This lora is a full all-round nsfw lora with furries as well, it's not exclusive to furries though.
This is the best T2V capable lora I have released so far, at 31 epochs with 21111 steps the model has been trained for a long time to get a good grasp of the concepts, the model might be a bit overfitted though, using a lower strength (0.8 or 0.9) can help if prompt following is struggling.
The model can do T2V and I2V, and technically some R2V although that's limited, characters will usually work.
Previews were generated with lightx2v distil lora as it is the fastest way to generate.
Concepts improved and introduced
Anthros (best captioned as "an anthro furry [species] [man/woman]"), aka furries.
Male/female nudity is improved, not quite 100% there, but the penis won't detach, and penetration and stuff looks better close up than the base model
Penetration/sex positions, the model understands vaginal penetration, blowjobs, deepthroats, handjobs, titfucks, most of the things you want will generally work if you prompt right, it's best to describe a position over just giving the name, that way the model understands it better. Cumshots sort of work, although they aren't really spectacular.
2d, 3d and realistic styles were included in the dataset, prompting for a "cartoon style", "3d animation" or "realistic video" will work
pov, side view, other angles. Generally angles will work if you prompt for them, including things like a moving camera.
Limitations and workarounds
Trained with the training adapter, it still can't fully prevent the model from undistilling, there are a couple ways you can run the model to counter the slight undistilling that has happened.
(fastest) With lightx2v distil lora (use low strength, 0.5 or even 0.2, 8 steps works), larryvrh's distil lora (link to converted comfy version) (might be even better) or another distil lora (test different strengths). High strength seems to have a chance to deform genitals more, reducing the model's capability of generating those.
(regular speed) Standalone, this generally still works like normal, although if you see noise or warping, try using the distil lora mentioned in the previous bullet point at a very low weight, still at 20 steps. Or alternatively, use a different distil lora. It seems to affect the generations less.
(slowest, but most prompt control) Use with ostris' training adapter, to unlock the ability to use CFG (around 3 works well). This allows for using a negative prompt, it's the easiest way to work around watermarks/signatures in t2v.
The main other issue is watermarks, as I did not remove them, or mask them out during training, some generations will have a watermark in a corner of a video, this happens more with the distil lora at a high weight than at a low weight, and isn't as common as in previous loras I've trained, but it still happens. I will look into what I can do to fix this without starting over from scratch.
I personally use euler/euler ancestral with the beta scheduler, although all samplers that work with regular H3 should work with this lora.
Side note
There are 2 ways to use this lora for reference to video, if you want to use a character as a reference, you can simply use a reference workflow with the fl2va model, with the lora added. The fl2va model has basic reference capabilities. Alternatively you can use the lora on the ref2va model, although it might not fully match expectations.
Disclaimer
This lora was finetuned off of Minimax H3, no alterations were made to the architecture and no safety filters were bypassed. This lora simply enhances the look and motion of concepts the model already understands.
Terms
You are not allowed to sell this model or finetunes based on this model. You are allowed to sell your generations made with this model however. Additionally the MiniMax H3 Community License terms still apply.
Description
Initial release, raw version (no changes made after training)
FAQ
Comments (17)
Ohhh yes, welcome back! Definitely eager to try this one since 2D, 2.5D and 3D with mostly furries and similar is my jam.
does this lora do the canine / knot penis well?
While it did have some vids in the dataset I don't expect it, not for t2v at least. It can barely do humanoid penises.
It would be helpful if you included a few non-furry examples - people who don't Focus on Furry might just skip it otherwise.
Yeah I probably should, I'll probably do that.
I did upload it but it got flagged for review, so did the original i2v examples video, I don't know how long it'll take before a moderator approves it, I don't see why they wouldn't get approved though, just guessing the automatic detection flagged something. Also strangely it seems the t2v is approved (doesn't have a warning) but doesn't show anywhere. CivitAI's weird, but if it gets resolved the t2v examples, i2v examples and human t2v examples should be up soon.
@mylo1337 can you share the prompt? would be great if these video including workflow meta data
@unheaded For the compilations can't really, they're multiple videos that I've combined into a longer video, the single video examples do have prompts included. Also workflow wise I used swarmui, but basically same as default workflow with my lora (0.8), distil lora (0.4), euler a + beta, 10 steps.
I trained on 1000+ clips in h3 with 90% of the data having watermarks and to remove watermarks if you caption them in the captions and then not mention them when generating it should 100% not show when generating. At least in my case. Do you mind if I ask: what trainer, what LR, and what dataset type (images, video, audio?)
21k steps is a lot, mine stopped learning at like 5k steps so maybe my settings just wrong. Thanks
I used diffusion pipe, a learning rate of 2e-5 iirc. Watermarks didn't start appearing until around 17k steps anyway, they are captioned so using cfg (so with training adapter added 20 steps 3 cfg) will allow prompting against it with negative prompt, at the cost of speed. Despite being captioned I did still end up getting occasional watermarks, captioning will not fully prevent the bias towards watermarks. In this case it depends heavily on prompt structure. Some prompts will never get watermarks, others will almost every time, afaik prompts following the prompting guide will almost never have watermarks, so generally those are fine.
The dataset consists of a bunch of videos, mostly furry with a couple non furry/realistic in there as well, and a bit over 100 pictures for training genitals better.
I'm also considering making a rank 1 lora with the sole purpose to counter the watermarks and possibly act as a baked in negative prompt, if I can find a way to track activations of the term watermark/signature I can possibly engineer a lora to counter it.
@mylo1337 thanks for sharing, maybe mine didn't progress enough for the watermarks to appear then. I appreciate all the info. Good luck in next versions!
@tazmannner379 The best way to prevent watermarks over the long run is to either completely remove them from all videos in the dataset, or to mask them out. However to fix it in post I'll have to do something different, like de-biasing the model away from watermarks again.
💕👍🏻
why ur loras are always the best. They dont pulverise over the aesthetics of base model and always play well.
I use fairly varied datasets to avoid overfitting and allow me to train more concepts and blends between. Basically since one of my earliest loras where I added some non furry videos and realized that both furry and non furry videos became better, and the nature of my loras is that they're basically multi concept mini finetunes.
@mylo1337 This is exactly why consumer models are also better than most people's loras. A lot of people just take old SDXL photos and throw them in loras that are for new models. It ultimately ends up just making the new model look like SDXL and takes away the quality of the new model. Using a varied data set allows the model to not overspecialize and become some inbred beast.
I've seen so many loras and their examples that instead of women or men having varied facial and body features, they all end up having the same body, plastic skin, and SDXL face. I understand wanting to be first, but...
I mean there have been several loras that embarrassingly put up "without lora / with lora" comparisons and the "with lora" examples all look far worse.