V2.4 Update:
Please, check on "About this version" for detailed changes on model.
Positive keys / triggers:
[score_9_up, score_9, score_8_up, ultra detailed, high quality, source_anime], Add rating_safe / rating_explicit (explicit can be combined with NSFW for a change in composition oriented towards western poses and angles in NSFW content)
Negative keys:
low quality, worst quality, jpeg artifacts, fused elements, blurry, bad anatomy, worst detail, censor, lowres, watermark, bad quality, score_6, score_5, logo, artist name, bad hands, patreon, <Capitan_error>Samplers (from better to regular):
Euler_A, Euler, DPMpp SDE, DDIM
Schedulers (from better to regular):
Simple, Normal, Karras, SGM Uniform
13 ~ 26 steps
4 ~ 7 cfg
___________________________________________________________________________________
V2 Update:
Please, check on "About this version" for important changes on positive, negative and schedulers.
Thank you for using the unmeasurable strength of VALKIRIA
___________________________________________________________________________________
Thanks to FeuerWalkure this model is possible, and that's why it's called... Valkiria
I started finetuning and making PonyMan what she asked, and realized that, at some point, I got two different safetensors... one more anime-aesthetic with 2.5D style and NSFW focused (PonyMan v3) and this one, a fine extra-detailed with more vivid colors and a large range of styles and colors.
Valkiria can understand a bit better prompts mixing natural language as well as tag-style.
Can do manly bara and feminine girls without mixing them together.
I didn't tag a single character on dataset so I WILL struggle to make "popular" character, just use a LoRa.
Just try the Valkiria, you won't get disappointed.
It origin was Pony (almost forgotten by now) so you could use
score_9, score_8_up, score_7_up,for more "anime, pony models aesthetic" outcomes.
rating_safeWill guide your picture in case you asked for spicy content but are getting just XXX
rating_explicitwill do the opposite.
NSFW,Not only changes the theme, but also has impact on colors and lights. (client asked this)
Real, 3D,Can make things a bit more realistic, but you will need to use lower CFG (0.5 to 1 point lower tha usual, never below 3)
Sampler: Euler, DDIM, Euler_ancestal
Scheduler: Simple, Normal, Karras
Steps: 10 to 22~24 (less: artifacts - more: mutations, usually)
CFG: 3 to 7 (only use 7 with more than 20 steps)
Negative words:
bad quality, worst quality, worst detail, artifacts, signed, censor, (bad hand, bad feet), long chest, chibi, cute themed, jpeg artifacts, lowres, watermarkDescription
+ Foco en estética 2D anime / western
+ Mejor acabado de imagen ( ~ 22% )
+ Mejor nivel de detalle ( ~ 6% )
+ Mejor comprensión de promtps ( ~ 12% )
+ VAE mas flexible para diferentes escenarios ( ~ 5% )
- Ya no hace realismo
- Requiere más pasos
FAQ
Comments (3)
I don't know why it only works with your pictures, I mean, I download them, drop on comfyui and change prompts and get amazing results but when I try my own setup it ends up with extremely low saturation and almost grayscale help
Strange since I don't use custom nodes. Just to be sure, check on this: 1) You downloaded the very last safetensor, 2) You are using the suggested amount of steps and CFG, 3) You are using the suggested dimensions, sampler and scheduler and last 4) Your denoise is set 1.0 if running txt2img
@capitansergio Denoise at 0,6 lol Thank you







