LTX2.3 Latent Upscaler (CRT AutoDL)

LTX-2.3, Lightricks' March 2026 release, is the version that finally made LTX a genuine Wan alternative rather than just the fast-but-rough option - and the headline change was a rebuilt VAE. The new latent space is sharper, especially on hair and fine edges, but it broke things that depended on the old one: TinyVAE preview decoders stopped working overnight because they were built for the previous latent format. That detail matters here, because it means an upscaler built specifically for LTX-2.3's latent space (rather than a generic image-space upscaler you'd apply after decoding) isn't a nice-to-have, it's the only kind that actually understands the format it's working with.

What it does

Zero inputs, one output: a LATENT_UPSCALE_MODEL. Like the pack's other AutoDL nodes, the empty input signature points at this being a self-fetching downloader - the CRT-Nodes README confirms this exact behavior for its LTX 2.3 AutoDownload node (built-in HTTP fetch, plus API routes for checking download status), and this node's naming and category (CRT/AutoDL/LTX2.3) strongly suggest it follows the same pattern to get the upscale model in place without you sourcing it yourself.

This node is one piece of a small internal pipeline the pack ships specifically for LTX 2.3 - alongside an LTX 2.3 Unified Sampler, an LTX 2.3 US Config node, and an LTX 2.3 US Models Pipe node in the same category. It's meant to be used together with those, not on its own; if you're not already running CRT's LTX 2.3 sampler chain, this node has nothing to plug into.

The output that matters

Just model, typed LATENT_UPSCALE_MODEL - wired into whichever node in the CRT LTX 2.3 pipeline expects a latent upscale model, typically downstream of the sampler and before final decode.

How to install it

Part of CRT-Nodes. Via ComfyUI Manager, search CRT-Nodes and install. Manually: git clone https://github.com/PGCRT/CRT-Nodes.git into custom_nodes, pip install -r requirements.txt, restart.

Common issues & troubleshooting

Official VRAM numbers for LTX-2.3 look scary. Lightricks' listed minimum is 32GB+, and it's widely regarded by the community as an overstatement - it describes the unquantized dev model with no offloading, which almost nobody actually runs. If you could run LTX-2 (19B), GGUF-quantized LTX-2.3 (22B) is reported to work at comparable VRAM tiers; treat the official number as a ceiling, not a floor.

Previews go blank or look wrong mid-generation. LTX-2.3's new latent space broke the old TinyVAE preview trick entirely - this is expected behavior across the whole LTX-2.3 ecosystem, not something specific to this upscaler, and there's currently no fix beyond generating without live previews.

Resolution or frame count errors. LTX-2.3 enforces hard constraints - width and height must be divisible by 32, frame count must be divisible by 8 plus 1 - and values outside those fail rather than silently rounding. Check your upstream sampler config before assuming the upscaler is at fault.

Node doesn't load. As with the rest of CRT-Nodes, check the ComfyUI startup log for a pack-wide import error first - this is a niche, late-added corner of a very large single-author pack, and it's more likely to hit a dependency conflict than most of the pack's simpler nodes.