LTX-2 Extension Module (Disk / Low RAM) ??

Long LTX-2 extensions have a dirty secret that has nothing to do with the GPU: the system RAM. Keep enough frames around as IMAGE tensors - thousands of frames at 1080p in float32 - and your 64GB workstation starts swapping to disk mid-run. That's what the Disk variant of the Extension Module exists for. Instead of stitching IMAGE tensors in memory, it reads and writes PNG frame directories. Same overlap/blend logic as the in-memory module, zero giant tensors in RAM.

This is squarely aimed at the low-RAM LTX crowd - the KB's own guidance is that 64GB system RAM is the practical floor for comfortable LTX-2 work, and below that, every frame you hold in memory is a frame that's about to page. The pack's broader low-RAM story pairs this node with VAEDecodeToDisk (decode one frame at a time to disk) on one side and a disk→latent reader on the other.

How it works

Where the in-memory module takes IMAGE inputs, this one takes directory paths:

• source_dir - where your accumulated frames live (default iamccs_vae_frames/seg0).
• new_dir (optional) - the newly generated frames for this pass.
• output_dir - where the stitched sequence gets written (default iamccs_extension_disk/extended).
• start_dir - where the overlap/start frames for the next pass get written (default iamccs_extension_disk/start).

Everything else mirrors the flagship module: overlap_frames, overlap_side, overlap_mode (cut or one of the crossfades), enable_math/math_operation, safe_mode, start_frames_rule (to keep the extracted start frames at 8n+1), and the same preset list. It processes the frames on disk and returns extended_dir and start_dir so downstream disk-based nodes know where to look.

One implementation detail worth knowing: this node returns IS_CHANGED = float("nan"), meaning ComfyUI always reruns it - by design. It mutates directories on disk, so caching would leave you with stale stitched sequences or stale start-frame folders. If it seems to "rerun every time," that's correct behavior, not a bug.

The inputs and outputs that matter

You set the paths once (they're relative to your ComfyUI output directory by default) and mostly forget them. The values you'll actually touch: overlap_frames (default 9) and preset (start at target_extension_ltx2 or the 24fps presets). Outputs: extended_dir, start_dir, then overlap_frames, calculated_frames, extension_frames INTs and a report.

Installing it

Ships in IAMCCS-nodes. ComfyUI Manager → "IAMCCS" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Restart ComfyUI, under IAMCCS/LTX-2. No extra dependencies - but the workflow around it (VAE-decode-to-disk, disk-to-latent) needs the pack's sibling nodes and the ComfyUI-LTXVideo nodes.

Common issues

The big one is directory collisions: this node always reruns and writes to fixed paths, so running two of them in parallel (or two workflows sharing a source_dir/output_dir) will corrupt each other's state. Give every project its own subdirectory. Second, new_dir is optional - if you leave it empty, the node stitches only what's in source_dir, which is a legitimate workflow (extend the accumulated clip without a fresh pass) but easy to mistake for a bug when you expected new frames to appear. And because it's disk-based, the 8n+1 start_frames_rule matters more than ever - a wrong start-frame count means the next encode fails a minute into the run, and you only find out when the loop stops.