Please remember to reply and/or post your gens if you like the model. Click "Expand" to continue reading for best results using this LoRA. Do not complain about horrible results if you are unwilling to take the time to read this through!
DO NOT USE FACE RESTORATION WITH THIS MODEL! USE ADETAILER INSTEAD (SEE BELOW)
ALL MY MODELS ARE TRAINED SPECIFICALLY FOR AUTOMATIC1111/FORGE/REFORGE USING ADETAILER TO INCREASE REALISM & RESEMBLANCE. THEREFORE, TENSOR.ART ONLINE GENERATIONS & COMFYUI WILL PROBABLY LOOK LIKE GARBAGE.
This is my first foray into character LoRAs. It was surprisingly simple compared to the photoreal models I typically train. My training method was simple. I extracted Alexstrasza's character model from the game, rendered it in 3D, and exported 814 images using various poses, angles, lighting, etc. I then grabbed 20 2D & 3D images from across the web, cleaned them up in Photoshop, and...viola! This came out much better than I expected considering that no cosplay or other photos were used in the character training data.
Checkpoint Note:
This LoRA works incredibly well with WaffleMix_v7 for realistic renders, while AniMerge v2.4 works well for 2D and some 3D.
MANDATORY Triggers
Alexstrasza, red hair, golden eyes, horns, long eyebrows, headdress, slim body
OPTIONAL Triggers
upper body, purple jewel, cleavage, fantasy armor, tabard skirt, pauldrons, gauntlets, glowing eyes
Prompt Format
Keep prompts and negatives as short and simple as possible. The format should look like:
a photo of ##Mandatory Triggers##, ##Optional Triggers##, ##the rest of your prompt##
NOTE #1: Disable face restoration if you wish to prompt "glowing eyes". It will negate the prompt. NOTE #2: The following prompts are unreliable and may be fixed in a v2: gauntlets (sometimes show fingers), pauldrons (don't always get rendered). The following prompts are unreliable but are still required triggers: long eyebrows (often brown instead of red, sometimes way too short for canon), headdress (only partially renders sometimes)
Recommended Settings
Checkpoints: WaffleMix v7, epiCPhotoGasm version X, AniMerge v2.4. Other photoreal, realistic, & 2D models may work
Clip Skip: 1 (realism) or 2 (flexibility)
CFG: 3-8 (lower=more realism, higher=more flexible)
Steps: 30-60 (less than 30 isn't recommended, except for 2D)
Sampler & Scheduler: DPM++ SDE Karras
Hires Fix scale: 2-4x
Hires Fix Steps: 12-20
Denoising Strength: 0.15-0.30
Upscaler: 4x_foolhardy_remacri
NSFW Support
I trained NSFW body parts seperately using real photos to increase detail and realism. These can be prompted but work much better when used in aDetailer.
aDetailer Female Breasts model v4.7 - Denoise Strength: 0.25-0.30
best quality, <lora:Alexstrasza_v1:1>breasts, nipples, 8K, high quality photography, Canon EOS R3
aDetailer Vagina Model v4.1 - Denoise Strength: 0.40-0.55
best quality, <lora:Alexstrasza_v1:1>p ussy (DELETE SPACE), high quality photography, Canon EOS R3