Loading
Have a ComfyUI workflow? Get a readable Workflow Doc in your browser.Make a workflow readable.Try it now ↗
Because sometimes, what you really need for your mermaid OC is that extra bit of uniqueness to it. I tried to aim for as much diversity with this dataset as I could before getting to work, with as many types of...
Model source
Source description
Second Anima release, with a few new additions and musings to note.
Let's start of with the new types included before moving onto the v1 types, from the types that work the best to the ones that don't:
Seahorses! It took quite a bit of training to get this one to work properly, a lot of training steps before it finally played ball, but it finally did. Some extra data for sea dragons were included as well, but they didn't exactly stick much in training.
Cephalopods/scyllas are also included. They were probably the type that got learned the quickest, to be honest. Dunno how much control you'll have over the exact amount of tentacles, but that's an experiment in itself.
Sea snakes: Were pretty easily learned, and surprisingly didn't interfere too much with the eel data.
Sources
1 sourceModel artifacts
1 artifactExotic_Mermaid_Types_epoch_18.safetensors
safetensors · BF16 · 132 MB · SHA-256 b46552a5baf1…55aa · Civitai
DownloadSource excerpts
1 excerptSource context: 278 downloads · Type LORA · Base model Anima
Sirens/harpy hybrids were an easy train. Guess that comes with harpies and mermaids having anatomies affecting vastly different body parts.
Bettas are included, but a bit wonky and inconsistent in body proportions. They work, but not consistently so.
Rays and mantas are where the training consistently began struggling. I blame lack of proper standardization of style in dataset. Case in point, I wanted both hip fins and arm fins included as options, but the only one that ended up even remotely giving good results was the hip fin variant. Though I guess in its current state it does have a tendency to create very alien and uncanny results, so take that for what you want.
Sea angels/sea butterflies barely produced any results beyond the bioluminecence. I blame a lack of dataset diversity for this type.
Melusine/two-tailed mermaids... Yeah, erm, this one didn't work. At all. At least, the tests didn't yield anything noteworthy. Maybe once in a blue moon this might work, but don't count on it.
As for what may be different with the ones included in v1:
Sharks and lionfish seem to be rather improved in quality
Eels seem a bit more stable, but the lack of distinct head fins from v1 is a sad loss
Random pattern mermaids seem more keen on giving distinct patterns rather than doing a clown blanket like in v1
Clownfish and koi don't really have much change to them
Anglerfish leans a lot more into the bioluminescence part of things. The lure/lantern can be a bit inconsistent, but not by much.
Jellyfish seems to lean more into the inhuman side than v1, can just produce a regular old jellyfish if not careful lol. Other than that, no issues to report.
From my tests, 0.7 should be the sweet spot weight to use this at. I'll update this with the tags I use as I go, but you'll have to wait and see.
