see sections below for Anima, LTX2.3 &amp; SDXL<hr /><h3 id="anima-base-1.0">Anima Base-1.0</h3>Use _f31ic1a in the prompt.Trained at 1024x1024 resolution using the same dataset as the sdxl version but using diffusion-pipe. training used ~10gb of vram without optimization and takes a bit less than an hour for 30 epochs.<hr /><h3 id="ltx2.3">LTX2.3</h3>character lora for LTX video 2.3. Initially trained it for t2v but it works well enough for i2v. It's my first LTX2.3 lora. it will produce inconsistant results with other loras and at irregular aspect ratios. Trained on dev but works better on distilled, not sure why.Use "Felicia" in your prompt's sentences. Strength from 1.0 - 1.3. t2v works surprisingly well, but you can generate better quality nsfw scenes using i+t2v because LTX2.3 is hard to get to generate nipples and genitals without i2v. I found increasing CFG and lora strength helped reduce noisiness when motion was a bit blurry. (e.g., situations when she is facing away or the camera angle is not straight on). The male influence seems overfitted, so you may have to fight a little if you require a specific looking man in your scene.7600 steps, 52 videos clipped from the anime and captioned manually. ~2 hours to train using RTX 6000 PRO with LR 2e-4. trained with audio.resolution buckets: 384x384x25;384x384x33;384x384x49;384x384x73;384x384x97;384x384x12126/03/2026: Hi! :)) I just today uploaded the training videos and captions used for training in the training data. It's quite time consuming creating your own dataset for LTX-2.3 training so I thought releasing it may be helpful to anybody interested in trying. happy training<hr /><h3 id="sdxl">SDXL</h3>I uploaded a multi-character Hyakuren lora before but I decided to create just this individual character model since it's easier to create single character sets with and more flexible for changing outfits n stuff.From my testing using different checkpoints, I've found it's not very compatible with checkpoints other than WAI-Illustrious-SDXL so I only recommend using it with WAI. Experiment with the settings and find out what works best with your setup, but I've had the best results with these: Resolutions: 816 by 1232 or 1232 by 816 or 1120 by 1120Lora Strength: 0.85 Sampler: Euler ACFG: 4 - 6 Scheduler: Karras<hr />additional: I recently made a nsfw detection model which is more accurate than old ones we use regularly for detailing. please try it out if you're interested and leave some feedback!

Files with SHA256: 47838a7ea021666ecc302a83ecd30301e1da941ede3bc3b87014997e32c9a59c