Introducing my first LTX LoRA release, adding inches to bustlines anywhere LTX 2.3 dares to go.

Important: As always with my models, this LoRA does not represent any real person in any way. If this model produces a video with any resemblance at all to a real person, that is purely an accident of AI random generation.

Update: I found a combination of settings that very much improves the look and consistency of videos with this model for both T2V and I2V. See below for details.

Based on an entirely new videos-only dataset created using the latest Busty Women Wan model, this LoRA for LTX 2.3 works for both image-to-video and text-to-video. Although the model is primarily created to provide bustier women with bra sizes pushing into the middle of the alphabet, it works equally well in I2V scenes with women of any bust size. The model works very well as a general-purpose motion video for female characters, as illustrated by the variety of shapes and sizes in the samples above.

All of the samples include full caption text so you can see how everything is made, and I2V vs. T2V are labeled in each sample as well. I've also tested applying this model to existing LTX character models with very good results, example videos to be posted momentarily.

My LTX workflow is nothing fancy, based almost entirely on the sample workflow provided by ComfyUI for T2V and I2V with LTX. I use the latest distilled LoRA (version 1.1) and I tested with a variety of different samplers and LoRA strengths. The Busty Women model works best between 0.75 and 1.00. You can go up to about 1.25 with mostly decent results, but above that and the output is almost always distorted. I liked my results best with 1.00 for T2V and 0.75 for I2V. I prefer the LCM sampler for the first pass, and for the second pass (upscaling) it doesn't make much difference which sampler you pick.

Settings and workflow recommendations:

• If you use comfyui for LTX 2.3 and you still have the node called "ManualSigmas" where a series of decimal numbers are listed in two locations in your workflow, you should replace the first of those nodes with the "BasicScheduler" node. In the first part of the workflow, you set the scheduler to linear_quadratic and 8 steps, this duplicates the numbers from the original node. In the second part of the workflow, you can leave the manual sigmas node in place with the default numbers. Using a scheduler node here adds denoising ratios to the process and that can add static to your audio. It works best if you leave that second node the way it comes with the template.

• The BasicScheduler node includes a "denoise" setting which affects how strong the effect of the rendering is compared with the random noise (or starting image) that you begin with. First part scheduler: 0.80

• These settings work well for both I2V and T2V and they will essentially eliminate all of the artifacts that I was seeing sometimes in my videos with extra nipple bumps. It also drastically improves the look of naked breasts with this model, making it safe to do topless videos too.

• My preferred samplers right now are: first pass LCM when I want speed and res_2s when slower is ok. They give very different output so experiment with both. Second pass I use euler_cfg_pp for great detail in the finished video.

• Enjoy!

The good:

• Really great image-to-video performance, the motion is very realistic.

• When the character isn't moving, the model does not add any extra motion. Motion is only added when appropriate.

• Works seamlessly with other character LoRAs.

• Does not have visible effect on faces when added to a character or starting frame.

The bad:

• Naked breasts don't look as good as I hoped. The nipples and sometimes other anatomy isn't great. I'd like to fix this in a future version.

• Sometimes extra nipple bumps are added to clothing. When this happens, you can often correct it by using a different sampler or reducing the LoRA strength, although this can also change other parts of the video.

• My dataset videos were all square videos from the neck down in order to prevent having a face bias, but this means the output will often show the character from the neck down. This can usually be worked around with extra prompting for face details and camera angles that encourage the face to be in frame. I'm not sure how to do this differently in a future version since I don't think LTX training supports an alpha channel to censor out the face from the dataset videos. If anyone has any suggestions, I'd love to hear them. While I generally prefer the output from the LCM sampler, that sampler does seem more likely to give neck-down output. Switch the first pass sampler to something like dpm++_2m_sde ("dpmpp_2m_sde") or res_2s if you have this problem.

I hope you all enjoy creating videos with my first LTX model. It isn't perfect by any means, but it's pretty good, and lots of fun to play with. Please tag the model when you post videos because I can't wait to see what you all can do with this LoRA.