@c0ur4ge I wish I could donate the 3 yellow buzz i have, sadly its locked to 10 or something. I really dont wanna play with crypto or credit card on this site haha. Wish they had paypal.

@c0ur4ge I told you all about it way back. It was called LTXXX back then, but name changed to not piss off lightricks (my suggestion). There's been some debate as to whether training this type of stuff would work better by training on the sulphur base model because it matches latent data better and could inherit it for free, but obviously its more complex than that. Possibly worth messing with though, training has been done in testing successfully.

Does sulfur work with the normal ltx2.3 workflow?

still not better than wan though :(

@Jesus__Skywalker Only if your settings are off. Get the settings right and ltx 2.3 is way better than wan and quicker with audio. Prompting is like 90% of it with ltx 2.3 imo.

@noyart I tried sulfur lora with Q8 gguf, works good, I'm sure if I tried fp8 version it would work too. You sill may need an extra nsfw lora for certain situations, if you have everything in you starting image (all body parts) then you probably just need sulfur lora. It also does decent at t2v but you need a good prompt.

@Jesus__Skywalker yeah if you want 5 seconds with no audio I guess go for WAN. dreamlay just released new lora that, I know his preview lora paired well with sulfur, I didn't try the new one yet. 

@FantasticFever what's sulfur? can't find it with search

https://civitai.red/models/2594061/sulphur-2-base?modelVersionId=2914022

@flo11ok874 oh, gotta wait for GGUF then

the lora is in the files: https://huggingface.co/SulphurAI/Sulphur-2-base/tree/main

Just use it like any other lora in your GGUF workflow

@MisticRain69 All of the examples are inferior to Wan in terms of movement and physics - they're really stiff

@Jesus__Skywalker Wan wins because physics.

@Jesus__Skywalker I was certain it would never touch Wan, but I think it hit Wan-levels recently. I was surprised by the specific aspects it matched. For example, I thought it would never get fluid physics as well as Wan 2.1, but I think it can now hold its own against Wan 2.2 with the newest lora on top of Sulphur.

@Kingp0dd529t They have Q6 sulphur 2 gguf here. https://huggingface.co/smthem/LTX-2.3-test-gguf/tree/main

@FantasticFever wow you remembered lol thank you!

@Jellai Even male anatomy? I haven't tested LTX 2.3 yet but the samples I've seen don't seem very impressive for young attractive men with good-looking anatomy.

@ss9999 There are some decent loras, but it doesn't reach the best of Wan. When I say it "touches" Wan, I mean it does some things about as well. Not everything. And not really anything better, except for how it relates to audio. Still, it's close enough now that I'm no longer dismissing it like I used to.

Even when 2.3 was released, it didn't do anything about as well as Wan, at all (again, except for some audio-related things which it did better). I didn't want to waste any time with it. But now, I've gotten to the point where I respect it. I feel like whatever they build from here in a later version could possibly beat Wan at something.

cause of "girl" probably

@makiaeveli I tried but both girl or g1rl are fine. Have written a email to support to look into this

@c0ur4ge This one?

https://huggingface.co/RuneXX/LTX-2.3-Workflows/blob/main/LTX-2.3_-_I2V_T2V_Dev_Full-Steps.json

There are a lot of choices and the ones I tried previously did not work, but maybe I did not try the right one.

@Jianka try multiple. some workflows work better than others.

Make sure you check your case sampler settings as well because sometimes using a different sampler will make it blurry looking so you had to mess around that and the strength settings. I haven't used this new one yet but I am going to download it and see what it's like

@c0ur4ge Because you're hesitating. The 'ltx 2.3 sulphur dev' model has had additional training with 125,000 real NSFW videos from various sources, so the model already understands human anatomy, genitals, and NSFW/SFW actions in general much more efficiently. Therefore, training Loras on it yields MUCH better results because it already has hundreds of thousands of real NSFW data points.

The results I'm getting with 'ltx 2.3 sulphur' on T2V without any loras are simply incredible, very realistic, I've never obtained such amazing results before. When I add loras trained on 'ltx 2.3 sulphur' they become even more incredible.

@c0ur4ge I'm not 100%, but I think they were referring to this or similar: https://civitai.red/models/2598050/penis-lora-ltx-23-sulphur-lora?modelVersionId=2918443

Also, sulfur did ship as a lora at first, but updated their repo to include a dev checkpoint. Again, may be the reference

@c0ur4ge sulphur was trained on real videos. you assuming they used any of your lora is wild. this is something you could have easily verified yourself

Sulphur uses a 125k dataset and has been through 2 different full runs, first one was 60k. I'll give you possible overlap but no sulphur releases use dr3aml4y in any merge. Any semblance of that could be from just the way LTX interprets prompts with explicit data on a given seed lining up despite the where the data is from.

@c0ur4ge "for all I know their checkpoint contains parts of my lora". maybe you didn't mean it but that is an accusation dawg. it's all cool if you didn't mean it but you have to understand how that reads to someone who doesn't know better.

@c0ur4ge Training on sulphur is more about training for sulphur, it's actually undertrained and if you push targeted data past steps you'll over influence base and have a lot of improvement on what's already there. What's already there is mostly detailed nudity so you're not starting like in base LTX2.3 which is to have to correctly construct full nude anatomy, it's already there and you can refine it. Should know that though this is for other people who don't understand why to train on sulphur, you're not training for base LTX if you do. Also with what this lora is, unneeded imo unless it is tuned to focus only realistic motion and high res realistic anatomy to hone realistic outputs in sulphur, that version would also imo need less training time than base to produce helpful results.

