Ablyssal

The settings usedFairy: translucent wings, pointed ears, fairy wings, glowing particles, butterfly wings, flower accessoriesSpirit: translucent body, spectral glow, floating, wispy edgesSeraphim: six wings, multiple feathered wings, halo, golden light, floating eyes, divine auraMermaid: mermaid, fish tail, scales, fins, seashell accessories, webbed fingersSiren: mermaid, fish tail, sharp teeth, glowing eyes, dark scales, fin earsSea Elf: pointy ears, webbed fingers, gills, blue skin, fin ears, aquatic markingsEnt: humanoid tree, bark body, branch arms, moss, leaves, glowing eyesDryad: tree bark skin, leaf hair, branch antlers, vines, flowers, glowing markingsElf: pointy ears, long ears, slender build, ethereal, elegant, leaf accessoriesOrc: green skin, tusks, muscular build, pointy ears, tribal armor, scarsMushroom folk: mushroom cap, fungal skin, spores, glowing skin spots, mycelium patternsKitsune: fox ears, multiple fox tails, nine tails, fox mask, spirit flamesCentaur: Centaur, human upper body, horse lower body, four legs, horse tail, hooves, armorOgre: large muscular build, tusks, broad shoulders, protruding lower fangs, thick skin, hornsDwarf: short stature, stocky build, braided beard/Braids, thick eyebrows, heavy armorLamia: lamia, snake tail, human upper body, scales, slit pupils, forked tongueNaga: serpentine lower body, snake tail, scales, cobra hood, fangs, golden eyesDragon: dragon, scales, horns, large wings, spiked tail, fire breath, glowing eyesDragon kin: dragon horns, dragon tail, scales, slit pupils, claws, dragon wingsDemon: demon horns, demon tail, bat wings, red skin, glowing eyes, dark auraLizard folk: reptilian humanoid, scales, lizard tail, claws, slit pupils, spinesDjinn: floating humanoid, smoky lower body, glowing markings, magical smoke, ornate jewelryWraith: shadowy figure, glowing eyes, spectral body, tattered cloak, dark mistFallen Angel: black feathered wings, broken halo, dark aura, glowing eyes, black feathersNephilim: angel wings, demon horns, mismatched wings, cracked haloVampire: fangs, red eyes, pointy ears, gothic clothing, sharp fingernailsDark Elf: dark skin, pointy ears, long ears, white hair, red eyes, silver jewelry

IntroductionThis guide will walk you through the process of creating Krea 2 LoRAs on Tensor.art, from preparing your training images to configuring your dataset and training settings. The goal is to give you a clear, practical workflow you can follow even if you’re new to LoRA training. By the end, you’ll have a better understanding of how to prepare your data, choose the right settings, and train a LoRA that works reliably with Krea 2. If you have any questions, feel free to leave a comment and ill do my best to helpStep 1: Preparing Your ImagesThe first step in creating a good LoRA is deciding what kind of LoRA you want to make and preparing your training images accordingly. This could be a style LoRA, artist LoRA, character LoRA, or another type of concept LoRA. Once you know what you want your LoRA to learn, there are many different ways to gather and organize your training images, but my preferred method is to create a Pinterest board specifically for the LoRA and collect your training images there.Once you have your board ready, you can use Pinterest-DL Pinterest-DL to download the images in bulk (I used ChatGPT to help me set this up).For the best results, I recommend gathering at least 100 images that are reasonably consistent with the subject or style you want your LoRA to learn. The quality and consistency of your dataset will have a major impact on the final LoRA, so take some time to remove images that don't fit the dataset before moving on to training.Step 2: Tagging Your ImagesOnce you have your images, the next step is to tag them. Tags help describe the contents of each image and give the LoRA information about what it should learn during training.My preferred method is to use an automatic tagger. I personally use SmilingWolf/wd-eva02-large-tagger-v3 (I used ChatGPT to help me set this up), but you can use any image tagger that works well for your workflow.There are several ways to run an auto-tagger. You can set one up directly on your PC, even if your computer isn't particularly fast. You can also find an online auto-tagger that supports batch uploads, which can make processing a large dataset much easier.Tensor also has its own auto-tagging tools, but I don't recommend using them for this workflow because editing and reviewing the generated tags can be more difficult. Being able to easily go through your tags and clean them up is important before moving on to training.Step 3: Cleaning Up Your TagsIf you use an auto-tagger, your dataset will usually contain a lot of unnecessary or inaccurate tags. Before training, it's important to go through your tags and clean them up. I highly recommend using BooruDatasetTagManager for this, as it makes it much easier to review, edit, add, and remove tags across your dataset.The tags you keep or remove will depend on what type of LoRA you're making. For a style or artist LoRA, you generally want to keep tags that describe the actual contents of the image while removing tags that describe the style you are trying to teach the LoRA.For example, you would want to keep tags such as 1boy, 1girl, dress, trees, and other tags describing the subject or scene. You would generally remove tags such as realism, painting, or fantasy if those characteristics are part of the style you want the LoRA to learn.After cleaning up your tags, you'll want to add your trigger word. A trigger word is a unique word or phrase that you include at the beginning of ALL