rockerBOO/sulphur-2-nvfp4

Sulphur-2-base (NVFP4)

NVFP4 quantized version of SulphurAI/Sulphur-2-base, a 22B-parameter fine-tune/variant of Lightricks/LTX-2.3 used for image-to-video and text-to-video generation.

Files

sulphur_dev is the base (non-distilled) checkpoint; sulphur_distil is a distilled variant for faster (fewer-step) sampling. Each original bf16 checkpoint was ~46 GB; NVFP4 quantization brings each down to ~22 GB (~2.1x smaller).

Requirements

NVFP4 (FP4 E2M1 block quantization) requires a Blackwell GPU (SM >= 10.0/12.0) for inference (e.g. RTX 50-series, B100/B200). Converted and tested on an RTX 5070 Ti.

Load with ComfyUI (or any loader supporting .comfy_quant metadata for NVFP4 weights) as a checkpoint — these files bundle the transformer, video VAE, and audio VAE (same layout as the upstream ltx-2.3-22b-dev-fp8.safetensors checkpoint they replace).

Quantization method

Quantized with the convert-to-quant (ctq) CLI tool using NVFP4 (FP4 E2M1) block quantization, Comfy-quant format, and the built-in --ltxv2 exclusion filter (keeps select transformer blocks high-precision and excludes the VAE and vocoder from quantization — matching LTX-2's mixed-modal audio/video architecture).

Steps

1. 1.Download the checkpoint (single file, no sharding):

   hf download SulphurAI/Sulphur-2-base --include "sulphur_dev_bf16.safetensors"
   hf download SulphurAI/Sulphur-2-base --include "sulphur_distil_bf16.safetensors"

1. 1.Quantize directly (single file, no shard merge needed):

   uv run --with convert-to-quant --with triton --with safetensors --with tqdm --with numpy ctq \
     -i sulphur_dev_bf16.safetensors \
     -o sulphur_dev_nvfp4.safetensors \
     --ltxv2 --nvfp4 --comfy_quant --save-quant-metadata --simple --low-memory

(same command for sulphur_distil_bf16.safetensors -> sulphur_distil_nvfp4.safetensors)

1. 1.Delete the raw bf16 file, keeping only the quantized output.

sulphur_dev_nvfp4.safetensors: 9643 tensors, 1232 layers quantized (verified via the _quantization_metadata header).

Flags used:

• —--nvfp4 — FP4 E2M1 block quantization
• —--comfy_quant — Comfy quantization tensor/metadata format
• —--ltxv2 — LTXv2 exclusion filter (keep select transformer blocks high-precision, exclude vae/vocoder)
• —--simple — skip SVD optimization, use simple quantization
• —--save-quant-metadata — embed _quantization_metadata in the safetensors header
• —--low-memory — stream tensors during quantization to reduce RAM usage

Intended use with this checkpoint

See ltx2.3_i2v_workflow_requirements.md in this repo/directory for the full set of companion files (LoRAs, text encoder, spatial upscaler) needed to run an image-to-video workflow with this checkpoint in ComfyUI, adapted from Lightricks' official LTX-2.3 workflow.

License

LTX-2 Community License Agreement, inherited from the base model SulphurAI/Sulphur-2-base / Lightricks/LTX-2.3.