Naming notice (2026-04-10). The "PolarQuant" technique used in this model is being rebranded to HLWQ (Hadamard-Lloyd Weight Quantization). The change is only the name; the algorithm and the weights in this repository are unchanged.

The rebrand resolves a name collision with an unrelated, earlier KV cache quantization method also named PolarQuant (Han et al., arXiv:2502.02617, 2025). HLWQ addresses weight quantization with a deterministic Walsh-Hadamard rotation and Lloyd-Max scalar codebook; Han et al.'s PolarQuant addresses KV cache quantization with a random polar rotation. The two methods are technically distinct.

Existing loaders that load this repository by ID continue to work without changes. Future model uploads will use the HLWQ name.

Reference paper for this technique: arXiv:2603.29078 (v2 in preparation; v1 still uses the old name).

LTX-2.3 (22B) — PolarQuant Q5 (Bit-Packed)

PQ5 quantized LTX-2.3 — joint audio-video generation, 22B params.

46 GB → 15 GB (-68%) | cos_sim 0.9986 | 1,347 layers quantized

Download Size

Compression

Component	Original	PQ5 Packed	Reduction
Transformer (1,347 layers)	37 GB	4.6 GB	-88%
VAE + Skip (4,600 layers)	9.1 GB	9.1 GB	BF16 kept
Upscalers	1.3 GB	1.3 GB	BF16 kept
Total	46.2 GB	15 GB	-68%

Quick Start

# 1. Install
pip install safetensors huggingface_hub scipy

# 2. Download & setup (15 GB)
git clone https://huggingface.co/caiovicentino1/LTX-2.3-22B-PolarQuant-Q5 ./LTX-PQ5
cd LTX-PQ5 && python setup.py

# 3. Generate video
python generate_ltx.py --prompt "A cat playing piano in a jazz club"

Architecture

22B parameters — largest video model we've quantized
48 transformer blocks, hidden=4096, MLP=16384
Joint audio-video generation — synchronized audio + video
head_dim=64 (Hadamard-compatible)
5,947 tensors total (BF16)
Spatial + temporal upscalers included

Files

LTX-2.3-22B-PolarQuant-Q5/
├── setup.py                          # One-command setup
├── generate_ltx.py                   # Easy generation wrapper
├── polarquant/
│   ├── codes/
│   │   ├── chunk_00_codes.safetensors (1.5 GB)
│   │   ├── chunk_01_codes.safetensors (1.5 GB)
│   │   ├── chunk_02_codes.safetensors (1.5 GB)
│   │   └── chunk_03_codes.safetensors (0.1 GB)
│   └── bf16/
│       └── ltx23_bf16.safetensors    (9.1 GB)
└── upscalers/
    ├── ltx-2.3-spatial-upscaler-x2-1.1.safetensors  (1.0 GB)
    └── ltx-2.3-temporal-upscaler-x2-1.0.safetensors (0.3 GB)

Hardware

GPU	VRAM	Status
A100 (40 GB)	40 GB	Recommended
A100 (80 GB)	80 GB	Full speed
RTX 4090 (24 GB)	24 GB	With offloading

Citation

@article{polarquant2026,
  title={PolarQuant: Hadamard-Rotated Lloyd-Max Quantization},
  author={Vicentino, Caio},
  journal={arXiv preprint arXiv:2603.29078},
  year={2026}
}

46 GB → 15 GB with cos_sim 0.9986. Quantized with PolarQuant.