LAXMAYDAY/LTX-2.3-22b-dev-int8-tensorwise

LTX-2.3 22B dev — INT8 tensorwise (+ConvRot), stock-ComfyUI native

Offline INT8 quantization of Lightricks/LTX-2.3 (ltx-2.3-22b-dev.safetensors), produced with Comfy-Org/comfy-quants (export-model-int8-tensorwise). Loads natively in stock ComfyUI >= v0.27.0 (QUANT_ALGOS["int8_tensorwise"], SM >= 7.5 / Turing+) — no custom node required.

Contents

Single-file bundle (drop into models/checkpoints/, load via CheckpointLoaderSimple; LTXAVTextEncoderLoader and LTXVAudioVAELoader read the text-embedding projection / audio VAE / vocoder from the same file). Text encoder (Gemma-3-12B) not included — use e.g. Comfy-Org/ltx-2 gemma_3_12B_it_fp4_mixed.safetensors.

Quantization contract

• 1,496 quantized Linears (blocks 2–45 incl. to_gate_logits; blocks 0/1/46/47 kept bf16) — layer-for-layer identical to the official Lightricks/LTX-2.3-fp8 / -nvfp4 releases.
• Per layer: int8 weight + float32 [out,1] scale + comfy_quant marker {"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256} (byte-exact to stock ComfyUI's own save path). All 1,496 layers ConvRot-rotated.
• Quant math bit-faithful to comfy-kitchen >= 0.2.15 (quantize_int8_convrot_weight).
• VAE / audio VAE / vocoder / text-embedding projection copied verbatim (bf16/f32).

Measured (RTX PRO 6000 Blackwell, 512²×49f+audio, 20 steps, torch 2.10.0+cu130)

	bf16	this artifact
Disk	46.1 GB	29.2 GB
Peak VRAM (full pipeline)	53.7 GB	38.6 GB (−28%)
Sampling speed	2.37 it/s	2.73 it/s (1.15×)
Video PSNR vs bf16	—	45.4 dB

Best fidelity and the only faster-than-bf16 format among fp8/mxfp8/nvfp4/int8 on this model.

License

Inherits the LTX-2 Community License Agreement from the base model (free below $10M annual revenue; derivatives must remain under the same license). See the license link above and the license field embedded in the checkpoint metadata.