I'm a bit confused how to use the sulfur_rank768 lora, can you explain? If I use it in place of the distill lora, it generates garbage. If I add it in addition to distill lora, also generates garbage.

@Mohzaic use both and set the sulphur lora to 0.50 for both high and low passes. I use the ceil72 lora and 10 steps. ceil72 set to .7 on low pass and .6 on high pass

@delta45424155 What do you mean high and low pass? This is for LTX. Do you mean for the upscale?

@Mohzaic You need to leave the distill lora as you usually do, or use distill ltx model, I meant use sulphur lora instead of the sulphur base checkpoint.

@ajreyz1384 you can run the sulphur lora from .9 to 1.0 for 1st pass and upscale, thats what i did, better results for image to video. or experiment!

Clone comfy and learn it once, or spend your entire gooner career looking for ways to get X to work on Y.

@85ck014xun768 nah.. i prefere a gradio webui than that noodles nonsense

just rename it to whatever your current LTX model is named in wan2gp and it will load it

@85ck014xun768 Got ComfyUI, got even multiple self-modified txt2img and img2img workflows which are now my main image generation platforms after 2 years spent entirely in Fooocus. But unlike ComfyUI, Wan2GP at least can launch this thing on my prehistoric RTX 3080 Ti and generate 10 seconds long 720p footage in ~3 minutes using distilled model. My attempts to generate something in ComfyUI few months ago failed entirely.

you can load it into kolboldccp but you'll still want to provide it a good system prompt.

@delta45424155 Yeah, I am already using a vision enabled qwen3-based model via Ollama (for I2V prompt enhancement), I was hoping these models would in some way be more optimised/efficient for LTX with baked in system prompts (or similar).

I'd say Eros is better at nsfw t2v

You can use Sulphur for i2v just fine. I just uploaded 1 exemple.

From what I've seen theres motion differences between really long prompts and seeds I2V but I haven't seen direct T2V side by side tests yet. The real sulphur version is the legit model and fusioncow is the one who literally trained itl. You would get the most flexibility by running the actual data lora.

In my muggle eyes, not much differences between Eros and Sulphur in both T2V and I2V. But they say Sulphur's more complete.

@ephchat1004 Interesting, I'm sticking with Eros for the moment, don't see a real reason to have both models in my folder.

welcome to the LTX community, no one knows anything, people create stuff, put it out, no actual instruction or working workflows. Everyone assumes everyone else is a programmer or something

I use CheckpointLoaderKJ and VAE goes into Video Vae. Audio I use regular ltx2.3 but I think the original WF for sulphur used it also for Audio, but since I dont care too much about audio I dont mind. edit: In my workflow if I use it for audio I get an error : "
AttributeError: 'VAE' object has no attribute 'latent_frequency_bins'"

Mines on 0.6 and works fine.

@RadomAccessFor It's recommended to run it at 0.9-1.0 in the first pass, and 0.5-0-6 in the second pass.

@K0DEX i try 384 with 0.3 + condasafe with 0.5 (no audio) work well Reduced the feeling of ghosting and avoided the audio damage caused by the conda :)

@djjayem940696 tks

This is the distilled lora for the sulphur dev checkpoint. You can use the Sulphur lora rank 768 combined with any distilled checkpoint too (0.8-0.9).

it is censored as heck due to not being abliterated, use another model. https://huggingface.co/GitMylo/nsfwvision-v5_qwen3.5-9b-sft

i think you need to use 10eros for better i2v

yeah, I2V is weaker in this base, can confirm. That said, scale back your test to something simple, to prove it's not just your prompt. You can't really choreograph a whole bunch of stuff with LTX2.3 because it doesn't strictly follow your "plan". I learned that the hard way trying to do a I2V striptease sequence. Game up because that training data is super weak and it's a complex sequence of actions, which LTX doesn't like as much.

You might need more frames, I've noticed when watching the preview window using KJ node 'LTX2 Sampling Preview Override' that when there aren't enough frames to perform all those complex actions such as "the man lowers the duvet to reveal his erect penis" followed by "the woman positions herself lying prone horizontal with her legs pointing to the left of the screen." then pepe goes into mouth etc. Yeah it will on the first step show in the preview window it doing what you want but due to there not being enough frames it will mess it up in the next step. And then the woman starts kissing he man or sucking some penis in the wrong place like the mans next or arm lol.

Yeah just use more frames and make sure to follow the official prompt guide for LTX, you actually have to explain every action in sequence. I had been trying for hours to get this thing to work, I attempted reasoning lora's, hard cut lora, different prompts etc. I was not until i noticed the preview window that it clicked, i just needed to use more frames from 121 to 201 currently at a resolution of 1024 x 1024 and its working fine for complex actions.

I've seen gguf's but dont know if they work proper: Q6: https://huggingface.co/smthem/LTX-2.3-test-gguf/tree/main + many: https://huggingface.co/vantagewithai/Sulphur-2-Base-GGUF

use unsloth's UDgguf and try the lora

haha

@flo11ok874 those are still massive in size, more than regular unsloth.

@Starry_Eyes wdym? Q4-k-m is 14.3Gb same at Q4-k-m from Unsloth Ltx. Q3-k-m is 11.1Gb vs 10.8Gb Unsloth.

@FusionCow do you think you could you could make a smaller sized rank lora? 10gb might be too large.

sorry did you have a workflow for guff? I have a 3060 12gb ofvram card.

it's talking to you lol

Eros is more locked in than sulfur, but it still does get stuck on one kinda body per prompt. I can get it to make me maybe three or four different body types, but then it'll be locked there for the prompt. I feel like it does better when I start zoomed in on the face, which makes them more unique, then zoom out and maybe the body/breasts will be different. Honestly though if you want it to work just use I2V, it works much better with the starting image.

A rather large portion of the dataset did have women in it with huge chests, this will supposedly be fixed in a future version. But yes I agree. I typically do I2V on animated furries and waifus of various kinds, getting them to have the appropriate sized breasts when removing a bra, dress, kimono etc, has been a challenge and I have brought this up on their Discord. The big breasted dataset is likely to be evicted entirely (it was mostly overweight women with massive honkers, which I believe skewed the model a fair bit)

If you want to NSFW LTX-2.3 stuff, it's currently the best option (other than 10eros, which includes this).

The lora is on Hf, linked in the text