Model details
Merge of the following models, credit goes to their authors (permissions & commercial use inherited from the strictest source model):
For exact recipe of each version, check the version descriptions.
Usage
It is highly recommended to download and use the following extensions:
https://github.com/Bing-su/adetailer - improves face/hand details when generating character images
https://github.com/mcmonkeyprojects/sd-dynamic-thresholding - improves behavior of the CFG scale, allows for more detail without loss of quality
Recommended LoRa for further details in images: https://civarchive.com/models/153562?modelVersionId=171989
Recommended settings
Clip skip: 1 or 2
Steps: 25
CFG Scale: 7 (without dynamic thresholding), 20 (with dynamic thresholding)
Sampler: DPM++ 2M Karras or DPM++ 2M SDE Karras
Upscaler: 4x-UltraSharp or R-ESRGAN 4x+ Anime6B
1st ADetailer: face_yolov8n
2nd ADetailer: hand_yolov8n
Mimic CFG Scale: 7
Top percentiles of latents to clamp: 100
Mimic Scale Scheduler: Half Cosine Up, min value 5.0
CFG Scale Scheduler: Half Cosine Up, min value 5.0
Hires. fix (upscaling) helps with enhancing the images, but results are fine even without upscale, using the above settings (so you only need to use Hires. fix, if you need the higher resolution). Example (made with v1.0):
Description
Further checkpoints & LoRas merged, check description for details. Pruned, f16 safetensors format.
v3.0 Recipe:
BMRA = brickMortar_v10 + restlessanimance_v10 (weighted block merge)
weights: 1,0.800968123070988,0.674238040123457,0.575358072916667,0.501543209876543,0.450008439429012,0.41796875,0.402639130015432,0.401234567901235,0.410970052083333,0.429060570987654,0.452721113040124,0.479166666666667,0.50561222029321,0.529272762345679,0.54736328125,0.557098765432099,0.555694203317901,0.540364583333333,0.508324893904322,0.456790123456789,0.382975260416667,0.284095293209876,0.157365210262345,0
brickMortar_v30 = BMRA + seizamix_v2.safetensors (weighted block merge)
weights: 1,0.800968123070988,0.674238040123457,0.575358072916667,0.501543209876543,0.450008439429012,0.41796875,0.402639130015432,0.401234567901235,0.410970052083333,0.429060570987654,0.452721113040124,0.479166666666667,0.50561222029321,0.529272762345679,0.54736328125,0.557098765432099,0.555694203317901,0.540364583333333,0.508324893904322,0.456790123456789,0.382975260416667,0.284095293209876,0.157365210262345,0
Additional modifications to the final merge via LoRa merging:
FZB_line_weights:0.35 (https://civitai.com/models/81238?modelVersionId=86180 )
outline:0.2 (https://civitai.com/models/81340/outline )
FAQ
Comments (3)
I love this model. It has a very distinct style and always does well. Great job!
Just did a few with v3 and I gotta say: this feels more like a step in a whole different direction, than a v3 (2d/anime as opposed to just a plain upgrade over v2) and in all honesty I liked the more 2.5d/painterly v2 better (just my personal taste, ofc). What I want to say with all this: I'd name the versions accordingly. Most people will assume a v3 is just a better version than a v2 and not even look at the v2 anymore. Imo in this case that would be a shame. ^^
This is a very good solid model. A good selection of models for merging gave excellent flexibility, expressive 2d stylistics and good stability.
Like.