Hi, everyone. Back again in the post about my activities using Stable Diffusion. This time I will discuss some things that I usually encounter when trying to mix a large amount of LoRA...
Stable Diffusion users use LoRA when they have difficulty bringing up "something" just by relying on the prompt. LoRAs are generally specific things such as certain clothes, a pose, an artist's style, and others that don't have many sources.
(according to my knowledge) they will usually only use 1-3 LoRAs each time using Stable Diffusion either with a local computer or a website that provides models and LoRAs to generate images with supporting prompts.
However, because I am not an expert in processing prompts, I rely more on LoRAs so that the results of the generated images are more varied. Unfortunately, I am not always satisfied if I only rely on 1-3 LoRAs, so I mix more LoRAs until I find the results I want.
Below are the events that I experienced this week (September 23-29) when mixing many LoRAs...
(unique case) Different from others
Normally I would create 4 images at once in "one generate". But this time I accidentally created a "creature" even though there was no creature in the prompt. While the other 3 images show normal results (displaying 1 girl because the prompt says 1girl)
 |  |  |
|---|
![prompt]
# Prompt 1girl, upper body, looking at viewer, <lora:skijocek-10:-.6:lbw=ON> <lora:banana-10:.3:lbw=ONL> <lora:gyokai:.6:lbw=1,0,0,0,0,0,0,0,0,0,0,1,-1,1,0,0,0> <lora:maimans-08:-.3:lbw=1,0,0,0,1,1,1,0,0,-1,0,0,-1,1,0,0,0> <lora:fairi-us-08:0:lbw=1,0,0,0,1,1,0,0,0,-1,0,0,0,1,0,0,0> <lora:llsoccer-08:-.6:lbw=OL> <lora:flat2:0:lbw=ONL> <lora:artbytokiame:-.6:lbw=ONL> <lora:Houtengeki:-.6:lbw=ONL> <lora:majoneko-08:-.5:lbw=1,0,0,0,0,0,1,0,0,1,1,0,0,1,0,0,0> <lora:maimans-08:0:lbw=1,0,0,0,1,-1,1,0,0,-1,-1,0,0,-1,0,0,0> <lora:fairi-us-08:-.3:lbw=OUTL> <lora:hike-vanitas-lora-10:-.5:lbw=ON> <lora:warboob-lora-10:-.5:lbw=ONL> <lora:yuzusoft-chibiV2:-.6:lbw=INOUT>, absurdres # Negative Prompt ((normal quality, low quality, worst quality):1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), lowres
Slightly Overcooked
Overcooked or Overbaked in LoRA usually refers to an image result that looks strange or too much. For example, if a LoRA has a soft style, then if it is too much, the image will look crispy or will look blurry if it is too little. *You know what I mean?
Well, these 4 images are the results if the x value of one of the LoRAs (lora:Lora_name:x:lbw=???) is too large. And to overcome this, of course, by lowering the value little by little until it produces an image that is no longer strange.
 |  |
|---|---|
 |  |
! [Prompt]
# Prompt 1girl, upper body, looking at viewer, <lora:skijocek-10:-.6:lbw=ON> <lora:banana-10:.4:lbw=INOUT> <lora:gyokai:.7:lbw=1,0,0,0,0,0,0,0,0,0,-1,1,-1,1,0,0,0> <lora:maimans-08:0:lbw=1,0,0,0,1,1,1,0,0,-1,0,0,-1,1,0,0,0> <lora:fairi-us-08:-.7:lbw=1,0,0,0,1,1,0,0,0,-1,0,0,0,1,0,0,0> <lora:llsoccer-08:-.8:lbw=OL> <lora:flat2:-.5:lbw=INOUT> <lora:artbytokiame:-.8:lbw=INOUT> <lora:Houtengeki:-.6:lbw=ALL> <lora:majoneko-08:-.4:lbw=1,0,0,0,0,0,0,0,0,1,1,0,0,0,0,0,0> <lora:maimans-08:-.3:lbw=1,0,0,0,1,-1,1,0,0,-1,1,0,0,1,0,0,0> <lora:fairi-us-08:0:lbw=OUTL> <lora:hike-vanitas-lora-10:-.8:lbw=ON> <lora:warboob-lora-10:-.6:lbw=ONL> <lora:yuzusoft-chibiV2:-.1:lbw=INOUT>, absurdres # Negative Prompt ((normal quality, low quality, worst quality):1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), lowres
Then I changed the values of Lora majoneko and hike-vanitas to 0, so the result became like this...
