Stable Release - LTX 2.5 (22B) FP8 (e4m3fn) for ComfyUI
Selective FP8 (float8_e4m3fn) quantization converted directly from the official LTX 2.5 22B BF16 Transformer weights.
Key Features:
- Selective Quantization: Converts 2D weight matrices to FP8 (e4m3fn) while preserving LayerNorms, AdaLN modulations, and dual-stream audio/video conditioning blocks in original BF16 to eliminate visual noise and audio-video desync.
- Fits 100% inside 24GB VRAM (tested on RTX 4090 + 64GB RAM).
- Fast generation time (~99s) with ZERO PCIe RAM offloading.
- 100% native PyTorch implementation (no experimental torchao or Blackwell dependencies required).
ComfyUI Pro-Tip: Place a VRAM Cleanup / Free GPU VRAM node right before your sampler to purge Gemma 4 12B Text Encoder from VRAM and achieve full generation speed!
Hugging Face: https://huggingface.co/guillaume127/LTX-2.5-FP8
GitHub Code: https://github.com/Guillaume-127/LTX-2.5-FP8
Description
Initial release of LTX 2.5 (22B Distilled) quantized to FP8 (e4m3fn) with selective BF16 preservation for ComfyUI.