This is Part 4, and yes the last part, of this introductory article guide to LORA creation. Here I will present links, many to LORAs and images I created while writing this, to provide a comparative sampling of what can be made without needing to know a lot of technical details. Seriously, I basically just uploaded the images and let the tensor.art training platform do its thing. So now you need to decide whether to...
Step 3: Dive in head first or check out one of the tensor.art specific step-by-step guides for LORA creation.
If you have a dataset ready and feel like you're good to go then get started, you can peruse this article at your leisure while waiting for your LORA to train. Or, if you're still uncertain, there are many great informational articles published here on Tensor.art. To begin I’d suggest skimming at least one of the step-by-step guides before diving in blindly. Here are links to a few:
https://tensor.art/articles/875734357225025459
https://tensor.art/articles/806929088523547379
https://tensor.art/articles/849095063096129962
https://tensor.art/articles/732090335657730218
All steps in the process ultimately lead to...
Choosing your LORA
Bottom line the best LORAs cost the most to train. There are a few that you can bank credits for in a short period of time, depending on how many images you have in your dataset. The two best for this are probably the Lightning FLUX and SD 3.5L. But they are very different.
How different?
As a test I created a dataset comprised of 6 Dystopian, 5 Fantasy, and 9 Wasteland warrior AI generated images using the comedian Katherine Timpf as my character template. All that means is, instead of describing a character in detail, my prompt ask was “Katherine Timpf as a” or “depicting Katherine Timpf as” kind of syntax. To round the dataset out I included 6 images of a character holding a sign reading “Kat”, 4 ‘photoshop’ style AI retooled (background cutout/replaced, indrawn touch ups, etc) genre themes images, 2 upscaled headshots, and 10 random other images. That should make for a solid, if not better, dataset than using just 10-12 images, right?
Let's find out. I have generated comparison samples for both versions of the WASTELAND WARRIOR KAT LoRAs: https://tensor.art/images/876085333362493423?post_id=876085333362493424&source_id=njeyo1nrnEW3oPYsaX309xkk
The first image sticks close to what the image set was intended for. The Witch art, which actually looks pretty amazing, was the best test I could think up that had completely no aesthetic similarities with the source images. I think they prove the Flux LORA is fairly versatile with prompt interpretation on its own whereas the SD 3.5L version seems less so. Flux costs more but the images should speak for themselves.
For example this is what we get when combining WASTELAND WARRIOR KAT with other LORAs:
Wasteland Warrior Kat FLUX
Not terrible. Now let’s try…
Wasteland Warrior Kat 3.5L
Obviously the first couple with the polar bear were a quick test. Simple prompt. No reference to the character or character description. (She even disappeared entirely from a couple image generations, like the 3rd image above, for some odd reason.) Now here’s four with more or less the same prompt, but generating very different artistic styles:
With SD 3.5L the prompt has to be worded very specifically whereas with FLUX you can write a few words and probably get a decent looking pic. Even the polar bear prompt generates a slightly different looking image without adding extra descriptors when using the FLUX version: https://tensor.art/images/877593127802635927?post_id=877593127802635928&source_id=njeyo1nrnEW3oPYsaX309xkk
Of course you don’t have to use either. But, by now, you should have a dataset ready (if not already uploaded) and all that’s left to do is train your chosen LORA then…
Step 4: Name and Publish your LORA
I know what you’re thinking... What did Mudge do that necessitated spelling out the obvious final step as if it’s difficult?
I chose a LORA name “Bare Ten” (because it only used 10 images), which sounded better than “Just Ten” in my head. Also I thought it fit well with my low-key “Barefoot” naming meme. The problem? I forgot “bare” is one of those words that is a red flag for auto-mod filters everywhere. It’s okay to laugh because that wouldn’t be a problem, except for the fact I often use my LORA names as trigger tags.
You see where this is going. Too bad I didn’t. But I quickly discovered the problem as I had to log in to see images created using those tags. So now I feel kind of silly because it just never occurred to me. The fix was simple, rename the tags. Which I did. I even came up with a more descriptive name for the LORAs. The links to them (and a few more I created) are below.
Ten Scenes Emilia: https://tensor.art/models/871749988005217621?source_id=njeyo1nrnEW3oPYsaX309xkk
https://tensor.art/models/872118629343169865?source_id=njeyo1nrnEW3oPYsaX309xkk
Ten Scenes Pirate Girl:
https://tensor.art/models/871704761999581375?source_id=njeyo1nrnEW3oPYsaX309xkk
Vintage Photo Nichelle Nichols:
https://tensor.art/models/872788354331067482?source_id=njeyo1nrnEW3oPYsaX309xkk
https://tensor.art/models/872497183465720805?source_id=njeyo1nrnEW3oPYsaX309xkk
What the above all have in common is they’re either trained on 12 images: 10 main scenes and 2 poster art mockups with title text or fewer than 20. I don’t know if it helps to train a LORA including images with text or not but I figure it can’t hurt. Besides I am testing my work prompt anyway to be sure it creates a decent looking title so might as well use the sample images, right?
Well, that’s it. Thanks for reading. I hope this helped. All that’s left is a refresher on the…
Simple Annoying Stuff:
IMAGES: Whether curated from an public domain image archive or sourced from free range pictures you corral and lasso in an image search, there’s always two things of primary importance. First, clarity. Second, quality. The better your source image the better your training will be, which means a better model result.
UPSCALING: Not necessary for images of decent quality. However, for the best results, use a tool that outputs images of high-quality but not necessarily large file sizes.
IMAGE OPTIMIZATION: This is sort of related to the above. If you have a tool that outputs too large of a file size this may become necessary. However, and here’s a trick I discovered by accident. Some tools will only ever produce images in the 1+ MB size range. Problem is optimization often loses some of the quality. What to do? Find a tool that outputs smaller already optimized files. Often you can upscale images through another tool and not only get an optimized file but a fairly decent image with more pixel resolution. Doesn’t always work but when it does, golden.
RETRAIN: Before your LORA expires you can RETRAIN it, meaning use all your already uploaded pics and generated tags, to create a NEW LORA. And, no, it doesn’t have to be the same model you already trained on.
Now Let's Go Make a LORA!
Tap the menu tab at upper right of screen. On the drop down menu you should see "Training" (just above "Creator Dashboard"). Click on "Training". This will take you into the first of the LORA training pages. From here you begin the process of uploading your images, or a prepared dataset in a ZIP archive. For noobs like us we will be batch uploading images and let Tensor auto create tags. Now you can add specialized tags and prune the auto-generated tags, but that’s a whole different article. Once the images are uploaded and the tags generated pick a LORA. Do you have the credits? If so, great! If not you have a couple choices: 1) Bank more credits. 2) Pick another LORA. 3) Use fewer EPOCHs. Once your settings have been adjusted to your liking and you've filled in all necessary information begin a training session. If it fails don't panic. Try again in a few hours when there's less traffic on the site.
Congratulations! You've made a LORA!
