Nah it's not hard just maddening fidgeting with weights. The hard work is done by the guys manually captioning the Sulphur dataset to improve it right now.

github? The model's not on github, at least officially from me.

With version 1.2 you may consider actually training on it. It's kind of becoming it's own branch now. If the loras are just i2v motion enhancers or motion concepts they will work, but if it's something like a t2v style or a unique character and voice, then you want to use 1.2 bf16. That all needs further evaluation and testing though. They're all ltx2.3 models so any trained on base will still work but I imagine with less effectiveness on 1.2 because of all the other weight changes in the model.

@EpicStuffs yes the bf16, and maybe with an abliterated encoder too I've seen people recommend that.

@ElGordoAI The one I put on the eros huggingface page isn't, it was a newer protected method that I combined and made fp8 scaled. But it also doesn't do much differently compared to the normal encoder, maybe if the prompt is super vulgar it will do dialogue better. I used it in all the previews.

I keep getting asked about wan2GP, I have no clue. Managing comfy integration to get people actually sampling the model correctly is enough on my plate.

@Phac123 LTX only falters on smooth complex motions and physicality of objects in the scene and with conditioned transitions; physical physicality in nsfw terms: clothing physics and squirting liquids. Outside of those two things Wan is just 5 seconds of the same 3D sex motion for 2 years straight with the same blink garbage. All Wan outputs actually just look the same when you average them out and it's just the garbage "come fuck these AI girls!" advertisement slop model now.

@Lora_Addict JoyAI Echo has an i2v version that fixes that coming out that I'll attempt to make a new Eros with. Right now I almost have their memory system that transforms parts of the model into conditioning memory holders reverse engineered with nodes so that might get fixed soon. Of course the new LTX version probably also will fix that they've seen that feedback, Sulphur does play some part in changing the face too it probably didn't get enough i2v training steps.

@aurelius I already took care of the with a reference memory patch. Also stop acting like Wan2.2 doesn't literally do the same thing with lora use. I never trained my loras on faces but everyone else did and they all change likeness too, and so does pretty much every video model except Sora and Seedance.

@felRuss the distilled lora LTX released is trash it causes every single issue.

@felRuss use the v5 workflow, I patched the model with a reference mode that makes it actually an i2v model now, the reference is stored and compared against transformer-side like any actual i2v model would work. Extra anatomy being added is normally at upscale and the fix is to just keep turning the distilled strength down until it goes away, on v5 I went all out on that and it uses 0.38 strength upscale with extremely weak sigmas and has very low chance of randomly resampling extra nipples and navels.

Where'd you find that? I added it to huggingface but can't tell what it really does besides just generally enhance prompt sometimes.

https://huggingface.co/TenStrip/LTX2.3_JoyAI_Lora_Extracted

Base LTX2.3 repo or any of the versions of them that are around. It's just an LTX 2.3 model. The VAE is inside the model and the video VAE is sent out of checkpoint loader, the audio VAE needs to be loaded with the LTX audio VAE loader and select the checkpoint.

You have conditioning too high or some other workflow thing going on. If you only ran one gen on one seed and then came and posted that that's pretty asinine.

@MCSizeMatters If you are using my tiled sampler on upscale it will drop lipsync based on what tile audio is being applied too, for vertical ration the audio_carrying tile should be the first tile (the head is usually in the top part) and if audio is passthrough it won't do any lipsync.

@lantoilee543 You need to pass the custom audio in to the second pass, the same way as the first for it to properly resample it and sync.

@lantoilee543 if it was tiled sampler then it might've not had the audio target tile set up, it can only do lipsync on one tile and it has to be targeted with first,middle, or last. It can be bypassed and then it turns into a normal sampler.

I use and used 5090 on all my gens and this one is faster for me since it makes better outputs at lower steps. The full model is not loaded for inference unless you messed your comfy up, you don't need to load any of it into VRAM you can fully offload it if you want.

