Yeah I'm committed to this one that it won't use any on-site loras so that there's no need to worry about overlapping into it. I'll train patch loras for it myself, but we still have RL training on 768 accuracy data for this for anatomy clean up and detail.

There's only LTX 2.3. 1.1 is their update to the distillation which still needs to be updated as I've clearly exposed it has major flaws.

@kronos1959777 what was weird though was I got much better results with distilled 1.1 lora so I reranked that one. Overall they aren't transparent about what the distilled is, but it's probably the usual teacher training. I think they train it on T2V latents at 30 steps so it learns to accelerate latents. Ruins I2V because it wants conditioned I2V latents to look like T2V, also rank 384 is insane. I think they thought it would work both ways but it doesn't. Gutting it like I did helps alot.

@FantasticFever I haven't actually found a setup it can't do. It can even to hentai tentacle stuff. Actually, the hard poses are female POVs. It tends to want to turn it into a male POV threesome.

Don't need to, but when comparing that new distilled lora to the other large distilled loras used at 1.0 strength first pass you'll see that the other ones reduce the nsfw motions a lot and ruin anatomy.

v1 fp8 is a transformer only diffusion model yeah. Beta was just hard cast to fp8 which is why the audio is very flat. This fp8 should have better audio since it's mixed precision. You can offload both vaes and text encoders since they're all split up and it should be able to run on 16g cards or even lower with more offloading.

@ArtificialOtaku well now I actually uploaded an fp8 checkpoint version to replace it. So you have the older diffusion model version. You might want to redownload the fp8 updated one now and try that one too.

I mean I'm waiguoren and know 0 chinese. I would say go to our discord but discord has no auto translator for some reason. Maybe chinese users can gather around it and form their own group and reach out if you have questions.

That was kind of the point of this, through weeks using it in-training I put together this merge which basically hit that perfect balance imo. This was out of 17 versions and was even down to tweaking weights -/+ 0.03 and this version still had the best sample outputs.

50 steps is kinda crazy for distilled. That many steps will cook the audio for sure. The max steps you need on distilled is like 13 sigmas/steps.

It's on both model pages, https://github.com/TenStrip/10S-Comfy-nodes

I've heard that a lot of workflow makers are trying it now they'll have stuff updated soon.

@dav79mail156 The edit-anything does a different and much better conditioning but the trick is also that the upscale pass is then forced back to a high strength uncompressed i2v conditioning.

It's hard to be clear but it is LTX2.3 dev straight up, you can probably take ltx2.3 dev seeds and prompts and get almost the same output. However if you use it you'll find it has anatomy capability and sexual motions the base doesn't have because I layer scale merged in Sulphur2 data to make it behave like that. Sulphur official is a powerful full trained version where as this is a custom mix that uses it.

Make sure you don't load fp8 model as actually fp8 inside the loader, leave it a default so it casts to bf16. But I will say that for some reason the fp8 dragged down the overall 'soundscape' compared to bf16. The bf16 will actually layer multiple sex sounds and voices and it's just harder to get with the fp8 version. It may have to do with the clip being the normal default text projections and instead you need to pull text projections from the merge. This is definitely a known issue though.

Try the new Fp8 should be fixed.

@ArtificialOtaku Yeah and that's the version on here now as "fp8." Transformer version is on HF still, some people prefer it.

@Goblintide I had that fixed for me by silverco1n on the discord credit to them.

@wzrdking something like- In her cute aroused hispanic voice with thick spanish accent she talks flirty Spanish dirty talk, saying: Ay papi, you fuck me so good.

which version? If it's old fp8 that version can have issues with some pytorch versions it seems.

@SantanaiR34 New one is a correct checkpoint version again so it loads like beta.

They're testing it now I think.

@K0DEX Thanks to Sulphur by Fusioncow he made the data.

I mean maybe I oversold it, the problem might've been unique to this because it's very different than base LTX 2.3, but if you're only doing nsfw you're probably better off. The motion blur issue was 100% because of the distilled loras though and thats for both models.

@ArtificialOtaku this model can do SFW, avoid nsfw prompts and things like "sexy" or "nude" and add some anti-nsfw negatives with NAG node if needed.

https://github.com/TenStrip/10S-Comfy-nodes

@waifuliberator yeah but there's like 10 links on every info page. they're barely necessary and only for the middle pass really but sometimes with anime stuff the latent sharpener can be useful to return crisp linework to the upscale.

There will be sulphur loras that shouldn't do that, there's already a few. Weirdly you actually don't get that from older LTX2.0 loras at low weight at leastr when I tested mine. Most loras low weight like 0.3-0.5 can be used to help reinforce a motion.

