Sulphur 2

An uncensored, open-weights AI video generator for local GPU use.

Sulphur 2 is an uncensored, open-weights AI video generator fine-tuned on Lightricks LTX 2.3. It is a community-released model designed to run locally on consumer GPUs. The release includes base weights, distilled LoRAs for faster generation, and specialized ComfyUI workflows. It focuses on high-fidelity realism by training on approximately 125,000 video clips while intentionally excluding 2D animation. As a derivative of the LTX 2.3 DiT-based foundation model, it supports both text-to-video and image-to-video tasks and includes a local prompt enhancer based on Qwen 3.5 to maintain a fully on-device creative pipeline.

How to User

Download the model weights from Hugging Face, import the provided JSON workflows into ComfyUI with LTX Video nodes installed, and run generations locally on your GPU.

Key Features

Uncensored video generation with focus on realism
Open-weights distribution for local deployment
Native Text-to-Video and Image-to-Video support
Distilled LoRAs for faster sampling speeds
Integrated Qwen 3.5-based local prompt enhancer
Compatibility as a drop-in replacement for LTX 2.3 workflows

Use Cases

Creating realistic AI videos without cloud-based content filters
Running high-quality video generation on consumer-grade hardware (16GB-32GB VRAM)
Developing custom video workflows using ComfyUI and LTX nodes

FAQ

What does 'uncensored' mean for Sulphur 2?

'Uncensored' means the model does not apply content filters or restrictions commonly found in cloud-based AI services. This allows for the generation of a wider range of video content, including mature or adult-themed material, based purely on the user's prompt. However, it places responsibility on the user to comply with their local laws and ethical guidelines.

Can I run Sulphur 2 on an 8GB VRAM GPU?

Running Sulphur 2 efficiently on an 8GB VRAM GPU is likely to be very challenging and may result in slow generation speeds or out-of-memory errors. The model is optimized for consumer GPUs with 16GB to 32GB of VRAM to handle its high-fidelity, DiT-based architecture and the associated workflow in ComfyUI. For lower VRAM setups, you might explore using the distilled LoRAs for smaller model sizes or limiting the generation resolution.

Is Sulphur 2 a brand-new AI architecture?

No, Sulphur 2 is not a new foundational architecture. It is a fine-tuned, community-released derivative model based on the Lightricks LTX 2.3 framework, which itself is built on a DiT (Diffusion Transformer) architecture. The innovation lies in its open-weight, uncensored nature, specialized fine-tuning for realism, and the inclusion of LoRAs and workflows for efficient local deployment on consumer hardware.