This was a labor of love. This is my first LoRA and was a great learning experience.
Trained on 133 manually captioned images with an average resolution of 3500x3500.
The majority of this dataset is in a portrait orientation; over half of the images have an aspect ratio between 1.5 and 1.75, so you will get the best results by adhering relatively closely to this ratio. Roughly 1/4 of the set is between 0.67 - 0.8, so it can handle wider ratios, though they may be somewhat unstable. Most example images use 832x1216, though it can easily generate at 1536x1536 thanks to the high resolution of the training set.
Changelog:
V1.6 NoobAI: Once again, better in every way compared to 1.5. Add "realistic" ("intricate detail" optional) for a more painterly effect. Euler for rougher/painterly effect, Euler A for a more typical anime effect, DPM++ 2m for a middle ground. Weight 1 works fine, may drop to 0.8 for more detailed/complex prompts if anatomy gets weird.
V1.6 Illustrious: Significantly improved colors, style and prompt adherence. Generic, open-ended prompts will show the base style more heavily - prompt "1girl, solo" and you'll see what I mean. Guiding the prompt will lessen this. I recommend using "Realistic" and possibly "intricately detailed" for best results. V1.5 Noob: You asked, I listened! See the model changelog for details. Please leave feedback - I'm curious to see how it works with merges and finetunes. V1.5: Better than V1 in pretty much every way. I tested primarily with PrefectIllustriousXL V1.0 and WaiNSFWBranchRouWei_epred1 and it works great on both. You may benefit from adding "realistic" to your prompts with either model, particularly WAI as sometimes the base anime theme can overpower the lora. This version may be slightly overfitted, so dropping down to 0.8 can help in some cases.
Does well:
General character features and poses
Outfits, accessories and materials (see below for recommended keywords)
Facial expressions
Fluids
General piercings
Is capable of / generally does well:
Fetish gear
BDSM-specific poses and equipment
Yuri
Penetration (use femdom for more consistency)
Hands (using hiresfix solves this for me in most cases - not tested with adetailer)
Struggles with / not guaranteed:
Males - they simply don't exist in the dataset, as the artist does not draw them. I have no interest in adding them.
Significantly complex positions - I'm working on improving this with better captioning
Needle play - Despite featuring prominently in the dataset and careful tagging, this has proven to be very inconsistent due to how niche this fetish is. The model tends to conflate "piercings" with "breast piercing", so you'll usually get nipple rings etc instead of the needle actually penetrating through the breast.
Catheters - see above, difficult to prompt despite appearing a lot in the training set, somewhat more stable with "urethral insertion"
Corset piercing - can SOMETIMES work when paired with "body piercing" but very inconsistent
RECOMMENDED TAGS (pick and choose, don't use them all):
Concepts: bdsm, bound, bondage, yuri, femdom, saliva, drooling, gagged, frogtie, hogtie
Clothing and Accessories: latex, ballerina, nurse, maid, nun, fake animal ears, oxygen mask (use with "tube"), respirator, toeless legwear, stirrup legwear, frills, hair ornament, hair flower, detached collar, bobby socks
BDSM Gear: ball gag, bit gag, spider gag, harness, stationary restraints, **** tail, speculum, ballet boots, spreader bar, nose hook, linked piercing, blindfold, shibari
Special (proceed with caution): urethral insertion, catheter, asphyxiation, needle play, speculum, gaping, whip marks
NEGATIVES:
All the normal no-nos from Civitai generation. Use common sense and be responsible.
RECOMMENDED PARAMETERS
Lora Strength: Works great at 1, but dropping to 0.8 can help if it's struggling with anatomy in certain poses
Dimensions: 832x1216, 768x1344, 800x1296, 1024x1536, etc.
For landscapes, just reverse the resolutions.
Samplers: Euler, DPM++ 2M, DPM++ 2M SDE
UPSCALING
I recommend VIPSMagicKernelSharp2021 if you have access to it. If not, ESRGAN 4x YandereNeoXL also does well. Other popular upscalers (RealESRGAN 4x+ Anime6B, ESRGAN 4x Ultrasharp, etc.) did not work well, smoothing detail and corrupting hands. VIPS and Yandere consistently improved finger structure in my testing.
Resize Scale: 1.5
Refine Sampler: Generating with Euler and then selecting DPM++ 2M SDE here allows for a nice blend of creativity and added detail. Otherwise, same as primary.
Strength: 0.3 if you're happy with the current level of detail, otherwise 0.4.
Steps: Roughly half of your primary steps.
WHAT'S NEXT?
V2 will focus more on concepts. I'll try to improve the nuance of the art style later, but for now I'd like more consistent prompting of things like needle play, body piercing, asphyxiation and breath play, etc. I have a few ideas for how to accomplish this, so stay tuned.