That's one of the negative sulphur artifacts but also a base model accuracy issue. Anything involving feet works better with this: https://civitai.red/models/2588526/ltx23-feets-and-shoeplay

Is the upscale taking the longest, or using my workflow? My workflows are slow, v4 maybe the fastest. Lowering resolution and removing loras or using fp4 encoder, offloading more to get more VRAM for inference all increase speed.

Increase conditioning (max if it works) and lower distilled strength or use a low ranked one to decrease random stuff like that. Don't use preprocessing on the start image either.

completely blurry is distilled lora not working.

Yes eventually it'll all shift on to that. Right now I'm working on an in-between version built on the JoyAI Echo ltx 2.3 base that is looking like it fixes a lot of issues.

@roboticone a lot of better data trained into it, they turned part of the transformer into a conditioning memory bank that you can then chain up to 5 minutes of outputs from with the same character and voice, but its wrapped up in their inference pipeline. I haven't really seen any examples of it idk if anyone is even bothered sitting down and doing that. I'm repurposing that memory system with patches and nodes for better i2v conditioning and transitions, or trying to at least. They state they'll put out and i2v model as well so this would be a beta version until that one.

I'll add that but it's pretty simple to do, just load a custom audio clip loader and plug it into the audio latent masker at the middle top, path that latent audio mask latent into all the latent combines before the samplers instead of just the second one, and then pipe the trimmed audio clip into the video combines.

@Makaka7 yeah you need to directly provide the trimmed audio clip into the video combines, don't let any audio actually get decoded and used, you just track the output to it and use the straight audio clip as the audio for the video combines. I'll clean that up and add it some time.

@opensource4life you do an i2v workflow but instead of an image input you upload the video and set it to the matching framerate and frame count, or you have to get a wan video up to 97 frames instead of 81 since it has to be in the 24+1 system. Then you just run the gen it will sample audio but it's bound to any mouth movement in the silent gen lipsync override is more complex.

@ArtiOtaku You want to use interpolation even if you lower it to 24 fps, even at 24fps the interpolation makes the video look much smoother and LTX for some reason will be triggered to actually make smooth motions based off the input. Also if it's a simple repeating motion you can easily extend the wan video to 10 seconds of 241 frames easily with no real transition or loss, just leave all conditioning at 1.0 and don't add any noise or preprocessing.

With lower conditioning and the normal distilled lora LTX resamples the whole style of the input. Mainly it's the distilled lora that's why I made all those stripped out small ones to fight against it. Almost all problems with LTX and this model come from those. Next sulphur version I pushed for an additional self distillation as part of it to avoid that.

Not sure how you managed to make it do that but I added a reference memory in my nodes that locks down major shift changes like that.

bf16 on huggingface is a 1:1 with the ltx2.3 dev model, but there's no distillation in any of them. The ltx2.3 distilled lora is terrible and pretty much causes every issue the model has, but its still the only way to make it fast.

@markovochka280 If you do 50 steps you don't want to use the distilled lora. That defeats the purpose of it. You do 50 steps and 4-5 CFG without it and just the model. My older workflows have 13 step sigmas set up but I don't think v1.2 needs any higher than 8-10, version 1 did and benefited more from higher step counts with distilled.

Right now I'm actually building and entirely different distillation away from the lora that uses echo's dmd deltas instead. Seems promising and it's working but needs more testing and relies on my nodes to make it work.

I have v6 coming and it pretty much antiquates those, and almost all other current workflow since it uses an alternative distillation that's much better.

Something inside the model, it's training, or architecture creates a learning block where the model interprets the data as an amalgam. All male orgasms end up as white appearing everywhere with random fluid noise flying out. The closest one is the futa cumshot one which actually does proper squirts and if you let them fly out of frame it looks okay. It'll probably take a dedicated IC Lora for it, or their next version with experts to have the proper motion depiction.

Ignore all that and try the DMD workflow I added instead.

The DMD method seems to have the most identity preservation but still working on refining it's motion and audio quality. If you just follow on HF there will be updates consistently, once I put a WF out I'm usually just on the next iteration.

