Sulphur 2 Base is an uncensored, community-built video generation model on top of LTX 2.3. It runs text-to-video and image-to-video natively and supports every other LTX 2.3 format, with an optional prompt enhancer and distill LoRA in the same release.
Originally released by SulphurAI on Hugging Face. All credit for the model goes to the SulphurAI team and contributors below. Civitai is hosting a mirror so creators can run it on-site - please head to the original repo for weights, updates, and to support the project directly.
Built by
SulphurAI - Hugging Face | Discord | Ko-fi
TenStrip - testing and merging (LTX2.3-10Eros i2v merge)
silveroxides (
@s1lv3rc01n) - testing, merging, quantizing@mov7162- Musubi Tuner guidanceTwo anonymous funders who made the original Sulphur and Sulphur 2 possible
Many other community contributors
Native t2v and i2v on LTX 2.3
Sulphur 2 inherits the LTX 2.3 pipeline end to end, so anything an LTX 2.3 workflow can do (t2v, i2v, the rest of the 2.3 formats) works here without extra glue. No converter nodes, no sidecar checkpoints.
Versions on Civitai
Dev -
fp8mixedbuild for local runs that need VRAM headroomDistilled - the included distill LoRA, ideal as the recommended starting point
Pair the Dev version with the Distilled LoRA for the lightest setup. Existing workflows still reference sulphur_final. Use the LoRA or the full model, never both at the same time. The HF release also ships a bf16 build, which is not currently mirrored on Civitai.
Built-in prompt enhancer
A multimodal prompt enhancer ships with the model and is designed to drop straight into LM Studio:
Open your LM Studio model folder.
Create
Sulphur/promptenhancer/.Drop the
.ggufandmmprojfiles inside.
LM Studio will pick it up on next launch. No system prompt is required, send the text (and an optional image) and it returns an enhanced prompt.
Uncensored by design
Sulphur 2 is intentionally unfiltered for adult and creative work. Generations on Civitai still respect the site's content rules.
Links
Discord: https://discord.gg/GSXJhKZ9V
Support the project: https://ko-fi.com/fusioncow
Description
FAQ
Comments (41)
will we be able to use it on the site?
the option at the bottom of the page show green for "generation services allowed"
very interested if they put up a create option for it
Only reason I see them for adding it. Funny too because unless I got contacted by them with assurances I'll never allow generation for anything I upload since they have a habit of randomly banning things for loose nudify violations. Meanwhile they leave all the Flux Klein straight nudify tools uploaded and readily available.
@tenstrip Don't give them any ideas. Last time I saw someone mention this it only led to the removal of the authors Lora AND a competing one, which meant they both lost. And the rest of us too. They're tools and should be treated as such.
If Justin was serious about opening the door slightly to a more spicy CivitAI now that Red is here, this is their chance.
Sulphur 2 Distilled is up for use, I just made a couple of vids. Came out better than the base for sure. And absolutely better than LTX 2.3 . I will have to grab the ggufs for safekeeping.
Hey, this is the maker of sulphur, could you provide a link to the huggingface in the description? it has things like a prompt enhancer that'd be useful
https://huggingface.co/SulphurAI/Sulphur-2-base
@FusionCow Pardon me, but I don't understand - Why is this even here, under "CivitaiOfficial"? You've posted the full checkpoint to your civ profile, and all the variants to hf 4 days ago. With actual quality generation samples, not some random nonsense a throwaway would attach to a model they'd plagiarised.
From your comment, I presume you were not contacted beforehand to ask your permission. Is civ now intent on taking not only credit for community published checkpoints, but the associated revenue from generation usage (lest we forget, 25% creator tip most of which goes to the base + direct tipping)?
The video model creator community which, if anyone needs a reminder, it has not fought to keep around and even actively taken action against in the past?
And why is dev set to generation only? I can perhaps understand wanting to save some bandwidth on a heavy model; However, a layman might be led to believe this variant is one of the cloud service only models (flux pro type deal) when nothing could be farther from the truth.
I am merely an observer, I do not wish to put words in your mouth, but I would appreciate your take on this matter, and @tenstrip's as well.
@firemanbrakeneck Last time I enabled anything for generation it got removed in a day so I won't unless they directly give me protection from vibe mods. I already have a bunch of hallway monitors following me who spam report all my realistic style I2V videos trying to farm protection ranks.
@tenstrip Thank you, this is exactly the sort of thing I'm talking about.
If civ genuinely wishes to reestablish rapport with the creators who built it up and sustain it, it needs to make a show of good faith. This looks to me more like a cash grab than respect.
@CivitaiOfficial @CivitaiOfficial @CivitaiOfficialTeam <- Do better, be better!
I am kinda a newbie with LTX Generation. Using it with Comfy, what folders do I put your models in?
@pjwhoopie checkpoint
@firemanbrakeneck I literally don't know I wasn't made aware of this
is there a working workflow out there somewhere? the one on huggingface is a broken mess
Yeah, Sulphur is cool but that is kind of embarrassing for the creator. Kind of basic, too.
