SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI videoSulphur 2

Sulphur 2

Sulphur 2 is a 9B-parameter AI video generation model based on the LTX-2.3 Diffusion Transformer (DiT) architecture, capable of turning text prompts or still images into 5-15 second videos with a cinematic quality. Users can either generate videos instantly through a cloud studio in the browser, or download the open-source weights from Hugging Face and run them locally in a ComfyUI environment with an NVIDIA GPU. The cloud version offers fixed credit pricing, concurrent tasks, and automatic refunds on failure; the local version supports BF16 and FP8 mixed precision, LoRA adapters, and custom node graphs, enabling uncensored, completely private video generation. The model weights are released under the LTX-2 Community License, permitting commercial use. The official website also provides hardware requirement guides, ComfyUI deployment steps, and prompt examples.

Open website

About this tool

Overview

Sulphur 2 is a high-performance open-source AI video generation model with 9 billion parameters, based on the LTX-2.3 Diffusion Transformer (DiT) architecture. It is fine-tuned for smooth motion dynamics, cinematic camera language, and high temporal consistency, and can turn natural language prompts or still images into realistic, seamless 5- to 15-second video clips, reducing the jitter or deformation artifacts common in earlier video diffusion models.

Key Capabilities

  • Text-to-Video: Turns descriptive text prompts into high-definition cinematic videos with natural camera movement and physical realism.
  • Image-to-Video: Turns still images, product photos, digital illustrations, or Midjourney concept art into smooth videos while maintaining consistent subject identity.
  • Camera Trajectory and Motion Control: Supports camera angles such as 360-degree orbiting pans, drone flyovers, slow push-ins, crane rises, and tilt-shift zooms.
  • Diverse Application Workflows: Includes e-commerce product video ads, furniture and home display, dynamic clothing showcases, and more.
  • LoRA Adaptation: Natively supports rank-16 to rank-768 LoRA adapters, allowing fine-tuning in ComfyUI for custom characters, stylized animation, or branding.

Technical Architecture

Sulphur 2 uses a Diffusion Transformer as its foundation, employing decoupled spatial-temporal self-attention layers to separate high-resolution spatial detail from temporal physical motion, eliminating jump artifacts in long frame sequences. Its causally optimized latent video VAE encodes frames with 8x spatial compression and 4x temporal compression, controlling VRAM usage while maintaining 4K upscaling clarity.

Usage

Two creation paths are provided:

1. Cloud Video Studio Generate AI videos directly in the browser without needing an expensive high-VRAM GPU. The hosted studio leverages high-performance managed infrastructure, offering fixed credit pricing, a fast render queue, and reliable downloads, with no need to install Python, CUDA, or ComfyUI dependencies.

2. Local ComfyUI Download the official Sulphur 2 9B checkpoint from Hugging Face and run it locally on your own NVIDIA GPU. You can build custom ComfyUI node graphs, attach community LoRAs, and enjoy 100% private, uncensored video generation with no ongoing subscription fees.

Local deployment steps are as follows:

  1. Install ComfyUI and custom nodes: Make sure ComfyUI is updated to the latest version, and install PyTorch 2.4+ and CUDA 12.4; clone the required LTX extension nodes into the custom_nodes directory.
  2. Download the Sulphur 2 weights: Download the checkpoint from the official Hugging Face repository into the models/checkpoints folder; you can choose the full BF16 or the optimized FP8 mixed precision.
  3. Load the official workflow JSON: Open ComfyUI in the browser, drag and drop the official Sulphur 2 T2V or I2V JSON workflow onto the canvas, enter a prompt, and click Queue Prompt to start generation.

Hardware requirements

  • Minimum configuration (FP8): Can run the quantized FP8 mixed-precision checkpoint (sulphur_dev_fp8mixed.safetensors), requires ComfyUI memory offloading to be enabled, suitable for GPUs with 12 GB to 16 GB of VRAM (such as RTX 3060 12GB, RTX 4070, RTX 4080).
  • Recommended configuration (BF16): Runs the uncompressed Sulphur 2 BF16 checkpoint; an NVIDIA GPU with at least 24 GB of VRAM is recommended (such as RTX 3090, RTX 4090, A5000, A6000).
  • Production cluster: Multi-GPU configuration.

Applicable scenarios

  • Independent filmmakers, animation and VFX studios performing local uncensored video generation.
  • E-commerce product video ad production, generating Shopify, Amazon, TikTok ad videos from static catalog photos.
  • Commercial display videos for furniture and home goods, apparel and fashion, etc.
  • Content creators and AI artists rapidly prototyping in the cloud and batch generating locally.

Limitations and notes

  • Running Sulphur 2 locally requires substantial GPU compute power and VRAM; if you do not have 24 GB of VRAM, you can use the cloud studio.
  • The cloud studio uses prepaid fixed-credit quotes, with automatic refunds if a task fails; actual credit consumption varies by model, duration, resolution, and audio, and the final quote will be displayed before submission.
  • The model weights are released under the LTX-2 community license; commercial applications must comply with the community terms, and enterprise or commercial pipelinesthe full license terms should be reviewed before deployment.
  • More details on pricing, performance, or compatibility are not yet fully provided on the official website.

Related tools

Browse category