Winnougan/Sulphur-2-base-INT8

Sulphur 2 — INT8 Quantized (Dev & Distilled)

INT8 row-wise quantized versions of Sulphur 2, an uncensored video generation model based on LTX 2.3, supporting both T2V and I2V natively along with all other LTX 2.3 formats.

These are community quantizations of the original Sulphur 2 release — full credit to the original authors. This repo provides faster, smaller INT8 variants for users running on lower-VRAM GPUs (tested on an 24GB on a RTX 4090).

Original Model

• Base architecture: LTX 2.3
• Original release: Sulphur 2 (Dev & Distilled, bf16)
• Discord
• Support me on Ko-fi

All credit for the base model, training, and architecture goes to the original Sulphur 2 authors. This repo only provides INT8 quantized derivatives for faster local inference.

What's Included

File	Source	Quantization
sulphur_dev_int8mixed_tensorwise.safetensors	Dev (bf16)	INT8, tensor-wise, LTX2 layer filtering
sulphur_distil_int8mixed_tensorwise.safetensors	Distilled (bf16)	INT8, tensor-wise, LTX2 layer filtering

Both were converted using convert_to_quant with the --ltxv2 model filter, which keeps sensitive layers (VAE, vocoder, embedding/connector blocks, and select transformer blocks) in their original precision while quantizing the rest of the transformer to INT8.

Why INT8?

• Roughly 2x smaller file size compared to the original bf16 weights
• Noticeably faster inference, especially on lower-VRAM GPUs
• Quality remains very close to the original bf16 model in practice — the --ltxv2 filter excludes the layers most sensitive to quantization loss

Getting Started

Follow the same setup as the original Sulphur 2 release:

1. Download either the Dev or Distilled INT8 file from this repo (or both — don't load both at once).
2. If using the Dev model, you can optionally pair it with the official distill LoRA from the original repo. Don't combine the LoRA with the Distilled model — use one or the other.
3. Load the model into ComfyUI using a QuantOps-compatible UNet loader node.

Prompt Enhancer

Sulphur 2 ships with a prompt enhancer. The easiest way to run it is through LM Studio:

1. Open your LM Studio models folder in your file explorer.
2. Create a folder named Sulphur, then a subfolder inside it named promptenhancer.
3. Place the GGUF file and the mmproj file from the original release inside that folder.
4. Load the prompt enhancer model in LM Studio.

There's no system prompt required — just send the text (and optionally an image) you'd like enhanced.

Notes

• These are INT8 quantizations only — no changes were made to the underlying model weights beyond standard scale-and-clamp quantization with row-wise scaling.
• If you run into quality issues, compare against the original bf16 release to confirm whether it's a quantization artifact or an unrelated workflow issue.
• Better setup instructions and training guidance are expected from the original authors soon — check their Discord and Ko-fi for updates.

Credits

• Original model & training: Sulphur 2 authors
• Quantization tooling: silveroxides/convert_to_quant
• INT8 quantization (this repo): Winnougan