Sulphur 2 looks to be the MF GOAT. It's the closest we have to Grok unmoderated
Can we use it for training on Civitai?
So i have a question, how is this with anime based animations?
What about GGUF version?
@FantasticFever Thank you!
@alexds9 I didn't try it, been using fp8, recommend this lora for distill https://huggingface.co/Seregil13th/Sulphur-2-base/tree/main/distill_loras
@FantasticFever Thank you :)
I've tried so many things to get T2V to work. Different sigmas, steps, using the distilled version or dev version with lora. I just can't seem to get good T2V results... Is this mainly for I2V, is it just not there yet or am I doing something terribly wrong? I downloaded the workflows from their github but ya... Nothing 🥲
It's undertrained right now really, a bit early to pull the shot and offer on-site generation but nobody contacted about this. Needs a ton of lora stacks currently for T2V. Right now dataset is being pruned down for high quality reinforcement runs.
Something is wrong with your workflow. I generated a dozen videos with this text to video model yesterday and everything turned out the way I wanted. True, there is a minus - in all the videos, if you can see women's breasts, you can 'see' nipples through the clothes. Not the nipples themselves, but not quite natural-looking mounds. And the movements, the environment, the camera movement - no problems.
UPD. I just generated text to video in old side workflow - it works. Could you tell me what specifically isn't working for you? What query are you giving? Well, I didn't generate anything strictly "nsfw", maybe that's why everything works?
A Reddit discussion said this model can't handle men well and that every scene has to have a female present. That sounds like BigLove for SDXL where I would have to put like 30 negative terms in to avoid man-man scenes transforming the men into women. However, I see two examples of man-man here (albeit with one of them feminizing the middle guy, especially the butt) so perhaps that comment was about version 1.
@ss9999 It's just undertrained and mostly generalized and not able to fully target very specific prompts like that, but if you do i2v that shouldn't happen with the right detailed prompt and image. Seems like a weak prompt and bad WF to cause that. There's way more gay and male-male interactions than lesbian scenes in the dataset.
@ss9999 I dont have problems with man when make SFW video.
@herkus_baronas631 Thanks Herkus for trying to help. I was generating "******" to use your terminology 😂. I wanted to see how close I could get to wan NSFW t2v. I'm a hound dog, what can I say. Normal prompts do seem to work better. Like if I type in "beautiful naked women," it always get's that right. But if I type in "beautiful naked women doing ****", it's hit or miss it seems. I'd be interested in your workflow if you can share but no biggie if not. I know it's a lot of work. Either way, this model is a fantastic start. Thanks again.
Oh, for the record, I hate sigmas 🤣. How the heck are we supposed to know what numbers to choose? 🫠
@BigSad11 Its not mine wf. I use wf from this video. https://www.youtube.com/watch?v=3HXCeSGnoq0 . Just change models to sulphur's models.
hmm... interesting choice
How do I add this to wan2gp? Just dragging it into the checkpoint folder does nothing.
Wan2GP has its own set of instructions on how to add new models. It's more complicated than dragging and dropping. Go to the Wan2GP github page and look for the instructions on 'finetunes'.
That said, I haven't as of yet been able to get good results out of this checkpoint in Wan2GP. Still experimenting though.
@ExgProp The only reason I'm using wan2gp is because I was following a youtube tutorial saying it's the easiest way to use ltx. I'm wondering if comfy is better for using Sulphur.
I added a .json file to the finetunes folder with the following text:
{
"model": {
"name": "sulfur2Base_distilled",
"architecture": "ltx2_22B",
"description": "sulfur2Base_distilled",
"URLs": [
],
"group": "ltx2_22b",
"ltx2_pipeline": "distilled"
}
}
Good results were achieved using LoRA LTX 2/2.3 [I2V] NSFW (+furry) and [LTX 2.3] Mystic XXX at low strengths.
@AKdesign thx bro. Gonna try later.
@myjokeaccount386 Wasn't trying to be rude in my response, I use Wan2GP as well and it is easy, but to do custom models correctly it helps to have an understanding of the process, which is why I suggested checking out the finetunes page. AKdesign has evidently solved the issues I was having, which is great.
@ExgProp The distilled version doesn't have CFG settings :) and that's what I used. The Dev version apparently requires a Lora like ltx-2.3-22b-distilled-1.1_lora-dynamic_fro09_avg_rank_111_bf16. I tried installing Eros that way, but it often gave me an out-of-memory error. Distilled Sulfur works reliably. Also, a lot depends on the Lora's strength (it's better to set it lower at first) and Prompt, you'll need to experiment. Good luck!
Oh wow, they change the creator from CivitAI to FusionCrow. W
As it should have been.
Can you imagine if the model had been up for 2 weeks, attributed to some user whose only contribution was reposting it from hf and enabling it for generation, then once it peaked the charts (hf's #1 trending), reassociated it with its creator, but didn't reward the proceeds gained nor kept it generate-able?
Good thing that didn't happen, eh.
Does this also work with S2V? Adding custom audio along with first image?