 |  |
|---|---|
 |  |
! [Prompt]
# Prompt 1girl, upper body, looking at viewer, <lora:skijocek-10:-.6:lbw=ON> <lora:banana-10:.4:lbw=INOUT> <lora:gyokai:.7:lbw=1,0,0,0,0,0,0,0,0,0,-1,1,-1,1,0,0,0> <lora:maimans-08:0:lbw=1,0,0,0,1,1,1,0,0,-1,0,0,-1,1,0,0,0> <lora:fairi-us-08:-.7:lbw=1,0,0,0,1,1,0,0,0,-1,0,0,0,1,0,0,0> <lora:llsoccer-08:-.8:lbw=OL> <lora:flat2:-.5:lbw=INOUT> <lora:artbytokiame:-.8:lbw=INOUT> <lora:Houtengeki:-.6:lbw=ALL> <lora:majoneko-08:0:lbw=1,0,0,0,0,0,0,0,0,1,1,0,0,0,0,0,0> <lora:maimans-08:-.3:lbw=1,0,0,0,1,-1,1,0,0,-1,1,0,0,1,0,0,0> <lora:fairi-us-08:0:lbw=OUTL> <lora:hike-vanitas-lora-10:0:lbw=ON> <lora:warboob-lora-10:-.6:lbw=ONL> <lora:yuzusoft-chibiV2:-.1:lbw=INOUT>, absurdres # Negative Prompt ((normal quality, low quality, worst quality):1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), lowres
It's too much
Just like above, these 4 images fall into the overcooked category. But the difference is that in this case, the LoRA value is too large so the result becomes like this...
 |  |  |  |
|---|
! [Prompt]
# Prompt 1girl, upper body, looking at viewer, orange \(theme\), <lora:skijocek-10:-6:lbw=ONL> <lora:banana-10:.4:lbw=ONN> <lora:gyokai:.4:lbw=1,0,0,0,0,0,0,0,0,0,-1,1,-1,1,0,0,0> <lora:maimans-08:0:lbw=1,0,0,0,1,1,1,0,0,-1,0,0,-1,1,0,0,0> <lora:fairi-us-08:-.7:lbw=1,0,0,0,1,1,0,0,0,-1,0,0,0,1,0,0,0> <lora:llsoccer-08:-.8:lbw=OL> <lora:flat2:-.5:lbw=INOUT> <lora:artbytokiame:-.8:lbw=OUTD> <lora:Houtengeki:-.9:lbw=ALL> <lora:majoneko-08:-.4:lbw=1,0,0,0,0,0,1,0,0,1,1,0,0,1,0,0,0> <lora:maimans-08:-.3:lbw=1,0,0,0,1,-1,1,0,0,-1,1,0,0,1,0,0,0> <lora:fairi-us-08:0:lbw=OUTL> <lora:hike-vanitas-lora-10:0:lbw=ALL> <lora:warboob-lora-10:-.4:lbw=ONL> <lora:yuzusoft-chibiV2:0:lbw=INOUT>, absurdres # Negative Prompt ((normal quality, low quality, worst quality):1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), lowres
The value of Lora skijocek here is -6 whereas the ideal value of a Lora is around -1 to 1 (in some Lora even 1.5 can still produce an image that is not overcooked). I finally replaced it with -.6, resulting in the image below...
 |  |  |  |
|---|
So those were some of the events I experienced over the past week while using Stable Diffusion and playing with LoRA. Maybe next week I'll tell a similar story but with equally unique (maybe?) images.
Thank you for visiting and reading my post this time. See you in the next post...
All images and dividers were created by me using Stable Diffusion and various LoRAs.