Everything here is free and stays free — the format spec, the nodes, the workflows, the cartridges, the LoRAs. If it saved you a night of debugging (it contains several hundred of mine), tips keep the 5090 warm:
🔁 Liberapay (recurring)
⚡ Or right here: the Civitai tip button on this page sends Buzz directly.
LTX-2.3 Companion
VAE + Vocoder + Connectors
9 GB instead of 43
***Not a model on its own Nothing quantized***
The Problem
If you run an LTX-2.3 GGUF DiT on a 64 GB machine and ComfyUI dies while loading with Windows fatal exception: access violation — this is the fix.
A GGUF carries only the transformer blocks. The loader still needs a checkpoint_path for everything else, and it materializes that WHOLE file in system RAM as owned copies — roughly 60 GB peak against a 43 GB bf16 checkpoint. On a 64 GB box that usually falls over, and it looks like a VRAM problem even though the GPU is idle.
What's Inside — 9.02 GB, cut byte-for-byte
Embedding connectors + adaLN singles
The DiT machinery GGUFs don't carry
4.89 GB
Text embedding projection
Keeps voice and identity conditioning wired up
2.31 GB
Video VAE
1.45 GB
Vocoder
0.26 GB
Audio VAE
0.11 GB
The Result
~11 GB
Peak system RAM
60 GB → 11 GB
Before → After
Output is identical — nothing is quantized, and the config metadata rides along so the loader configures the transformer exactly as the full checkpoint would.
Usage
Put it wherever your checkpoints live. In JoyEcho_ModelLoader: your GGUF in model_file, this file in checkpoint_path. Nothing else changes.
Pairing Note
Cut from the e50 merge — the exact match for the e50 GGUF family. It loads with any LTX-2.3 GGUF, but the connectors and gate tensors differ between merge variants, so for a different checkpoint family, cut your own with the included one-command script (on the HuggingFace repo).
Not a model on its own.
There is no DiT in here — it is the other half of a GGUF.
Also on HuggingFace
joeygambino/ltx23-gguf-vae-companion
Description
FAQ
Comments (2)
@joeygambino Can You share workflow for this?