Those are here. https://huggingface.co/TenStrip/LTX2.3-10Eros_Workflows Right now in how I understand it it's the way that LTX works. The latent has anchor points in a left to right flow, that conditioning for i2v/IClora is carried only from the first one onwards. If you do transitions, camera movements, expressions and talking, the future anchors will lose more and more likeness as those motions change them and they do not reinforce themselves to match any kind of "likeness" That is something that needs to be made and integrated into the model.

Only way to do it is to have a last frame Guide with the same character and face so that there is data on both ends for it.

@pirijir875978 just remember to add 'LTX crop guides' node on the final video latent before the VAE decode node if you get a slideshow at the end of the output.

I attempted a gguf conversion but it has tensor dimension errors on load because of the merge but anyone who can manage to scale it without messing it up and make it gguf can go ahead and add it on huggingface.

I never tried them but for I2V I would assume they will make it worse. For T2V they may be okay at half weight.

Those are in these nodes. https://github.com/TenStrip/10S-Comfy-nodes

Which version? You shouldn't need any loras with this version really at first, try to see what it can do already. Most loras will make it weird.

@Sgt_Fartz Yeah it is better than 384 but the 72_condsafe ones I linked are even safer at least for first pass at (0.8-1.0). Also faster. https://huggingface.co/TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments/tree/main

Yeah add it on huggingface I'll link to it. Not having the good distilled loras kinda sucks but better than no working model at all I think.

@MisticRain69 its probably because how it does distilled especially if it uses the 384 distilled lora or the other ones. They mess the model up usually.

I've seen someone suggest to not use " " with dialogue. She says: oh wow! He says: cool...

There's a ton of 3D stuff? Try all kinds of styles, "blender porn SFM style", "a 3D porn video of:... " - Also because of LTX structure the prompt becomes a timeline that flows left to right, tags would have to go into an auto-interpreter that makes the prompt from them and that's doable already.

Yeah I tried and failed as well no worries.

She says: she says her lines. Then add a return (paragraph) and start the prompt, or describe a motion to break it up. Use punctuation and also prompt enhancement you end up with issues like this with manual prompting.

If you look at video model training, or specifically DiT models, it's just like that, but with different data that isn't cats and random videos.

@itsreallytrue Depends on if it's T2V or I2V Lora. I2V can have lazy captioning, you mostly only want to describe loose context; camera angle and figures, then describe motion as cleanly as possible without embellished descriptions, and also use the exact same phrase to describe the motion acting as a keyword. All the extra stuff becomes noise tokens that unwind stuff the model can do better. T2V, idk much.

Yeah I use Grok for images, prompt, and start videos as well. Hopefully they see the commotion though and decide to fix their odd approach of censored fake nudity.

Doesn't load it all at inference. Fp8 is much better for gen time bf16 is a dev usage.

This is doable with prompt and image, my workflows with 13 sigmas help alot with keeping the mirror physics correct too and it tracks as she moves. "Realistic solo video of a solo real woman with long blonde hair holding her cell phone in her hand and filming herself in a mirror facing the camera. She...."

I'll look at that. I'm working on improvements to it right now including solving the vertical aspect color shift with nodes.

@ddng533 I tried to alter latent timing with nodes but it doesn't work. The most the model will do is around 28-30 fps it only knows how to do motions at 24 fps, 25 fps is actually a little bit better sometimes. You can interpolate 25 to 48 nicely.

It's not supposed to with the temporal inpainter, but if you delete that node it will. I got a replacement for those workflows already that middle sampler isnt really needed.

I'm not sure why some people have this error. Updating LTXvideo comfy extension would help and make sure that the edit-anything IC-Lora is in comfyui/models/loras. You can put a LTXVImgToVideoInplace node instead and put the pink latent string through it with the same image and red vae hookup, set it to 0.75 or more/less depending on how much you want movement.

Don't connect the audio to the video combine from audio vae decode to mute output is one way. Instead of bypassing the audio you just set the first empty audio latent to 1 frame. Or you can take 10 seconds of silence as a sound file and encode it with a latent mask at 0 strength and put that in to force the model to generate silently maybe.

The fact that the normal 384 contained feed forward and attention layer influence makes no sense, but that only became apparent when using an actual fine tune of the their model and seeing it directly counter prompt and data it wasn't distilled on. Also from 384 to 72 and it's still a bit strong at 1.0, might've been a little overkill by them.

@Dnai100 Sounds like distilled lora problem. If it is just a cloudy blur with just audio noise that's distilled lora.

