SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.
見出し画像

How to Easily Try the Intelligent Video AI 'Sulphur 2' with ComfyUI on Google Colab

サファ

I recently tried Sulphur 2, and it was quite good.

I've quickly made it possible to try Sulphur 2 with ComfyUI on Google Colab, so I'll introduce how to use it.

If you use the setup code for Colab that I've published, you can try it just by launching ComfyUI on Google Colab and loading the workflow for Sulphur 2.

If you find it difficult to set up a local environment or just want to try out Sulphur 2, please check it out.

What is Sulphur 2?

Sulphur 2 is a video generation model based on LTX-2.3. It supports both Text to Video and Image to Video.

I've explained the overview and more on YouTube, so please take a look there.

How to try Sulphur 2 on Google Colab

First, paste the contents of LTX-Video_ComfyUI.sh from the Gist you can reach from the post below into a Google Colab cell.

As shown in the image in the post, please launch it with the checkmark on the Sulphur 2 side. It uses quite a bit of GPU and memory.
(I am using it on an A100)

Main files to be downloaded

A quick note: for Sulphur 2, the following files are mainly used.

  • sulphur_dev_fp8mixed.safetensors

  • ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors

  • gemma_3_12B_it_fp4_mixed.safetensors

  • ltx-2.3-spatial-upscaler-x2-1.0.safetensors

Basically, if you run it on Colab, it will be automatically placed in the necessary location, so you don't need to worry about it.

Launching ComfyUI

Once the setup is complete, ComfyUI will launch.

After waiting a short while, a URL for accessing ComfyUI will appear in the Colab output field.

Opening that URL allows you to operate ComfyUI in your browser.

Loading the workflow

Once the ComfyUI screen appears, load one of the following workflow JSON files.

Click here if you want to try T2V.

Click here if you want to try I2V.

You can load the workflow by dragging and dropping the JSON file onto the ComfyUI screen.

Once the workflow is loaded, enter a text prompt and execute it to start generation.
You can change settings such as the duration from the configuration items below the red frame.
For I2V, a node for video uploading will also be displayed, so please upload the image you want to animate there and generate it in the same way.

After that, just adjust the prompt, image, video length, etc., and generate.

By the way, I am currently testing with a length of about 15 seconds. I haven't been able to try many settings yet.
I would be happy if you could leave a comment if you discover things like 'I found this setting to be stable!'

About the workflow JSON

The workflow JSON distributed this time is based on the official ComfyUI workflow example for LTX-2.3, with model specifications and some settings adjusted to work with Sulphur 2.

Therefore, the JSON contains metadata added by ComfyUI when saving, initial prompts derived from the official sample, sample image names, etc.

When actually using it, please rewrite the initial prompt to match the video you want to create.

Sulphur 2 is good.

Due to the nature of the model, it is difficult to provide actual examples, but as a local video generation AI, I get the impression that the quality and expression have improved significantly.

However, the required specs are quite high, so while it won't be completely local (since it's a Google Colab environment), if you just want to try running it, I hope you will use this article as a reference!

いいなと思ったら応援しよう！