V2: This is a much better version that has better prompt adherence and plays better with other LoRas. The showcase has generations that are with and without other LoRas.
Prompt: axolotl
If used with other LoRas and you aren't getting the desired effect
pink frilly gills
round black eyes
V1.1: This is a version with about 3 times the steps as the previously published one. It's more responsive and makes less anatomical mistakes. From my previous experience, I usually choose the first epoch where the concept is consistently applied across different scenarios/checkpoints/lora combinations due to fear of overtraining, a habit I developed from my first few extremely overtrained models lol.
That's what I did for V1.0. The last epoch finished training and I tested it on my favorite generations/some that had mistakes that I didn't publish. I was surprised to see that it was even better at following the prompting and didn't make the mistakes the previous version did. After some thinking, this time must be different because this is the most extensive captioning I have done and the most selective I've been for training data. Flux training has differences I don't understand yet, but luckily all those extra steps gave it time to learn the captions better and it didn't get overtrained. The outlined below is the same, it just does everything better. Hope to see your axolotl generations!
V1: I have the workflow I used for all the images in the Axolotl Wildcards linked in the suggested resources. It also includes 11 different axolotl-specific wildcards 😂.
It seems like Flux also doesn't fully understand what an axolotl is. Just like Stable Diffusion, it just knows it's a pink salamander. Luckily, now we can generate axolotls on Flux as well!
I trained this model with 116 images I created of axolotls in a variety of different art styles. Each one was fully captioned, I had Microsoft's copilot analyze each one and then corrected any mistakes it made, one-by-one (because it was free lol). I think it was well worth the effort because it seems very responsive and I'm happy with the quality.
I won't be charging buzz through the early access model features. The only thing I want is for people to make funny axolotls and post them on this model page 😁 Thank you!
Description
V1.1: This is a version with about 3 times the steps as the previously published one. It's more responsive and makes less anatomical mistakes. From my previous experience, I usually choose the first epoch where the concept is consistently applied across different scenarios/checkpoints/lora combinations due to fear of overtraining, a habit I developed from my first few extremely overtrained models lol.
That's what I did for V1.0. The last epoch finished training and I tested it on my favorite generations/some that had mistakes that I didn't publish. I was surprised to see that it was even better at following the prompting and didn't make the mistakes the previous version did. After some thinking, this time must be different because this is the most extensive captioning I have done and the most selective I've been for training data. Flux training has differences I don't understand yet, but luckily all those extra steps gave it time to learn the captions better and it didn't get overtrained. The outlined below is the same, it just does everything better. Hope to see your axolotl generations!
It seems like Flux also doesn't fully understand what an axolotl is. Just like Stable Diffusion, it just knows it's a pink salamander. Luckily, now we can generate axolotls on Flux as well!
I trained this model with 116 images I created of axolotls in a variety of different art styles. Each one was fully captioned, I had Microsoft's copilot analyze each one and then corrected any mistakes it made, one-by-one (because it was free lol). I think it was well worth the effort because it seems very responsive and I'm happy with the quality.
I will be updating my axolotl wildcards page right after posting this model, it also includes the workflow I use.
I won't be charging buzz through the early access model features. The only thing I want is for people to make funny axolotls and post them on this model page 😁 Thank you!
Final note: I also had the trigger word "axomodelxl" but it doesn't seem necessary at all. Maybe if you use multiple loras and you can try that out if you're not getting the results you want. Just saying "axolotl" has been enough for me.
All the images in the showcase are using the lora at strength 1.