Today's simple discovery5. (Lora and TI(textual inversion) mix combination)

In the middle of generating veriety version of the vampirella.



parameters


(full body:1.2) photo of Vampirella, solo focus, look at viewer, very long hair,

micro bikini,

muscular, curvy, (huge breasts:1.35), abs, narrow_waist, huge pelvic, (thick thighs:1.35), high_heels, boots,

(dynamic pose, fighting stance:1.2), blood_puddle, partially submerged, body half-submerged,

(full moon), forest, Blood everywhere,

(best quality, realistic) <lora:VampirellaV2:0.4> v4mp1r3ll4, <lora:angel_wings_v10:0.6> (angel_wings)


Negative prompt: (breasts out, nude:1.4), extra hand, extra foot, close up, (multiple view), missing finger, extra fingers, fused fingers, cropped, (crossed eyes, worst quality,bad quality:1.4),monochrome, greyscale, (text, signature, watermark:1.4), mutated


Steps: 15, Sampler: DPM++ 2M, Schedule type: Karras, CFG scale: 7, Seed: 1489340367, Size: 512x640, Model hash: b89ba7f884, Model: meichidarkV38_pruned, Denoising strength: 0.55, Clip skip: 2,



This one

As you can see, [face, hands] are different quality from others. 

In my setting, most of cases has not this much of a differences.

I wonder why in this case?




Steps up to 20 denoising 0.55

Still different.





Denoising 0.5 15steps

Still different.





Denoising 0.5 20steps

Not the steps and denoising issues.





Remove angel_wings Lora. (denosing 0.55 15steps)

Slightly better.






How about remove vampirella Lora? (denoising 0.55 15steps)
That breasts parts bad.





Without TI (Textual Inversion) : v4mp1r3ll4  (denoising 0.55 15steps)
Still bad.





Remove all Lora and TI  (denoising 0.55 15steps)
Not much as I think?
 But Slightly better.






Maybe Tiled VAE cause this problems??

Tiled VAE on





Tiled VAE off

Same result tiled VAE on







Conclusion : angel_wings Lora causes main quality issue, in this case. 


Not for Lora mix or not for meichidark check point (semi realistic model).









Every image has exif info

Comments

Popular posts from this blog

Solved: Facial Errors in Stable Diffusion Pony Models

Some of results (character full images)

Todays simple discovery (uncannyValley_ilxl10Noob)