@likebot97 The large 24g file is a model not a lora.

Biggest weak point yeah. Data is actually under trained and generalized and thats a very distinct and complex motion. I'll train my oral suite on it to see if maybe 768 and some 1024 short clips can get it to go past upsampler. The best lora to use for it is penile praxis v4.

Cum shots are weak data from just a generalized training. To be improved after pruning and more training.

That probably comes from it being a little ways towards being a different model. If you wanted to mix them you'd want sulphur lora 768 + that on dev model or retrain it on sulphur_dev bf16. https://huggingface.co/SulphurAI/Sulphur-2-base/tree/main

I got this linked to me, this is what you want I think. https://huggingface.co/adrepale/LTX2.3-10Eros-LoRA/tree/main

I've gotten it usually, it's not "zoom" more like "camera slowly moves in towards..." There's a lot of wild camera motions and transition data being pruned out of Sulphur for refinement a lot of that interferes but high frame conditioning and prompt can curb it.

she speaks in a suave British English accent saying: oh hello, my dear.... In my WF you have early step CFG which solidifies audio prompts like that I usually get the accent described.

Distortion to me just means the distilled lora. Oversaturated and changes the look of everything. It's why I use undistilled and with a different distilled lora that has it's weight split, you need much less weight on upscale pass or it will resample and ruin it.

@plz12345 I said just the undistilled model, you have to handle the distillation at inference with lora because the distilled lora they released does much more than let you run 1 CFG and low steps, it adversely effects pretty much everything. Just merging at 1.0 to make the 'distilled model' will make it unusable for best quality.

Yes it can handle mild female self filming, just dancing stuff, it can probably do the same output as default LTX2.3 if the prompt is basic enough. It's not overinfluenced to the point where people start thrusting or turn nude on their own. As far as pole dancing goes that's a very hard thing for LTX and it probably needs a targeted lora, although in that stripper video in previews I got a few motions with it but it won't do spinning moves very well she will clip through the pole a lot.

Not sure what that one does but 'creepy' sounds like prompt and over-ranked distilled lora usage-not optimal. My WFs are only designed around using this I2V and max quality and MrXin ones are good for general and fast outputs.

@KlaudM Heretic clip is bad and it will ruin dialogue a lot and make people sound drunk. That setup seems overly complex all you need is V1 inside a checkpoint loader and you can get both VAE's with just it like a normal LTX model. Then for clip handle that by loading it like default as well with the gemma_3_12b normal clip - fp8_scaled for quality or fp4 for memory saving.

@KlaudM If you're using tiled sampler on a landscape ratio like 1920x1080 pass, or T2V you should hit the bypass mode on it. It's for very large square and vertical ratio pass that are I2V conditioned.

Does drawthings do LTX at all? It would need a program-side mod or code change probably to run it correctly.

@aphoticroots advantage with drawthings is their metal quantizations but I assume the presets might not work with LTX models.

I ran no-prompt i2v tests with IC-Lora on and off and it significantly affected the output away from nsfw which booted it in my book. I think I found that the main benefit was having image-guided condition from the IC Guider node so I just put that in. Likeness drift needs more of a direct solution that I'm working on.

There is subtitles issue but in my WF it uses CFG and negatives to prevent overlays like that. Not sure how Chinese is written for dialogue but quote marks " " can help to cause it you only need - she says: speech. then he says: speech.

Are you using the old fp8 transformer? Use the newer fp8 if so. The audio is usually motion dependent, more motion makes more audio.

I haven't tried them yet I'm still working on nodes. It would load in a gguf loader but not sure how you pull the VAE and Clip off it.

@steezo_jones 3090 is perfectly adequate to run fp8 version.

@Kingp0dd529t yeah but just not the quantized versions. all of those are compatible.

If this happens after update you have to recreate the node. Some recent update to Comfy-UI has made nodes not auto-reset like they should this is happening on a lot of workflows.

That's not my workflow that's his. I use mine on my huggingface. Distortion and smearing is the distilled lora and my workflows split it up to change the weight of it between video/audio for both passes. You need at least 0.7 for first pass, my WF uses 0.8-0.9. But the model is at the mercy of being LTX and having to deal with that terrible distilled lora.

Just use fp8, bf16 is just for training or dev work fp8 is perfectly fine for generating theres very little output difference.

@ark_5 The seems like a clip or model load error make sure you load a normal working gemma3 version and have the fp8_learned or bf16 loaded correctly.

Use penile praxis v4 lora to help. I was surprised how good it as even with penis on it's own, that was not guaranteed and it still needs more training.