images in your training captions to identify the concept your LoRA is learning. During generation, using that same trigger word tells the model to activate the learned concept. Choose something distinctive that is unlikely to already have a strong meaning in the base model. For example, instead of using a common word like realism, you use a unique token such as myartiststyle for an artist or style LoRA.Character LoRA TaggingCharacter LoRAs require a different approach to tagging than style or artist LoRAs. I don't have enough experience with character-specific tagging to give reliable recommendations here, so I won't pretend otherwise. If you're making a character LoRA, I recommend looking for a dedicated character LoRA tagging guide to learn which features you should tag and which ones you should leave for the LoRA to learn.Dataset StructureOnce you're finished cleaning your tags, your dataset should consist of an image file and a corresponding text file for each image. The image and text file must have the exact same filename, with only the file extension being different.For example:image001.jpg image001.txt image002.jpg image002.txt image003.jpg image003.txt Most automatic taggers will create these text files for you, but always double-check your dataset before training. Every image should have a corresponding .txt file with the same name. These text files contain the tags and captions that the LoRA will use during training.Final Dataset ChecklistBefore you start training, it's worth doing one final check of your dataset. Training can use a significant amount of credits and time, so catching problems beforehand can save you from having to retrain.Every image has a corresponding .txt file.The image and .txt file have the exact same filename.Your images are relevant to the type of LoRA you're making.You've removed unnecessary, incorrect, or unwanted tags.Your trigger word has been added to all images.There aren't any obvious duplicate or broken images.Your dataset is reasonably consistent with what you want the LoRA to learn.You've checked your captions one final time before training.Once everything looks good, your dataset is ready to upload and train. make sure everything is in one zip fileStep 4: Training Your LoRAOnce your dataset is prepared, you're ready to start training. Click on upload dataset then click on your zip file. For Krea 2, I've found that around 10,000 training steps is typically a sweet spot for a dataset of roughly 100 images. However, the ideal number of steps can vary depending on the images you're training on, so treat 10,000 as a suggestion rather than a strict rule.My usual starting settings are:Repeats: 15Epochs: 7Save Every N Epochs: 1Training Steps: ~10,000Base Model: Krea 2 Raw, not TurboFor the best results, I recommend training on Krea 2 Raw rather than Krea 2 Turbo.If you don't have much experience with LoRA training, don't worry about changing all of the other training parameters. Base settings work well, so I recommend leaving the advanced settings alone until you understand what they do.Saving Credits with Epoch ContinuationKrea 2 training can be expensive, but there is a way to spread the cost out if you have a Pro subscription and are short on permanent credits.If you have a daily credit allowance, set your training to 1 epoch and adjust the repeats so that the total training cost is around 300 credits. Once your daily credits refresh, use epoch continuation to continue training from where you left off. You can repeat this process each day.This approach will take several days to reach your target training time, but it allows you to train using your daily credits instead of spending your permanent credit balance. It's a slower method, but useful if you want to reach around 10,000 steps without paying the full cost upfront.Step 5: Testing Your LoRAOnce your training is finished, it's time to test your LoRA. Don't assume the final epoch will automatically be the best one. Depending on your dataset and training settings, an earlier epoch may produce better results than the final one.After an epoch finishes, click Publish, then choose Create or Add to a New Project. Give the project a name and use it to test your LoRA.I personally recommend testing every epoch rather than only testing the final one. For each epoch, use the same prompt or a small set of consistent prompts so you can compare the results fairly. I like to run multiple generations using the same prompt for each epoch and compare the results as a group.Testing multiple generations is important because a single generation doesn't always represent how well a LoRA performs. By testing each epoch with the same prompts, you can see which checkpoint consistently gives you the results you're looking for.The last epoch isn't always the best epoch. An earlier checkpoint may have learned the concept better, while later epochs may begin to overtrain or produce less desirable results. Test them all and choose the checkpoint that performs best for your specific dataset.Keep Your Testing ConsistentWhen comparing different epochs, try to keep your prompts and generation settings the same. For example, if you're testing Epoch 1, use the same prompt and settings when testing Epoch 2, Epoch 3, and so on. I also recommend generating multiple images with each prompt rather than judging an epoch from a single generation.This makes it much easier to see which epoch is actually performing better instead of being influenced by random differences between generations. Once you've tested all of the epochs, choose the checkpoint(s) that consistently produces the best results for your intended use.