Having now trained three LoRAs which I've had to handtag (datasets collected by me from the original work), I've noticed many issues in the trainer UI. I've submitted improvement to these shortcomings as feature requests, but since there hasn't been a feature "approved" for public viewing in months, I'm assuming that whole section of the site is abandoned, so I'm going to post a public article detailing my concerns. I am a desktop user with a single 1440p monitor.
You can't see a full image as you're tagging it
This one is a critical problem: You can only see a small, cropped, version of the image you're tagging. This is annoying because you can't tell if a character's bottom garment is visible (and thus things like "pants" should be tagged) and more. A workaround is to copy and past each image into an editing program, but why should I have to?
The boxes for tags are tiny
Tied in with the above, Both the box for inputting tags, and the box holding existing tags are tiny.
The solution to this and the above is let a user focus on a single image at a time and have an appropriately sized box to input in while listing all current tags so they're visible at once.
Auto-tagging is all or nothing
Pretty straightforward. You can only use auto-tagger to tag everything. Can't ask to tag one image.
Auto-tagger does not check for redundant existing tags
Straightforward again. Auto-tagger will freely apply a redundant tag if you decide to tag everything with a particular tag (such as monochrome because you know your entire dataset was from a black and white manga) or append.
Auto-tagger should offer an approve mode.
Right now the auto-tagger just adds tags to an image. I'd like to see a mode that adds them to a "suggestion bucket" instead, allowing the users to easily add any tags from it they like and drop those they don't.
No easy way to purge duplicate tags
Again, a straightforward issue. The small size of the tag box means you can't easily see if a tag is duplicated. The count of tags counts instances of a tag instead of images tagged with it, which means you will see a tag listed as having multiple instances when there's actually only one image with it. You may want duplicate tags, but most of the time you don't and there's no good way to get rid of them.
No option to copy tags from an already tagged image
Again, straightforward. I had to resort to keeping a text document of common tags and copy pasting in as needed.
Image order is random
This issue also seems to apply to images posted to the bounty system. I literally DLed my dataset to take a break from tagging, then reuploaded it and got a totally different order of images!
Option for selecting Pony as a base model has terrible description
Pony is labeled as being XL but specialized in furry and other such stuff. This is completely misleading. Pony should be on its own row and have a description reflecting its nature as a finetuned XL for non-realistic use.
No suggestions for autocomplete
Again, straightforward. The autocomplete currently used for post tags is pretty bad and should not be ported: Poor at finding common tags, and really not friendly to anyone who keeps typing after it tries to suggest something then submit it with enter (it instead eats the thing you just typed out and tags only the part it saw when it first tried finding tags, so you get partial words).
No option to post newly trained LoRA as a version 2 (must be new post).
Straightforward, and rather annoying. Workaround is possible by DLing from training then uploading that, but that seems an excessive step given that most LoRA are 200+ MB.
Option to release dataset isn't written to accommodate public domain or CC images
A minor one, but public domain images aren't "owned" by the uploader (and there's a lot of stuff that's relatively new due to government works being PD by default in various countries), nor are Creative Commons images (for those who wish to make a LoRA and release under CC)
No clean way to attach credits for CC-By images used in training data.
Uploaded files only allow training data or the LoRA itself. Had to include in description.
Update (October 10th 2024): Fourth of a year and 25 LoRAs latter and I think every single thing here is still an issue. The image order thing might be changed but I haven't experimented with it enough.