Download sulphur2-workflow and get Sulphur 2 running in ComfyUI
This page collects the small workflow package, the actual folder map, screenshots, a video preview, and the fussy install notes that usually matter after the first failed Queue button press. It is a little messy on purpose, because ComfyUI installs are often like that too.
First frame captured from the video material. I use it as the visual anchor because it shows the exact Sulphur look without forcing the full video to load first.
Download the sulphur2-workflow package
The repo is the lightweight part: Python node files, requirements, and example workflow JSON. The large model files are not inside the download, so do that part slowly and do not rename files unless you enjoy debugging dropdown menus at 1am.
cd ComfyUI/custom_nodes
git clone https://github.com/mrchen1225/sulphur2-workflow.git
cd sulphur2-workflow
pip install -r requirements.txtWhat is inside this ComfyUI workflow
The download is mostly a ComfyUI custom node package. The useful bits are the LTX2 node loader, model loading helpers, requirements, and one example workflow JSON you can drag into ComfyUI.
example_workflows/ltx23.json
The drag-in workflow file for the LTX 2.3 graph.
LTX2_node.py
ComfyUI node wrapper and exposed controls.
load_utils.py
Loading helpers used by the GGUF and model path logic.
requirements.txt
Small dependency list, but install it before testing.
The model files you need before pressing Queue
The biggest trap is using a random Gemma 3 GGUF. This workflow expects the matching Gemma file with the tensor naming that the loader knows how to read. If you only remember one thing, remember that.
A local setup checklist that saves one evening
This is the order I would check the install in a real ComfyUI folder. It sounds almost too simple, but the errors here are usually path errors, then memory errors, then sampler tuning. In that order, more or less.
Clone it into custom_nodes
Put the repo under ComfyUI/custom_nodes, then restart ComfyUI after installing requirements. A browser refresh is not enough; do the boring full re-start.
cd ComfyUI/custom_nodes
git clone https://github.com/mrchen1225/sulphur2-workflow.git
cd sulphur2-workflow
pip install -r requirements.txtKeep GGUF files in the gguf folder
The Sulphur transformer and the Gemma text encoder both belong in models/gguf. If the dropdown looks empty, it is nearly always a folder typo.
ComfyUI/models/gguf/
sulphur_distil-Q6_K.gguf
gemma-3-12b-it-qat-Q4_0.ggufConnector is a checkpoint
connector.safetensors is easy to misplace. I would not put it in clip or text_encoders; the loader expects it under checkpoints.
ComfyUI/models/checkpoints/
connector.safetensorsVAE files are seperate from checkpoints
Video VAE and audio VAE sit in models/vae. If audio or decoding fails late in the run, check these names before changing sampler settings.
ComfyUI/models/vae/
ltx-2.3-22b-distilled_video_vae.safetensors
ltx-2.3-22b-distilled_audio_vae.safetensorsWorkflow checkpoints: T2V, I2V, upsampler, and IC LoRA
I am using the YouTube first-frame visual for these cards, then keeping the practical notes under it. The actual setup still comes down to names, folders, and the model dropdowns.
Text to video graph
The cleanest starting point: prompt, Gemma encoder, Sulphur model, KSampler, video/audio decode, save.
Image to video graph
Use this when identity or composition needs to survive the first few frames.
Upsampler graph
The practical clean-up pass. Keep it until the base clip is stable, then upscale.
IC LoRA graph
A more opinionated path. Useful, but do not stack random LoRAs until the base run loads.
YouTube video and first-frame reference
This section uses the exact embedded YouTube video you provided. The still image beside it is the captured first frame from that same video, so the page image and page video stay visually tied together.
Embedded YouTube video
This is the provided YouTube embed: kSHNnj5sCyo.
Troubleshooting notes from the bench
Before changing ten settings at once, fix the boring stuff. Names, folders, restart, then memory. The sampler can wait.
--lowvram --disable-xformers --use-pytorch-cross-attention --reserve-vram 2 --disable-smart-memoryUnboundLocalError: embed_tokens_key
Wrong Gemma GGUF. Use the Gemma file prepared for this LTX 2.3 GGUF loader, not a generic llama.cpp conversion.
Sulphur model missing from dropdown
Move the transformer into models/gguf. Do not put it in unet or diffusion_models.
Connector missing
Move connector.safetensors into models/checkpoints and restart ComfyUI.
OOM during model load
Expect a heavy RAM spike. A large Windows pagefile helps, and the low VRAM launch flags are worth using on smaller cards.
sulphur2-workflow FAQ
Where do I download sulphur2-workflow?
Use the public GitHub repo at https://github.com/mrchen1225/sulphur2-workflow. Clone it into ComfyUI/custom_nodes or download the ZIP, then install the requirements.
Does this include all model weights?
No. The workflow package is small. The model weights are large and need to be placed manually in the ComfyUI model folders.
Can it run on 6 GB VRAM?
It can be made to load on low VRAM with streaming and offload, but system RAM and pagefile matter a lot. Treat 48 GB RAM peak as a real planning number.
Should I use the LoRA with the Sulphur GGUF model?
No, keep the LoRA for vanilla LTX 2.3 experiments. Stacking it with the full Sulphur GGUF model is a common way to make the run worse.