In order to make the average usage have more detailed sexual motion, which is what you see in most of the shared high rated clips, the original training data is almost x1.8 strength across the blocks. It gives it a skew towards sex and the model is made to animate it. It's not a storytelling model or something like just an uncensored LTX model anymore, it's an explicit shot animator. I would consider making a much softer lighter version but LTX2.3 is very poor at even simple concepts so I just suggest that if intro scenes or similar mild content can be done by online services, just use them instead or the base model with low weight Eros delta loras. If you're talking about threesome or full body scenes, it also struggles with multiple females and characters in general due to the low resolution areas that usually come from those proportions.

noise and blur is a problem with distillation, I just added a DmD lora on huggingface but it's very untested. The other distilled loras are all on my huggingface too make sure they are correctly loading and working.

@shidoki95686 768x768 single pass or 1024x1024 - 1216x1216 upscaled ones

Usually an issue with a lora not working or the clip model. Make sure its using the same checkpoint in the clip loader with the right gemma.

that would be something in prompt but also probably the sulphur data, it has a lack of gay and male on male situations, that's been added to the next versions dataset though. You'd want to use LoRAs for it if there are some.

Thats a sulphur issue but make sure prompt has "vagina, innie pussy, shaved pussy" etc, something in it telling it what is there.

The models have the audio components and text parts of the clip inside them like normal LTX checkpoints. Gemma is still needed as the actual encoder but use the fp8 scaled ones, the full size gemma is oversized for that task.

I reworked the recipe with the DMD workflows for coming version 1.3. Now that I can actually output correctly with the model I can find it's bad base tendencies including that. It was impossible to work out what was the LTX distilled lora and what was the merge. If you're still using the distilled lora you're gonna keep getting that because that makes it worse. I reworked both merge steps with shape corrections and a different strategy, but previously v1 and v1.2 are made for explicit animations, and exclusively it seems like, they're not just uncensored LTX models that can do everything - but 1.3 should be more of a hybrid that can when used correctly.

All it can do is set it up and put her hand on it but you'd need a lora for the detailed motion. There's millions of data clips for that and there's still not a single lora for it yet. I'll do it I guess.

It's a pretty big change to make that version a director capable and sfw capable one, but uploading here means I need a bunch of previews and stuff so it lags behind.

@ArtiOtaku 1.3 is just a floor reference. It's the base model wrapped around 1.2 and reformatted, it only is used as reference shape in my finalization for merges so that it knows what shape is basically default. 1.3 is almost identical to base dev model just with sulphur data injected into it, but it's much less accessible than this version because I opened the gates slightly for this one.

That's one reason I pulled back 1.3 heavily. Using that with the posing-nude ltx 2.3 lora is they way to get those outputs.

Definitely don't pay to run any LTX models on runpod, it's terribly unstable and unreliable.

@MiraStudio It's the same, on local you're not paying that much to waste an hour trying to make some good outputs. Maybe if 5090 was cheap hourly but unless you're experienced with 2.3 using it it's gonna be annoying plus all of the loras and workflow setup have to be stored and paid for on a perm server.

@MiraStudio Training yeah but I mean paying to run comfy and generations is rough especially if you're just starting with LTX or the model in general and don't know what it does well or not.

1.3 is the opposite of that it's an experimental base model flip. It's not made to do nudity at all. That's why it's not uploaded here.

The DMD V5 workflow is the best working dmd workflow I have right now.

There are some on huggingface, I still don't know how to properly load them in the workflow. Most people use the conv int 8 base model and one of the lora extractions of Eros.

I use the DMD reshaped rank 256 instead now, 1.0 strength both passes with the v5 DMD workflow. It just respects conditioning better and the sigmas add a bit more detail and way less resample; https://huggingface.co/TenStrip/LTX2.3_DMD_Lora

If you're talking about the old V5 workflow, the image resize sliders were disconnected somehow, they should be wired into the first resize_v2 node's width and height but aren't. My huggingface has a newer dmd V5 workflow with it fixed.

The distillation deltas from Joy AI. I made them work outside that model and with the normal LTX workflows just with an extra step on first pass and with the normal distilled audio blocks.

@lavinia12345 It seems to have much more flow and structure than the distilled one, input image/video drift is also way better.