Vodka by FollowFox.AI
Vodka by FollowFox.AI
- TypeCheckpoint
- ModelSD 1.5
- TagCHECKPOINTanime base model fantasy midjourney midjourney v5 stylized woman
Overview
TLDR: We are launching Vodka_V1 by FollowFox.AI, a general-purpose model fine-tuned on Midjourney V5 images. And we are sharing all the details on how we created it on our substack. Our initial version is already quite fun to work with, and thus we decided to release it. We want to continue experimentation here, so please share your feedback and expect further improvements.
Please note that model is not particularly strong in photorealism, so if that’s your primary goal, you might have better luck with alternatives. It can generate some, but it takes effort, and they have a very specific style.
Outside of photorealism, we have been getting interesting results across all topics and variations.
You can read all the details about the model training process on followfox.ai (link to the post), as we can embrace the open-source nature of this community. You can recreate the process, see exactly how we got here, and provide feedback and suggestions on individual aspects of the protocol.
Parameters that seem to work well
We did some testing but hoped to hear more and update this based on the feedback from the community:
CFG ranging from 3-6 work well. For realistic human generations, lower CFG works much better, indicating over-training. For more creative generations where realism is not needed, a CFG value of 7 also works well.
Low step counts (0-30) seem to generate less interesting results, but starting from 40 and to 150, everything seems equally worthy.
If a prompt is overly lengthy, lower CFG values seem better, or otherwise, images look fried.
Almost all samplers seem to generate interesting results.
Next Steps
There is a ton to try and improve from here, from a few long-hanging fruits, such as lowering the learning rate to make the model less fried to experimenting with captioning images, segmenting by different classes and weighing them by the current weakness, cleaning the dataset, using more advanced training protocols, so on and so on.
For now, we will release the model and observe what happens. If we see some traction and interest, we will continue iterating. So please share your thoughts with us!