LTX-2.3 GGUF (Q8_0 / Q5_0 / Q4_0) echoVid + ltxAud Surgical Merge

**CIVITAI IS RENAMING EVERY DOWNLOADED FILE THE SAME THING. MAKE SURE YOU SAVE-AS AND NAME THE FILE SO YOU KNOW WHAT IT IS.**

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:

• ☕ Buy me a coffee on Ko-fi

• 💛 Sponsor on GitHub

• 🔁 Liberapay (recurring)

• ⚡ Or right here: the Civitai tip button on this page sends Buzz directly.

GGUF quantizations of the echoVid+ltxAud surgical merge — JoyAI-Echo's video branch married to LTX-2.3-distilled-1.1's audio branch, for multishot work where one character has to hold their face and voice across shots.

A GGUF here is DiT-only. It contains the transformer and nothing else, so you still need a full .safetensors checkpoint in checkpoint_path to supply the video VAE, audio VAE, vocoder and text connectors. That is the single most common setup mistake with these files.

Files

• LTX23-echoVid-ltxAud-Surgical-DiT-Q8_0.gguf — 23.1 GB, 0.36° angular deviation.
  The default. Take this unless VRAM forces otherwise — at 0.36° it is effectively indistinguishable from bf16 in output.

• LTX23-echoVid-ltxAud-Surgical-DiT-Q5_0.gguf — 15.5 GB, 2.82° deviation.
  For 16 GB cards. A visible but usually acceptable step down.

• LTX23-echoVid-ltxAud-Surgical-DiT-Q4_0.gguf — 13.0 GB, 7.1% spot error.
  12–16 GB cards. The quality cost is visible; take it only if the others will not fit.

“Deviation” is the angular change in what a layer actually outputs, measured against the unquantized bf16 weights — a more honest figure for a quant than file size alone.

Quick fixes

• Errors about missing VAEs
  Why: a GGUF is DiT-only.
  Fix: put a full bf16 checkpoint in checkpoint_path. The GGUF goes in model_file.

• Robotic voice, the model reading your prompt aloud, or a new face every shot
  Why: almost always missing text_embedding_projection tensors.
  Fix: check the tensor count — print(int.from_bytes(open("f.gguf","rb").read(16)[8:],"little")). For a JoyAI-Echo DiT, 4448 is good and 4444 is missing the projection.

• Lip sync breaks about 9.6 s into every shot
  Why: not the weights — the pack's video RoPE clock was hardcoded to 24 fps while audio runs in true seconds, so a 25 fps render drifts ~4%/s.
  Fix: apply the multishot patch (Bug fix #0).

Looking for something else?

• Full .safetensors checkpoints (bf16 / fp8 / TALKINGHEAD) are in the base merge listing.

• e50 — the reference-voice articulation rebuild — has its own listing with every format including its own GGUFs.

• INT8 ConvRot has its own listing.

Everything else I've published

• All models on Hugging Face

• JoyAI-Echo multishot patch — nodes, releases, issues

• LTX-2.3 Multishot Lite — the core-node workflow