't in that folder.
• dependencies (optional) - connect this to a previous loader's output to enforce load order. Leave it unconnected and it behaves like a normal loader; connect it and you get the sequential, low-peak behavior that's the entire reason to use this node.

Outputs are the familiar MODEL, CLIP, and VAE, so it drops into your graph exactly where a stock checkpoint loader would.

Installing it

ComfyUI Manager → Install Custom Nodes → search LTXVideo → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo and restart. It appears under LTXV/loaders. Put your LTX checkpoint in models/checkpoints; on LTX-2 you also need the Gemma encoder in models/text_encoders/. The pack officially asks for a 32GB+ card and 100GB+ of disk, though the community routinely runs quantized LTX on far less.

Where people get burned

• Not chaining dependencies. The node only lowers your peak VRAM when its dependencies input is connected to force ordering. Drop it in without that edge and you get none of the benefit - this is the step people miss.
• Expecting it to beat quantization. Ordering loads helps, but it doesn't shrink the model. If you're still OOM, the bigger levers are quantized weights (GGUF / fp8), ComfyUI's --reserve-vram 4+ flag, and disabling previews - the standard LTX-2 survival kit.
• Thin system RAM. Offloading moves weights to system memory, so 64GB RAM is the comfortable floor for sub-32GB cards. On 32GB you risk page-file swapping, which is slow and wears the disk.
• Ignoring the whole loader set. The README ships a family of low-VRAM loaders meant to work together in the right order. Use the pack's example low-VRAM workflow as your template rather than swapping this single node into a graph that wasn't built for staged loading.