Pal #094, Felbat
The one-eyed, one-horned flying pal and people eater. Sample pictures are all cherry-picked raw gens to demonstrate the model's capabilities in pure text2image. No inpainting or upscaling was done.
Prompting
Weight: 1.0 produces perfectly good results. You can go down to 0.9, or even lower, for more flexibility.
Trigger words: felbat, semi-anthro and felbat, anthro
Canon beauty: felbat, semi-anthro, ambiguous gender, mouthless, armless
Negatives: bat \(animal\) because the word "bat" always adds a bunch of little friends (base model issue)
Also does taurs with some difficulty, and may occasionally be able to produce a quadruped. Not sure why you'd want a quadruped felbat, but there you have it.
Reminder Tags
In harder art styles, or more complex situations, the model might forget certain stuff. I've compiled a list of the tags you can use to fix your felbat.
armless- Arms tend to show up when you prompt for the things they're used for, like gestures. This will fairly reliably get rid of them and at least sometimes try to perform the action with the wings.mouthless- Likewise, especially if you're prompting an expression, it might try to draw a mouth. This will strongly discourage it.black body, white hair- Sometimes using Danbooru artists gets anime girls instead of felbats. This fixes most of them.hair over one eye- If it tries to show both eyes, this'll cover one. They do have two eyes (I checked!), but they're typically depicted with one completely covered.cape- Especially helpful if you're trying to dress the felbat up without losing the wings.
Other Useful Tags
wing hold- It's a small tag, but I think it does reinforce the idea of holding something without hands. The model can occasionally draw things that way, anyway, with or without it.floating cape- Because tags like arms or wings outstretched can give it the wrong idea about their anatomy. Although they often don't.
Quality Indicators
Obviously, these are just my subjective ratings. But this is my experience so far.
★★★★★ Character Likeness
Good at what it does, at least at the angles it can handle. You ask for a felbat, you get a felbat. If you don't get a felbat because of other tag biases, a higher weight on felbat and/or few reminder tags will bring it back into line. Sometimes shows them with two eyes, which they actually do have, the one is just covered by the hair from most angles.
★★★☆☆ Flexibility
Somewhat stiff at full strength. Can definitely do clothing swaps, although it tends to need more reminders when it does. Images seen from the side or behind tend to get mixed up. It can only sort of do upside-down, which is unfortunately important for bat-like critters. The face can be fairly expressive with the right tags.
★★★★★ Style Compatibility
Extremely well-behaved in my experience. I haven't seen any major effect on styles.
★★★★☆ Other Checkpoints
Moving it to something as far away as a realism model takes a lot of reminders, but it can do it. The realism model I tried tended to think it's a costume, not a creature, which is certainly an interesting result.
★★★★☆ Other Characters
Sadly, it pretty much took over the word "bat". You can get other bats, but they tend to have felbat hair, or outfit, or horn, or elsewise. Maybe I should've created regularization images for bats. Other species are mostly unaffected, though, so I'm giving it 4 stars.
★★★★☆ Sanity
One star docked because, while it's technically a base model issue, the "bat" part of the name fills a lot of pictures with little friends. A few style-related tags I didn't create regularization images for might have gotten messed with. Lesson learned there. It seems to forget the second foot when the legs are crossed, perhaps more than the base model, and I think I know which picture did that. Also, maybe unsurprisingly, it seems to favor the character being seen from slightly below.
Training Info
Base model: NoobAI-XL EPS 1.0
Concept data: 12 human-created images with various styles and aesthetics (1/12 porn, 5/12 with minor anthro traits, 0/12 with hands)
Regularization data: 120 AI-generated images from NoobAI-XL V-Pred 1.0 with the same captions as the concept data, without "felbat" or certain style-related tags, with randomized aspect ratios and quality tags
Balance: 1 regularization image per 1 concept image, same loss weight
Total Steps: 3750
Notes: The "bat" token thing is annoying. I tried training with its Japanese name, yamitobari, but some part of that word was attached to an artist style or two so I gave that up early. Training a nonsense tag (f3lb4t) took forever to even pick up the likeness. So, felbat it is.
Things Nobody Asked (Yet)
Why are you training everything on Epsilon 1.0? There's a newer version!
Compatibility with the models I use. It's a common ancestor between Epsilon 1.1 and Vpred 1.0, and also 75% of an ugly semi-realism merge I made a while back and still kind of like. Not sure how important that is, but it just seemed like the best place to start, and I haven't had any problems using them on other models yet.
Can I lewd the felbat?
You can lewd the felbat. Just don't let it hug you.
It has no mouth and no claws. What the heck goes on inside that cloak when they feed?
The pointy gold things at the waist are its fangs. That's my theory and I'm sticking to it.