INT8 ConvRot build of the <a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino/joyai-echo-ltx23-echoVid-ltxAud-surgical">echoVid+ltxAud surgical merge</a>, exported with the official <code>comfy-quants</code> <code>int8_tensorwise</code> + ConvRot recipe.<h2 id="what-convrot-is">What ConvRot is</h2>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.<h2 id="files">Files</h2><ul><li><code>ltx23_echoVid-ltxAud_surgical_int8_convrot.safetensors</code> — 27.2 GB. The full single-file bundle: video VAE, audio VAE, vocoder and the four <code>text_embedding_projection</code> tensors are all inside, copied verbatim in bf16. No companion downloads.</li><li><code>ltx23_echoVid-ltxAud_surgical_int8_convrot_transformer_only.safetensors</code> — 25.5 GB. The DiT alone, for pairing with your own full-precision VAEs and text-encoder setup. It deliberately keeps the four tiny <code>text_embedding_projection</code> tensors — conversions that drop them cause the robotic-voice / prompt-reading / new-face-every-shot failure class.</li></ul><h2 id="speed:-this-is-not-automatically-faster">Speed: this is not automatically faster</h2>Worth understanding before you pick it, because it surprises people:<ul><li>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.</li><li>With <code>fp8_scaled_mm</code> 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.</li><li>Under stock ComfyUI, the native W8A8 loader runs the INT8 weights directly, which is where the format's own speed advantage lives.</li></ul>Do not pair the fp8 <code>.safetensors</code> with <code>fp8_scaled_mm</code> — that combination is a hard error, because the toggle downcasts from bf16 itself. bf16 or INT8 are the valid sources.<h2 id="looking-for-something-else">Looking for something else?</h2><ul><li>Full bf16 / fp8 checkpoints are in the base merge listing.</li><li>GGUF quants (Q8_0 / Q5_0 / Q4_0) have their own listing.</li><li>e50, the reference-voice articulation rebuild, has its own listing with every format.</li></ul><h2 id="everything-else-i've-published">Everything else I've published</h2><ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino">All models on Hugging Face</a></li><li><a target="_blank" rel="ugc" href="https://github.com/jlucasmcrell/joyai-echo-multishot-patch">JoyAI-Echo multishot patch</a> — nodes, releases, issues</li><li><a target="_blank" rel="ugc" href="https://github.com/jlucasmcrell/ltx23-multishot-lite">LTX-2.3 Multishot Lite</a> — the core-node workflow</li></ul><h2 id="support-this-work">Support this work</h2>Everything I publish — workflows, merges, quants, patches — is free and stays free. The models are non-commercial by license and I have no plans to change that.If something here saved you an evening:<ul><li><a target="_blank" rel="ugc" href="https://github.com/sponsors/jlucasmcrell">GitHub Sponsors</a> — recurring, no platform cut</li><li><a target="_blank" rel="ugc" href="https://ko-fi.com/joeygambino">Ko-fi</a> — one-off, no account needed</li><li><a target="_blank" rel="ugc" href="https://liberapay.com/joeygambino">Liberapay</a> — recurring, open-source and non-profit</li></ul>Bug reports are worth as much as money. Most of the fixes in these packages exist because somebody posted a log.

Files with SHA256: fd0dfe2556234c678a097de2a07a2c87ac2adec1c8dfa3c5a5c986a14666467e