One Merge, Five Builds

The JoyAI-Echo × LTX-2.3 surgical merge — JoyAI-Echo's video/conditioning branch (cross-shot character memory) married to LTX-2.3-distilled-1.1's audio branch (the voice) — published in five encodings, one for every tier of hardware. Every clip below was rendered by the build it sits under.

bf16 — the reference

43.0 GB · exact weights · the ground truth every number below is measured against

THE WITNESS (bf16 take) — the five-location identity demo on the reference weights. (This take has a brief voice drift in shot two; a replacement render is on the way.) For reference renders on 32 GB+ cards with sequential offload. model repo

fp8 (e4m3) — the fast path

23.4 GB · ~1.7° weight deviation, ~2.3° on the full scaled-mm compute path · needs RTX 40/50-series

THE WITNESS — five locations, one woman, one voice, one repeated sentence. Rendered on the bf16 checkpoint under the same fp8 scaled-mm compute path this file uses (hybrid: fp8 linears, bf16-exact modulation). Native fp8 tensor-core matmul for iteration speed. model repo

Q8_0 GGUF — the fidelity pick

23.1 GB · 0.36° deviation, 44 dB SNR — functionally the same model as bf16

ONE ROOM (final) — one den, fixed tripod, she walks, sits, and returns. For the JoyAI-Echo multishot pack; weights dequantize to bf16 compute. Runs on any GPU. model repo

Q5_0 GGUF — the 16 GB pick

15.5 GB · 2.82° deviation · samples differently rather than worse

THE WITNESS (Q5_0 take) — the same five-location piece, Q5_0's own sample. For 16 GB cards on the multishot pack. model repo

INT8 ConvRot — the stock-ComfyUI pick

27.2 GB full bundle · W8A8 with Hadamard rotation · stock ComfyUI ≥ 0.27, SM ≥ 7.5

Low-light stress test — she says “even in the dark, the gradients hold,” and they do. Standard LTX-2.3 workflows, no custom nodes; 1.5–2x faster on 30-series. model repo

The measurements

Identical activations pushed through the real weight matrices — 60 layers, six types, five depths, both the audio and video branches — compared against bf16:

buildcosine similarityrel. errorSNRangular deviation
Q8_0 GGUF0.9999800.62%44.2 dB0.36°
fp8 as stored0.9995642.94%1.69°
fp8 as computed (W8A8-style)0.9992153.95%2.27°
Q5_0 GGUF0.9987574.95%26.2 dB2.82°

Three findings that were not givens: quantization error does not compound across the 48 transformer blocks (flat from block 0 to 47); the audio and video branches degrade identically (the audio branch is this merge's reason to exist — it is not more fragile); and Q8_0 is close enough to bf16 that the sampler cannot act on the difference.

Why these are exhibits, not comparisons. Diffusion diverges rather than degrades: a few degrees of per-step deviation, compounded over 20+ denoising steps, walks the sampler to a different valid sample — not a noisier copy of the same one. Measured: Q5_0 vs bf16 at an identical seed scores 11.7–14.5 dB PSNR, while two unrelated shots score 12.2 dB. Side-by-side frames of two builds tell you about the seed, not the quant — so each clip above shows what its build produces, judged on its own terms.