JoyAI-Echo x LTX-2.3 Surgical Merge — INT8 ConvRot (stock ComfyUI)

**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.

INT8 ConvRot build of the echoVid+ltxAud surgical merge, exported with the official comfy-quants int8_tensorwise + ConvRot recipe.

What ConvRot is

Weights are rotated by orthogonal Hadamard matrices in groups of 256 before INT8 encoding. That spreads the row-wise outliers DiT architectures produce across channels, so the quantizer is not forced to waste range on a handful of extreme values. It is a large quality gain over plain row-wise INT8, and the rotation is undone at load.

Layer selection follows the official Lightricks recipe layer-for-layer: transformer blocks 2–45 quantized (1,496 Linears), blocks 0/1/46/47 and both embeddings connectors kept bf16 — the same selection as the official LTX-2.3 fp8 and nvfp4 releases.

Files

• ltx23_echoVid-ltxAud_surgical_int8_convrot.safetensors — 27.2 GB.
  The full single-file bundle: video VAE, audio VAE, vocoder and the four text_embedding_projection tensors are all inside, copied verbatim in bf16. No companion downloads.

• ltx23_echoVid-ltxAud_surgical_int8_convrot_transformer_only.safetensors — 25.5 GB.
  The DiT alone, for pairing with your own full-precision VAEs and text-encoder setup. It deliberately keeps the four tiny text_embedding_projection tensors — conversions that drop them cause the robotic-voice / prompt-reading / new-face-every-shot failure class.

Speed: this is not automatically faster

Worth understanding before you pick it, because it surprises people:

• In the JoyAI-Echo pack, INT8 is reconstructed to bf16 at load — the pack has no INT8 compute path. On its own that means you pay full bf16 memory plus ~40 s of reconstruction, and it runs no faster than bf16.

• With fp8_scaled_mm ON, that stops mattering: the reconstructed bf16 is downcast to fp8 for compute, giving ~22 GB resident and native fp8 matmul — identical to feeding it the bf16 file, at ~60% of the download. This is the configuration to use.

• Under stock ComfyUI, the native W8A8 loader runs the INT8 weights directly, which is where the format's own speed advantage lives.

Do not pair the fp8 .safetensors with fp8_scaled_mm — that combination is a hard error, because the toggle downcasts from bf16 itself. bf16 or INT8 are the valid sources.

Looking for something else?

• Full bf16 / fp8 checkpoints are in the base merge listing.

• GGUF quants (Q8_0 / Q5_0 / Q4_0) have their own listing.

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

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