@delta45424155 No praxis has targeted highres data that helps it draw better, just at 0.3/0.4 to correct when it will disfigure it on some prompts.

LTX can do much more complex stuff because it has a very long available frame count but it's not consistent or easy, requires a ton of prompting adjustment and manipulation. As for sulphur or this model Eros this is probably the most generalized nsfw video model right now, even uncensored version of grok lacks a lot of concepts like anal and certain positions.

I heard it was adding more lesbian, but still isn't trained enough to get to the individual concepts clearly. Future version should but that is also more of a motion for a lora to be used for.

This is an ltx2 problem the model gladly redraws the face at certain anchor points on the latent and treats it as just a loose reference especially with camera change, dialogue, and expressions. I think theres 3 issues; it's not a dedicated i2v model or structure, the feed-forward and self attention is especially compared to everything else. My nodes dampen it quite a bit with attention layer anchors but the problem is deeply embedded in the model's core behaviors.

@popeyer The real solution is to make an end frame with an image edit to give a future target of the same face/character and then the model won't drift off. The more usable way is to train a 15-30 minute character lora on images of the face to actually inject it into the model, there are some workflows being worked on to do that.

you mean workflow? thats on my huggingface links are all in info and I pretty much gen everything with those sigmas and distilled weights.

Grok or one of my SDXL models with pretty much everything I've uploaded recently.

@civitaicomsank205706 Yeah it's an anime structured model with realistic outputs. For anime just that new anima model.

Civit took that as their own from huggingface. They need to contact me directly with assurances before I add this to online gen. It would also need to use my i2v workflow for at least a guaranteed somewhat decent outcome, but I wouldn't pay any money to gen any LTX2 stuff it requires constant tinkering and quite a handful of bad outputs to actually get something. I also don't have commercial license and I don't know how civit handles that because they have ltx2 commercial license but I don't.

@Silicon_Mirage IF gen times jump up to multiple minutes on distilled you've maxed out VRAM. Use smaller model, smaller size, cut loras that aren't needed, close other programs like discord, and use a lighter browser for GUI. And use fp8 not bf16, don't load additional VAES use the ones inside, and use a smaller clip gemma.

Yes and all the other LTX conditioning tricks as well, it's the same model. That will probably have better consistency results for character likeness that way.

@Silicon_Mirage Sure it can. All of that sounds like workflow and prompting issue. 8 step workflows are going to have very inconsistent low quality movement my workflows use STG and 13 steps first pass you can get good complex movement but use the right enhanced prompts, for longer generations you can use time stamps as well ( 7-9 seconds: she moves this way)

Yeah if I had a 6000 or something I would do undistilled, with CFG you usually want 50 steps.

I didn't train it but I merged it differently and that'll be passed on to them for official versions. Soon I'll start training high rank data patches of my own for the model and see if I can improve it further.

Talking about chunk feed forward? I never used it, the only major impact on reducing motion blur was reranking the distilled lora and having more samples. The rank72 lora I did reduces motion blur by a huge amount compared to using the 384 at high strengths. One way to reduce fast motion blur is with temporal upscaling and getting more steps to render it.

@McClippy Yes. The rank 32 one removes it even more but it seems like it's a bit too weak to actually sample so I'll try 52 next time. Raising base frame rate to 25 or 26 also helps.

Are they stuck in a forever "analyzing" loop? Or does it just show the warning triangle? The analyzing bug happens on big files, the file size cap is more like 50mb now anything bigger takes forever to clear through. The warning triangle is bugged it'll show but it won't actually be hidden. I can see 12 vids on your page, last one 8 hours ago.

Wan2.2 doesn't have major finetunes because no one stuck the 10k+ up to train one and the license was restrictive so they all trained 2.1 instead I think. As far as I know MoE doesn't affect training at all, and they also said they were increasing tuning support. Results should be better since it works to isolate only the relevant data to prompt. However I don't have a ton of faith in Lighttricks at this point because they acted like Ltx2 was ground breaking but they undertrained it on a terrible dataset. Unless they've fixed their dataset it's going to have the same weaknesses and problems just less weird prompt feedback. Main issue now is it's going to take another 10-15k to get that version of LTX even to this point so we need to find crowdfunding or angel donators to fund that new run.

Yes, that's something many have attempted loras for. I think architecturally LTX is incapable of matching a model like Wan 2.2 with such a complex concept, at least on this version. The new LTX version can probably do it.

Both models don't need a vae, they have the audio VAE on board you can load them in the audio VAE loader.