Many video generation models still come with heavy restrictions, limited format support, or fragmented workflows. On top of that, prompting can be inconsistent.
Even powerful models may produce weak or unstable results if prompts are not structured properly, forcing creators to spend extra time refining inputs manually. Sulphur 2 (released by SulphurAI) focuses to provide a more flexible and creator-friendly solution by offering an optimized uncensored video generation model built on top of LTX 2.3 that gives you full control and flexibility.
It also includes a built-in prompt enhancer designed to improve prompt quality automatically, helping users achieve stronger outputs with less manual optimization. It supports text-to-video, image to video, distilled generation with loras enabled.
Installation
1. First, you have to do the ComfyUI installation. Older user need to update ComfyUI from the Manager to avoid any errors.
2. As the workflow is based on LTX2.3, you should have the basic LTX2.3 installation setup already done. These models include:
-ltx2.3 fp8/bf16 model (optional)
-text encoders (gemma3 and ltx audio text encoder)
-spatial upscaler
3. Download Sulphur 2 model from its Hugging face repository.
These are different Sulphur model variants / formats optimized for different use-cases like quality, speed, VRAM usage, and finetuning. Choose what ever your system resources suits.
(a) sulphur_dev_bf16.safetensors - for high 16-24gb VRAMs that give high quality generation
(b) sulphur_dev_fp8mixed.safetensors - for low VRAMs with nvidias blackwell supported for 4090/5090 users,gives some quality degradation.
(c) sulphur_distil_bf16.safetensors -distilled variant bfloat 16 for faster generation.
(d) sulphur_lora_rank_768.safetensors-lora model with rank768 used with base model. Used to modify style, improve motion, add Sulphur-specific behavior, finetune capabilities. Higher rank means- stronger learning capacity more detail preservation larger file size more VRAM usage.
(e) sulphur_distill_fp8.safetensors by Winnougan. Its the FP8 quantized variant of the distilled Sulphur 2 base model.
Save base models inside ComfyUI/models/checkpoints folder and save lora into ComfyUI/models/loras folder. Use base models or the lora models, its officially recommended to not use both at the same time, as the base model already loaded with loras.
Alternatives- Use Sulphur GGUFs for low VRAMs but quality will be compromised.
Make sure you already installed ComfyUI-gguf custom node by city96 using Manager. Replace the load checkpoint node with unet loader node if using gguf models. Use Q6/Q8 gguf with only 768 lora, gives you same result with low VRAM consumption.
4. Download ltx-2.3-22b distilled lora fro90 ceil72 condsafe (ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors) . Use it with strength value as 0.5 for the 2nd Sampler (in High Res setup). Save this into ComfYUI/models/loras folder.
5. Download sulphur_experimental_lora_v1 (sulphur_experimental_lora_v1.safetensors) This is experimental and optional lora. Save this into ComfYUI/models/loras folder.
6. Download prompt enhancer (mmproj-BF16.gguf & sulphur_prompt_enhancer_model-q8_0.gguf), its optional . You can use Prompt Enhancer into the workflow. It supports images
through an mmproj file, that usually means it is using a multimodal
projection model (commonly from LLaVA / vision-language models) to let
the LLM understand images along with text more deeply to enhance the
image generation.
7. Refresh and Restart ComfyUI to see the changes.
Workflow
1. Download the workflows from our hugging face repository.
There are four different workflows-
(a) LTX-2.3-sulphur_i2vbase.json (Image to Video base model)
(b) LTX-2.3-sulphur_t2vBase.json (Text to Video base model)
(c) LTX-2.3-sulphur_t2vDistilled.json Text to Video distilled model)
(d) LTX-2.3-suplhur_i2vdistilled.json (Image to Video distilled model)
The base variant consumes high VRAM. Use the distilled one if you have low VRAM setup. Any of the workflows has been divided into low/high resolution generation. Choose either of them to do you generation.
For setup, the recommended starting point is using-
-the distilled versions (fp8mixed or bf16), along with the provided distill LoRA.
fp8mixed supports RTX4090 and RTX5090 with NVIDIA's blackwell latest technology enabled.
2. Drag and drop into ComfyUI. If missing nodes found, then just select and install using the manager.
3. Load sulphur base/ distilled model into checkpoints loader node. lora_rank_768 (sulphur-final) into load lora node. Load other models(loras, vae, text encoders, clips etc) into their respective nodes.
4. Load image into load image node if using image to video workflow.
5. Put prompt into prompt enhancer box inside video settings group node if using prompt enhancer. It will enhance and convert it to detailed one. Otherwise use the default prompt box.
6. Set CFG, video width, height, fps and duration.
7. Hit run to start generation.
