Back to News

Sulphur 2: A 9B Video Generation Model Based on LTX 2.3

news

Sulphur 2 is a community fine-tune of LTX 2.3 offering text-to-video and image-to-video generation with a built-in prompt enhancer and distill LoRA, trained on 125K+ curated clips.

Sulphur 2 is a community-built, open-weights video generation model fine-tuned from Lightricks' LTX 2.3. With over 1,800 likes and 800K+ downloads on Hugging Face, it has quickly become one of the most popular LTX 2.3 derivatives in the open-source video generation space.

Overview

Released by SulphurAI on May 3, 2026, Sulphur 2 is a community video generation model fine-tuned on 125K+ curated video clips. Built on top of Lightricks' LTX 2.3 architecture (a 22B parameter diffusion transformer), it provides both text-to-video and image-to-video generation.

The model supports all LTX 2.3 native formats and comes with several companion components including a prompt enhancer, a distill LoRA for faster inference, and ready-to-use ComfyUI workflows.

Key Features

FeatureDescription
ArchitectureFine-tuned from Lightricks LTX 2.3 (22B DiT)
Task SupportText-to-Video (T2V) and Image-to-Video (I2V)
Model Size~9B effective parameters (dev BF16: 16.7 GB, FP8: 8.5 GB)
Prompt EnhancerBundled GGUF prompt enhancer available
Distill LoRA1.1 distilled LoRA for faster generation
Training Data125K+ curated video clips
LicenseCustom non-commercial / research

Model Variants

Sulphur 2 ships with several weight variants to accommodate different hardware:

  • sulphur_dev_bf16.safetensors (~16.7 GB): Full precision dev version
  • sulphur_dev_fp8mixed.safetensors (~8.5 GB): Quantized dev version for lower VRAM
  • sulphur_distil_bf16.safetensors: Distilled version for faster inference
  • sulphur_lora_rank_768.safetensors: LoRA for fine-tuning
  • distill_loras/ltx-2.3-22b-distilled-lora-1.1.safetensors: Additional distill LoRA

Prompt Enhancer

One of the standout features of Sulphur 2 is the inclusion of a dedicated prompt enhancer: a small language model packaged as GGUF that automatically rewrites and enriches user prompts for better video generation results. Two variants are available:

  1. Standard prompt enhancer: General-purpose prompt rewriting
  2. Alt prompt enhancer: An alternative variant without prompt filtering

The prompt enhancer can be loaded in LM Studio or any GGUF-compatible inference UI.

ComfyUI Integration

Sulphur 2 ships with ready-to-use ComfyUI workflows included in the model repository:

WorkflowDescription
T2V BaseText-to-video base workflow
T2V DistilledText-to-video with distill LoRA
I2V BaseImage-to-video base workflow
I2V DistilledImage-to-video with distill LoRA

Simply download the model weights from Hugging Face and place them in your ComfyUI/models/diffusion_models/ directory, then load one of these workflows and select the Sulphur 2 checkpoint.

Performance & Quality

Sulphur 2 was trained on a carefully curated dataset of 125K+ video clips focused on human-centric and diverse visual content. It handles:

  • Human motion and interaction with improved consistency
  • Improved prompt adherence compared to the base LTX 2.3 for certain creative use cases
  • Native portrait and character-focused generation

Availability

Sulphur 2 is released under a custom non-commercial / research license. Check the license terms before using the model in any commercial application.

Summary

Sulphur 2 represents one of the most popular community fine-tunes of LTX 2.3, offering creators an alternative for video generation with helpful tooling like prompt enhancement and distilled inference. Its combination of quality, flexibility, and ready-to-use ComfyUI workflows makes it a compelling option for video generation enthusiasts.

Sulphur 2: A 9B Video Generation Model Based on LTX 2.3 | ComfyUI Wiki