It also has problems with sound, which were not present in preview. Here it's almost silent.

@c0ur4ge when you say lower distillation str, between 0.25-0.35 distill strength, which node or item should i actually lower? is it the CFG value under KSampler? Thanks

had the same issue. it works again now.

@c0ur4ge haha! that's why I let the wolves calm down, it worked after an hour :P

If you dl most of the videos and drag them into comfy it will have the workflow and prompt included

could always download a milk cannon vae and emulate via shooting white vanilla pudding in asynchronous streams

@bamece3943525  I did eventually figure it out, it was more to do with coupling other loras with this one and blending their triggers with this one. For some reason it's not mentioned much but there seems to be a trigger "DR34ML4Y" that is used more for a broader usage then specific individual ones like are on the page.

@c0ur4ge Ok then that was my delusion that the "DR34ML4Y" actually worked lol. I seen a few ppl put it in prompts idky. I put it in mine for a couple gens and it seemed like it was actually making a difference. I take your word for it, its not in there. Thanks for the reply

I tested it again without the lora and the voices became distinct. No clue on what's happening.

Thanks @c0ur4ge . I'm now trying one of the workflows you suggested.

I'm still curious as to how I should get the image ready for I2V to make the best of each of the trigger words.

Your example prompts are great for T2V but trying to master how the image should be posing nude before running the flow has me stumped.

I will continue experimenting. There's a reason why this Lora is incredibly popular. I'm just yet to unlock the I2V 'magic' pose to make your amazing flow work like it does in the T2V examples.

Just to confirm, you have both LTX23 version and a WAN22 version (hi & low). You never need to have all three Lora's loaded in the flow - correct?

You shouldn't need to always have an image posing certain ways to get it to work. I've taken subjects that were just standing and had them move into different positions without any issue using standard WAN2.2 workflows. So, you're issue is very perplexing.

Are you SURE your workflow isn't doing something in the background like limiting tokens, lowering strength, distilling too much data? Are you using a lot of speedup loras (which you shouldn't use if your already using a model with them included)?

How many steps per model are you running? I'd recommend doing a total of 8 steps split between both the high and low model, 3 for high, 5 for low. That said, I can get great results with 2 steps for high and low both for a total of 4 steps. But, that's just to test before moving up to 8-12 steps.

Is your prompt actually describing what's going on or are you trying to prompt using image tags? Example: 2girls, 1boy, intercourse,blowjob,standing,doggystyle, or are you using narrative prompts? With WAN you want narrative prompts that describe movement, expression, and so on.

Also, your checkpoint, try using Daiswa's checkpoints and dr34m together. Lora strength depends on several things but for th emost part 40-75 work very well and if I add more loras I lower but sometimes raise the lora strength up.

Trial and error is all you can do in a lot of cases.

Another tip: Once you have a working workflow for a specific scene - save that workflow and name it after the subject matter. That way you can load that one for particular scene and then build upon it for other scenes while not ruining what you have working.


Ah, one other thing. Video length matters - if you're rendering a 3 second video and expecting a lot of dynamic movement and alterations then it could be difficult.

try cfg 1 on both ksampler. and to save time change the length to 17 so that you can see if the result is good or not. if you're satisfy change to 81 again.

It's compatible, i'm using it all the time with Wan2.2 4steps distilled lightx2v 1030 Q4_KM.

look out at my output here, most of them using this lora as a helper or not as helper.

@anonymous_time_420 You did say that. "This is just a merge of several loras."

Who's loras? Their own? If so, then their claim that this is all personally trained by their team is still valid and your claim, your obvious allusion to them using other people's loras, is a false allegation.

Your statement is claiming that the dream lora is "just a merge of several loras" which implies that it is nothing more than a merge of other people's loras. You didn't specify "a merge of this team's loras" or a "merge of other people's loras" so the context and tone of your message it's clear what you meant.

I suggest you take your own advice regarding logical deduction, context clues, and comprehension.

Maybe less 420 and more reading.

And you certainly didn't back up your claim with any evidence. Which is of course what you should do when you make such claims like, "this is just a merge of several loras." It's not up to courage or anyone else on their team to defend themselves against your claim until you provide evidence that supports your argument. Of which, so far, you don't have. You're just spitting bullshit into the faces of everyone and hoping no one notices. People like you are how misinformation spreads; you don't even take two seconds to fact check yourself.

So? Provide proof of your statement. The dream lora is one of the most commonly used loras in AI gen history and you'll find its use in a lot of workflows for a reason. Because it's very very good; and even outside of its intended scope it can be helpful.

Knowing how to use a tool goes a long way. You can't just always plug a lora into a workflow, crank it to 2.0 strength and get th eresult you're expecting. If only it were that simple...

@DaddyWolfgang Why write such a massive essay?

Also, who even said anything here about whether something is valid? He never mentioned anything in that regard. The only thing one could really challenge here is the statement, “You are better off using Sulphur or more targeted LoRAs anyway.”

He stated that various LoRAs were combined here, but he never said that LoRAs created by someone else were being used. If you are going to start quoting things here, please make sure you do it correctly.

Furthermore, it says so right here on the page: “This LoRA, incredibly, manages to handle ALL of our specialty LoRA concepts – Double/Single Blowjobs, Missionary, Reverse (and regular) Cowgirl, Doggy, wrapped in one model!”

Ultimately, he was simply reiterating what had already been stated here, so I do not see any issue with that. I am not sure why people immediately overreact and feel the need to defend everything. Being able to handle criticism should be normal.