AkashicPulse EQVAE is a model based on AkashicPulse series, but fine-tuned to utilize Anzhc's EQVAE as its VAE. (In theory,) This improve the model generation color and accuracy.
Recommended settings [UPDATED FOR V1.0]:
Same like v0.2, the model is in between needing RescaleCFG and not needing it. Expect to turn on and off this a lot.
The usage of artist tags IS HIGHLY RECOMMENDED. Put it at the FRONT of your prompt (and use BREAK if your generation platform support it).
Steps: 32-50
CFG: 7-10 with RescaleCFG on or off, or 6 if you don't wanna touch RescaleCFG
RescaleCFG: 0.7 OR OFF, inspect generated image quality if something is off.
Sampler: Euler a (or try AkashicSolver available on Adept Sampler with AkashicAOS).
Noise Scale: 0.9, using 1.0 will cause the model to over-generate noise.
Recommended negative prompt:
worst quality, low quality, bad anatomy, bad hands, text, signature, watermark, username, artist name, [things you want to remove]WARNING: Using excessive negative prompt might resulting the model to lose its quality.
Model Specs:
v1.0
Based on: AkashicPulse EQVAE v0.2
Dataset size: ~4.6k
Epoch posted: Epoch 15 out of 15
EQVAE used: Anzhc's EQVAE B7
v0.2 [CHECK ABOUT THIS VERSION FOR DETAILS]
Based on: AkashicPulse EQVAE v0.1
Dataset size: ~9.4k
Epoch posted: Epoch 7 out of 10.
EQVAE used: Anzhc's EQVAE B7
v0.1
Based on: AkashicPulse 4.1
Dataset size: 10.6k, repeated twice.
Epoch posted: Epoch 9 out of 10.
EQVAE used: Anzhc's EQVAE B7
Description
The model reaches its ambiguity point, between needing rescaleCFG and not needing it. Expect to turn this on and off often.
Training v0.1 for another 7 epochs with 9.5k dataset. This toned-down the model a bit, and some generations might not need rescaleCFG, while some might still need it.
Improvement in anatomy generation and stability without setting the CFG very high (8++) then rescale the CFG.
Really high steps to generate good anatomy is no longer needed. You can set steps to 28-32 and get decent results.