https://github.com/chrisgoringe/cg-sigmas

The next version of this is gonna be way better--better and more data and actually 2.3 trained.

@jeremiahomolewa56846 Only at 1800 steps last update I got. Aiming for well over 15k I think but also needs an i2v training run.

Sure for simple things it can continue and evolve motions but longer stuff usually needs a lot of prompting. If the motion doesn't work that good maybe lower frame rate to compensate for the added length, and interpolate it back up afterwards.

I mean every example I posted is i2v, so maybe.

Any image2video workflow can be turned into v2v extend. My workflow has a video loader node all you need to do is set frame rate and cap and then use that video instead of the image on the latent combines. I usually set the initial video frame load cap to 48 or 73 frames and then you get a 10 second extension of that, works great on nsfw videos or anything.

I've never used wan2gp but next version I'll see about compatibility and what I can do to integrate for it. That specific bug is pytorch related though something isn't using torch correctly system or program side, or a dependency issue. See if normal ltx2.3 dev undsitilled works first, if you can get that working this will just replace it.

@ConnoisseurOfHentai Idk if this is better maybe https://github.com/deepbeepmeep/LTX-Desktop-WanGP

That will definitely be the future in a few weeks/month when I can update it.

Probably related to the prompt. If you don't prompt anything for audio, background noises, or dialogue ltx2.3 will just fill in with whatever is related. The audio in this is nearly untouched from the normal 2.3 model besides some slight influences for female moaning and sex noises.

News is that it's training and still early on. Initial results show much better learning especially genitals and movements. For i2v it's probably already usable but I can test anything for 2-3 weeks.

@Honeyphoria Audio quality to me means the tone and overall sound fidelity, but idk if some people are using 'quality' and referencing just prompt adherence and available noise variety. The next training will have full sex noises and sounds along with the JOI and ASMR type stuff and orgasming voices and all that. They were already present but I toned down the audio layers in this version since it's 2.0 trained and it would drag the 2.3 model backwards on most of the layers. For audio fidelity theres two major things people miss, one is that if you pass the first audio latent into the upscale pass like normal-and that usually uses a lower strength distilled lora, the audio will always be completely ruined because it's being resampled too lightly. Inside my workflows there is a group of nodes that take the first pass audio and manually reencode it, providing the exact sound output as an audio-video layer to the upscale pass instead. This is how I'll always do it because it's just way better. From there you can start changing sampling and steps, different samplers produce way different audio outputs. My sampler swap workflow uses SDE sampling with 3 CFG for the first couple steps which provides a way better initial audio track, but I still didn't get it tuned enough sometimes it goes overboard and it's loud as hell.

T2V: yes, much better when you use 2.3 loras for the better anatomy as well. The new 2.3 training is approaching 10,000 steps with a high batch. The full version of that will be released by it's trainer and operator. I will find a training step and weight mix from the new run that I hold to be the best for i2v use and update this, then further updates will be trained/tuned by me or made from merging other loras in.

@yuduz367 https://github.com/chrisgoringe/cg-sigmas

@yuduz367  I'm not knowledgable about optimizing and low memory use. I run a 5090 with no real optimizations.

@yuduz367 It's an I2V model, it's meant to provide a better base model to animate pornographic images than the default LTX2.0 model, and it has a wide ranging variety of nsfw data inside it that does better than the base model. I still see people using the base model with nsfw loras and wonder why; this version of 2.3 is always going to be better for I2V and provide much more variance. I merged it in at a conservative level so that the audio wouldn't be interrupted as much and so that loras can override and shine through still without being disturbed by overlap. The actual sulphur project (not mine) is the full strength T2V model that will be released. I am basically just doing my own personalized I2V fork of it with permission since I'm involved with it.

The script auto converted it and I don't see any gain to leaving it at the full dev size for what it's supposed to do, but the next version I may leave a full size one available.

