A new character for the Cast. Dark hair, warm mediterranean features, that easy everyday look — versatile across daylight, urban night, and quiet indoor scenes.
Fourth Krea2 character in the Cast.
Specs
Trigger word:
K_Gu14_KSCBase model: Krea2
Recommended weight: 1.0 — sweet spot
Working range: 0.8 → 1.2
Notes
Stress-tested across a wide range of prompts and contexts — character holds, flexibility is there.
All preview images here were generated with Guia alone, no style LoRA stacked. What you see is the LoRA on its own.
Pairs naturally with the ShinobiSat Krea2 Workflow — same pipeline, tested together.
If you cook something nice with her, drop it in the gallery.
Guia is the ninth drop of the ShinobiSat Cast — a growing series of photoreal character LoRAs, now across multiple base models. More coming.
— ShinobiSat
Description
Trigger word:
K_Gu14_KSCBase model: Krea2
Recommended weight: 1.0 — sweet spot
Working range: 0.8 → 1.2
FAQ
Comments (20)
I like the model, but the LoRA gives me checkerboard artefacts, which are also present in your examples.
(Used gonzalomo v4 bf16, your LoRa at Strength 1)
Thanks for the feedback, and sorry about the trouble! The LoRA was trained on a dataset with an intentionally grainy look, but you're right that on this one the effect is stronger than I wanted, and it can read as a checkerboard-like pattern.
With gonzalomo v4, you can reduce it a lot by switching the sampler/scheduler to Euler + Beta and lowering the LoRA strength to 0.9 (0.8 also works, but in some cases you lose a bit of likeness). At 0.9 with Euler Beta it stays as faithful as at 1.0.
I'll work on a new, more stable version with a subtler effect as soon as I can. Thanks for pointing it out!
@ShinobiSat Thanks for your reply and thanks for taking care of the issue. Love your support! :-)
Will try with lower strenght tomorrow. Btw. what is the training resolution of the LoRA?
@Nino26 My pleasure; the training dataset was created entirely at 1024x1024.
Hi @Nino26, I’ve retrained the LoRA and am running some tests now; it looks like the issue is resolved. There’s still a tiny bit of something if you zoom in really close, but that’s definitely due to the grain present in the dataset.
In any case, I want to run a few more tests to see how it performs overall so I have more info ready for the release—I think I’ll be able to put it out as early as tomorrow.
I’d really appreciate getting your feedback on the new version once it's out.
Thanks for now.
@ShinobiSat Great work. I'll be happy to give it some testing :)
Hi @Nino26 , the new version is available; I hope you like it, and I look forward to seeing some of your work using it ;)
https://civitai.red/models/2947377/shinobisat-cast-guia
@Nino26 Oh, I almost forgot: I ran some tests with Gonzalomo v4, and I like it with Euler Beta at a strength of 1.1–1.2.
@ShinobiSat Thank you very much. Already downloaded. Best and fastest "customer service" on here! :-)
I still have some 400+ renders in queue so testing will be a bit delayed - I will get back to you ASAP!
@Nino26 😉
Hey, @ShinobiSat - I have done some renders with the v2, (Set at strength 1.1 or 1.2) on Models SickRealism v2, GonzaLomo v4 and CyberRealistic v4. (all BF16) - See Guia Tests by NN | Civitai (the models are tagged)
When I keep the image size small the LoRA works now. But as soon as I go high res (16 MP 4885 x 3261), I get the checkerboard again. The effect is stronger with "darker" Base-Models like GonzaLomo or SlickRealism. Its weaker with CyberRealistic (which is producing light images compared to the other two)
The stronger the LoRA gets, the stronger the checkerboard effect is. At 0.8 the checker board effect is acceptable, but the LoRA looses (eg. the piercings go missing). At 1.2 the LoRA fully produces the intended character, but also the checkerboard again.
Hi @Nino26 , I’ve looked at your results, and they do indeed still show that effect quite noticeably.
I did notice something, though: when using your exact prompts and seed, I get a different result—both visually and slightly in terms of quality—which is very strange.
I checked the file hash and it matches, so a corrupted file is ruled out...
That said, I want to clarify that you currently can't get a result completely free of that "grain," because it's in the training dataset. Of course, it shouldn't look like that—it ought to be a fine grain, not the checkerboard pattern you're getting. Also, I would definitely lower the step count; I never go above 8.
In the tests I ran using your prompts, the result with 8 steps actually seemed better—not perfect, but better than with 12 steps—though that isn't the main issue here.
Right now, aside from the oddity of getting a different result than you, nothing else comes to mind, but I’ll run some more tests.
Maybe in the next few days I'll try training a new LoRA without that grain to see if the checkerboard pattern disappears.
Oh, which workflow are you using?
@ShinobiSat I guess the different result comes from the workflow. I use NegPiP and CyberKrea K Sampler for Krea2. (The NegPiP is probably the reason) - I can tell that the workflow itself works with other Character LoRAs like DomiK.
As a mixture of "grainy" models and this LoRa gets uber grainy I assume its just a bit too much ;)
@Nino26 Yes, models that are already grainy don't mix well together, but GonzaLomo v4 isn't like that—it should yield a good result.
I generated the dataset using Krea2-SAT-DirtyRealism—that's where the grain comes from—and indeed, I absolutely cannot use that specific model with these LoRAs; it produces an excessive effect.
@ShinobiSat hm If you feel happy, and try an "ungrainy" model for another character. I'll be happy to check it out again :)
@Nino26 Yes, I was also thinking about training Guia without treats...
@ShinobiSat She deserves it :)
@Nino26 ;)
@Nino26 Could you try this model? The dataset is made with another base model.
@ShinobiSat yes will do and get back to you