Have you noticed that when a wolf snarls you can see its gums but most AI models don't show this?
Trained on 689 photos of real wolves (no art, no werewolves nor AI gen), tagged manually and using GPT4 and joycaption to 275 tokens each.
The main trigger word is wolf so Just add this lora and prompt as usual and get - A Better Wolf™ -
Flux 2 [klein] 9b
Wow Flux 2 almost nails wolves right out of the box, but stuff like the teeth and eyes improve with this lora, snarl, angry, growling and interactions such as eating improve.
need a workflow? try this one which can handle the lora + 8 reference images
ZImageTurbo
This model is trained on the v2 dataset, in 6000 steps using default settings in Ostris AI-Toolkit. It started out great then at 3000 steps I got deformities which cleared up at 6000 steps. I'm pretty happy with the result. the fur improves from being ruffled to smooth more wolf like
Qwen
I've enhanced the dataset to a v2 using 689 images and joycaption to meticulously describe each scene. then trained I trained using Ostris new 3 bit quant and side chain "accuracy recovery adapter" . The LoRA is CLIP-heavy: it defines shape and features, not style. so load it onto the CLIP side of your workflow. i use a modified workflow from Aitrepreneur and can't really share it.
Flux:
We have 3 versions two of which works and the third is for creating nightmares.
Recommended is trained on 750 steps and gives a very impressive result
Light is lighter trained at 500 steps and can add detail to a scene
- I've kept all test images with seed = 1 and prompts in same order so you can compare the two models -
Bad Taxidermy can create good results, which is why I uploaded it but it turned out to be very over trained and give absolutely horrible images. so don't download it unless you need that.
Trained using ostris ai-toolkit
SDXL:
This LORA has been tested quite a lot now and it fixes a lot of the wolf related issues in SDXL
NB Training data included
Description
This version is trained with 750s and gives a significant improvement to the wolf image quality in terms of realism
FAQ
Comments (13)
Loved the very specific fix on the model! Looks good.
Thank you, I've updated the models to something more useful it turned out the first Flux upload was really bad.
This improved really fast, great job!
Thank you! It turned out it was
A: an issue with AI - Toolkit so i had to re-train
B: me over training the model. I've saved versions along the way so i did not need to retrain but I had to find the sweet spot. turns out it was around 500 and 750 steps
@Torque how long does the whole process take you? I would love to try it but don’t know if I have the time.
@DankGabrillo Oh man it's a time sink.
it took me 2 weeks to create the dataset from images. Captioning them using ChatGPT was a time saver and i had done the work already for the SDXL version. I also recommend taggui for tagging and managing tags
8 hour training overnight on my 3090, and uploaded the first model. then the bug in the AI-tools was discovered
8 hours of training overnight once more
excitingly uploading the Taxidermy version as soon as i wake up
back from work discovering that it is extremely bad and spending the rest of the evening figuring out the sweet spot for training
Knowing all this I would say the next time i will could train a new model using only 1000 steps and find the sweet spot. I could be all done in an evening.
that said, training humans such as celebrates seems very fast and only needing a few images. training complex stuff with multiple keywords seems to be very hard on flux compared to SDXL
If you are not up for training yourself you could try to set a bounty
@DankGabrillo hang on you are the one behind UncannyAF! That is an amazing model.
You know you can use SDXL loras on Pony if you bump up their weight to 2 or beyond.
@Torque Ai guilty as charged, though ye need a degree in aggressive prompting to really get something out of it. But yeah, thanks for the insight. Been thinking about using lora merges to finetune a model, but using something like krita to cut out the target for each lora and stick it onto an empty background. The idea being I could influence precise parts of the model without wrecking the rest of it. would save a ton of time on captioning too i guess, like you'd only have too describe the wolf for example without the background etc. It's just a theory atm, from your experience do you think it could make a valid lora if only the subject with zero other context was present?
@DankGabrillo Sorry for the late reply. either cut out the target, or use chatgpt to describe everything but the target in the prompt. the best image tagging software is github.com/jhc13/taggui which has a good offline auto image tagging for more nsfw stuff
Can you please include the trigger word "wolves" in your next version? thanks
the trigger word is already there. try "pack of wolves"
Seriously good stuff. I do love my wolves
ty continued updates.