caiovicentino1

/

LTX-2.3-22B-HLWQ-Q5

like 39

Yes please!

Hi @FearL0rd and @DeadMan3000 , thanks for the ping.

TL;DR — No native ComfyUI node today, but there's a clean workaround that lets you use this repo inside ComfyUI-LTXVideo right now, and a plan for native HLWQ loading.

Workaround (works today, ~5 min setup)

The HLWQ-Q5 format stores weights as Hadamard-rotated 5-bit codes + BF16-kept layers. Running setup.py dequantizes the codes back to BF16 and writes a standard Lightricks-compatible checkpoint you can point ComfyUI-LTXVideo at:

cd /path/to/ComfyUI/models/checkpoints
git clone https://huggingface.co/caiovicentino1/LTX-2.3-22B-HLWQ-Q5
cd LTX-2.3-22B-HLWQ-Q5
python setup.py   # dequantizes to ltx23_bf16.safetensors (~44 GB BF16)

After setup.py finishes, the file polarquant/bf16/ltx23_bf16.safetensors (+ the two upscalers already in the repo) can be loaded by the official ComfyUI-LTXVideo nodes like any other LTX-2.3 checkpoint. Caveat: this trades the VRAM/download benefit of HLWQ (15 GB on disk) for ComfyUI compatibility — the dequantized file is ~44 GB.

Native HLWQ ComfyUI node (planned, not built yet)

For users who want to keep the compression benefit inside ComfyUI, a dedicated custom node is the right path:

• LoadHLWQLTXVideo — reads hlwq_config.json + polar_state_chunk{0,1,2}.safetensors + bf16_kept.safetensors, materializes the UNet on GPU with per-block FWHT + 5-bit dequant inline (same pipeline as setup.py, but streaming instead of write-to-disk)
• Wraps as a standard MODEL output socket so it slots into existing LTXVideo workflows

Rough effort: ~2-3 days to build, plus testing against the LTX official workflow examples. The dequant kernel is already working in setup.py, so the ComfyUI wrapper is the main new code.

Would either of you want to alpha-test?

If yes, open an issue on github.com/caiovicentino/polarquant (or reply here with your ComfyUI version + GPU model) and I'll ping you when the node is ready to try. Realistic ETA 1-2 weeks — the mechanistic-interp work on Qwen3.6 is the current priority, but this is next in the queue once that lands.

Meanwhile the workaround above should unblock you today if you have 44 GB of disk for the dequantized BF16.

Well that kinda defeats the object. lol Anyhow I just found out it only works with 50 series GPU's and I have a 4080 Super. Oh well. At least I know now.

Thanks anyhow. I'm sure others will appreciate it.

BTW people can convert to NVFP4 in Comfy using the kitchen node. Only LTX 2 19B is listed as compatible in it though.

Hello. Anything new about ComfyUI node?