Not sure if you're talking about T2V or I2V. Most of the tricks with I2V to improve the output structure is usually with more sampling. In the workflow I posted the attempt I mad significantly improves evolution and structure with the 13 steps on the first pass. Important as well: anatomy distortion and mutation is also super prevalent in the upscale pass if you don't set it up correctly which is why I like previewing the first pass, especially with vertical resolutions, still. The model could also prone to that since it only had smaller resolutions on this first training run. Additionally for I2V; adjusting the constraint mechanisms can allow latent to evolve the image more both the preprocessing and conditioning constraint. Resolution is also super important there aren't many models that have small details, if you can you want to crop out most of the start image straight to the focus and action and cut most of the framing and background out that isn't needed. This first dataset itself also didn't receive enough I2V training on the first run to properly do complex things which should be better in the next version but it might not be enough to patch weaknesses in the LTX2.3 model itself.

Not to me for this project, I only train my own loras but will be adjusting around what's missing from the next version-which isn't much from what I've seen. The real sulphur model team that is doing the training will appear on here at some point and post the model, it will also be on hugginface and they may take donations to recuperate for possible future versions. It most likely will be imperfect since it's a 100k+ dataset but the I2V will be very good.

He nuked himself from what I saw.

@MisticRain69 something on reddit set him off, he got "one-guyed" and just said F all of you I'm out.

Turn LTXPreprocess img compression to 15-20 or as low as it will go until it just stops animating and make the i2v conditioning strength 1.0. Then increase the compression until it does just the right movement without introducing new elements. If you use my split sampling workflow then CFG is actually on for the first frames, so prompt all the undesired stuff in the negative as well.

A few ways: Make the audio latent 0 frames, or 1 frame if it doesn't let you and then skip 1 frame when you combine the video. After decoding you can also only save the video and unplug the audio from video combine. You can use a volume control to just make it silent.

This is just LTX2. You need to prompt some kind of audio, usually very descriptive. No audio prompt: the model will default to either random white noise or music.

the one in the checkpoint or the ltx2.3 vae.

@aardbark77431 blurry is a workflow/resolution thing. Fast motion will not be very crisp usually and it's hard to get it right with sampling and interpolation. If it's extremely blurry is a workflow related issue. For i2v the i2v_preprocess node literally blurs the input image so you want it as low as possible to still get an output with motion.

@NUGGZ1616 that sounds like some issue with the distilled lora or strength then, or possibly the way that i2v conditioning is set up or too much img_compression.

@aardbark77431 The video and audio VAEs are in the checkpoint. At least it does work in comfy like that. KJ has split ones off though here https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/vae

@NUGGZ1616 it should work good as a dev checkpoint that you use distilled lora on top of yeah, that's how I always use it. The new 1.1 distilled doesn't seem better but I use it first pass only at 0.9, and then I use the old ranked down distilled lora on the upscale pass at 0.7 strength.

The keywords can just be used to target types of motion or ignored but better to match the style, just need the word somewhere in the prompt. Punctuation doesn't do anything its ignored by tokens, but just use normal natural language. LTX loras would only need keywords in t2v, i2v doesn't need them just use lora strength.

That's definitely fixed in the new one. I can already add, change anatomy, do pull-outs, and camera and other reveals on it's own at a decent level. A lot of training left I think too.

2.3 is in training and already I think past the point where it's better. I just had my first hands-on testing. Actual anatomy is pretty learned and changes to poses and advanced prompting is going to be possible and audio is way improved.

Yes but a very specific mix with different training steps mixed together first. You can make a similar one by taking the old sulphur lora and using the script but it'll be different. I used the steps that had less t2v and more i2v training. The new mix in testing is a lot better and has all reasoning loras along with the actual 2.3 training run.

@N0n4m3 pretty much always because of the literal difference between loaded vs merged, but not significant per-seed with this merge. I've been changing the merge script into something completely different that uses Layer-Wise merging and full bf16 with actual critical layer highlights and protection this time and designed for 2.3, it's not quite phr00ts script anymore but mainly it's the actual 2.3 data making it work so much better.

It's a newer distilled lora. The distilled 1.1 lora is good for the first pass, but seems to include color changes and counters anatomy on second pass. The reranked ones are going to be better for this and i2v: https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/loras but I find that using the old distilled at weak strength on the second pass is still better.

not sure what that is, wrong page.