❌💲 Commercial use of this model is not allowed
I like this bird species, but AI simply can't generate it, the closest you can get is an ambiguous corvid. So I made this LoRA to see if I can make it work.
🏷 DATASET TAGS (doesn't apply to Flux and ZIT, those use natural captions)
Tagging is a mix of autotagging and manual.
The whole dataset is tagged with "jackdaw, western jackdaw, eurasian jackdaw, coloeus monedula, corvus monedula, corvid, bird, feral, animal, animal focus, realistic, no humans".
Some pictures focus on the bird's head, most of the others use "fullbody".
Most images in the dataset show solo birds (tagged "solo, 1 jackdaw, 1 bird"), a few images show two ("duo, 2 jackdaws, 2 birds").
🖼 DATASET
122 images which are photos only (doesn't apply to the Flux version). I don't own anything.
Images don't contain signatures/watermarks.
Resolution is all over the place, flying pictures are lower quality overall, though some images were large enough to be automatically sized down.
Images don't show any humans or other distinct creatures.
Only adult birds with normal colouration. No eggs, hatchlings or fledglings, no colouration anomalies.
⚠ ISSUES
The bird's colouration most likely will be lost in cartoony results.
BONUS: I tried SD3.5L, but it doesn't work and I don't know why.
I deleted the LoRA for SD3.5L since it couldn't produce the jackdaw. You don't need a LoRA for generic corvids. I thought I could make a proper LoRA if I used all the 122 images and maybe increased the number of Epochs. But seeing the sample images, I knew that nothing helped... I'm not putting these useless LoRAs on here, but if you can use SD3.5L locally, you can see for yourself. I took time to correct the captions on 14 pages of images so there wouldn't be any incorrect bird names, so it wouldn't be called "large" or its eyes "yellow", and it was all pointless... How come Flux works with about 30 images, but SD3.5 doesn't with 100+? And I wondered why SD3.5 LoRAs were so scarce. Ain't gonna bother with it again anytime soon.
Description
TL;DR Dataset is 122 images with natural language captions.
I wanted to try something relatively safe for my first ZIT LoRA. Just to be sure, I did a couple of test generations with the base model alone: I only got what I presume to be hooded crows so I knew it couldn't do jackdaws by itself. I grabbed the full dataset with natural language captions that I previously made for a SD3.5L LoRA (which was a complete failure), the captions were generated on-site and I corrected them, and I just tossed it into the generator.
Strange things happened when I trained this. When I downloaded the LoRA, I found its name was gibberish. I renamed it for clarity, hopefully that doesn't break it (I still have the file with the original name). The test outputs also had gibberish names, I changed them too. Also, some results with the 3rd prompt (including for the last epoch) refused to load, I checked in two different browsers, so I grabbed the 3rd image from a previous epoch.
Trained on-site on January 5 to 6.