Model Introduction
Welcome to KiraDepth, a v-prediction model that aims to create a saturated 2.25/2.5D aesthetic.
This is a simple merge I made for myself to do two things:
Learn how to merge models, and
Create a style of model I enjoyed yet couldn't find a v-prediction equivalent.
Recommended Settings
This is a v-prediction model. If you don't know what that is or don't know how to setup your local environment to support v-prediction models, please read NoobAI-XL's original model page here or the guide posted by L_A_X here.
Parameters
These are the settings I typically use when generating. Experiment to see what works best for you.
CFG Scale:
3-5Steps:
28-40Sampling Method:
EulerorEuler ASampling Scheduler:
NormalorSGM Uniform
Prompts
These are the prompts I use throughout all the example images. Again, experiment to see what works for you.
Positive:
masterpiece, best quality, very aesthetic, good quality, newest, very awa, absurdres, highresNegative:
worst quality, worst aesthetic, bad quality, normal quality, average quality, oldest, old, early, very displeasing, displeasing
Merge Recipes
Inside this dropbox contains the workflows for each version (minus some A/B txt2img testing). These images have a ComfyUI workflow embedded, so you should be able to import them directly into ComfyUI.
I'm continuously experimenting with new nodes and ideas to try to get the style I am looking for while hopefully not overcooking the model.
Note about v1.0 workflow: The node SDXL Preset Merger is a custom node I wrote (that I've since phased out) to emulate a similar functionality seen in supermerger's MBW preset. I have not published this nor do I have plans to.
Afterword
Thank you for checking out this model page. I am new to the scene of model creation and am pretty slow at figuring this stuff out.
This is simply a project I sought out for my own creation that I decided to share with the community.
If you have questions about my merging process or are willing to point me to resources/guide me in merging models, write a comment or DM me!
Description
Adds a bit more depth (namely better rendering/shading) back into KiraDepth while retaining most of the original 2.25/2.5d aesthetic. First experiment with adding in an Illustrious model, but I think it came out okay.
Models merged:
KiraDepth (v1.1): https://civitai.com/models/1503720?modelVersionId=1753103
Nova Reality XL (Illustrious v3.0): https://civitai.com/models/453428?modelVersionId=1605991
Plant Milk (Coconut): https://civitai.com/models/1162518?modelVersionId=1549172
Illustrious XL (v0.1): https://civitai.com/models/795765/illustrious-xl
noob_v_pencil-XL (v3.0.0): https://civitai.com/models/1027203?modelVersionId=1787376
FAQ
Comments (15)
Switched to a somewhat different workflow and was finally able to get Illustrious models merged in. I think I managed to get a similar feel to v1.1 while increasing rendering/shading performance to more accurately get that 2.25/2.5d feel.
v1.2 is I think a step in the right direction for the feel I am going for. Better shading, slightly more detail while not feeling messy or distracting. And better eye stability at lower resolutions (hopefully no more messy, jumbled eyes!).
Next steps I think is to look into Illustrious 2.0 and Rouwei, maybe start looking into more LoRAs that could enhance the feel. Unsure which I would want to tackle first, but I have ideas.
hi,is a EPS?
This is a v-prediction model.
While it does include weights from epsilon prediction models (NoobAI-XL EPS and Illustrious models, see each version for what models are merged), the resulting model is still a v-prediction ZSNR model.
There's a link in the description to the original NoobAI-XL model page that has a guide on how to setup your machine for local generation with v-prediction (Automatic1111 does not support v-prediction models. reForge and ComfyUI do).
Merge recipes have been posted for all versions. If you're curious about how I cooked (and maybe why I don't belong in the kitchen), check the version description for a workflow image.
EDIT 2025-06-04: I've changed to dropbox for easier uploading and access for you guys. Check the description for the link.
I guess ImgBB won't really work.
I'll create a dropbox with workflows embedded instead.
So I've tried for a while, and I hadn't really come up with anything that fits this "style" and improves substantially on this model.
I've also tried creating a new one from scratch, but was unsatisfied with the style compared to previous versions.
This model was an ultimately an excuse for me to dip my toes into model merging, and I'm glad I did. I'm having a lot of fun researching new and existing methods to get a style of model I truly enjoy.
That being said, v1.2 is the final version of this model. This page will not be getting updates except for answering questions.
I will still be merging my own models and maybe, when I find something I really like, I'll share it with the community like I did this model. I'm a lot more knowledgeable than I was before, but I still have a lot to learn and understand.
I think v1.1 is better. My Go to Model for now :)
Glad to hear it!
I can't believe that such a well-crafted model with such amazing results has only had 500 downloads in six months.
Thank you!
I have tested about 8000 different Illustrious and NoobAI models so far (https://civitai.com/user/Lizardon1025) but this one is still my favorite Anime model. It's level of detail and consistency is absolutely amazing and none of the most popular models even come close to this. It might be the most underrated model out there and still deserves way more attention.
The only tiny downside of this model is that it sometimes creates light particles in images (which might not be desired) but putting the tag "light particles" in negative easily fixes this problem.
@daddyq If you are still interested in creating models, I still think this model deserves newer versions. Although I don't really know what to improve at this point...
I am glad you still enjoy this admittedly old model.
If I were to update this, it would be to move to Chenkin as the base rather than Noob (would mean it would no longer be v-pred, but I'm find with that trade-off for newer dataset cutoff).
That being said, I'm currently into inferencing rather than merging/training rn. Anima is really fun right now.
@daddyq A newer dataset cutoff would certainly be a good thing. But while Chenkin claims to have a knowledge cutoff of Nov 2025, everything after spring 2025 seems to undertrained. I've written an article about different knowledge cutoffs if you are interested:
https://civitai.com/articles/14008/character-knowledge-cutoff-comparison
Anima is a really interesting base model, although it still needs a lot of refinement to reach the level of modern Illustrious models. As it understands both Danbooru tags as well as natural language, it has a lot of potential but it might also be harder to refine properly...
@lizardon1024 Thanks for the article.
It does seem to confirm my experiences with other models. Chenkin does seem better at reproducing artist styles than Noob, but that is irrelevant in a style merge (and when people inevitably use LoRAs).
WAI and Hassaku I was hesitant on using in the past as
1) They are epsilon-pred models which aren't the best models to merge into v-pred w/o losing what makes v-pred special (better dark scenes, true white and black rather than a gray mess, etc. Learned this the hard way with 1.2), and
2) Just merging in CLIP does give new knowledge, but the prompt adherence in my experience normally suffers as a result. I've had many woes trying to get RouWei CLIP merged without it destroying my style and having prompt adherence suck. Maybe I'm just doing it wrong?
JANKU's license doesn't allow merging, so I steer clear of that one.
In the end, this model may just have to stop being v-pred (and likely Noob in general unless 0.3 Chenkin improves) for me to look into updating it. Either that or I look into doing something like Cat Tower does and do a small fine-tune pass to keep the style I want.
@daddyq I agree that in the moment there isn't really a good alternative to NoobAI for KiraDepth without ruining its style. I guess that's just the fate for such an already perfect model like yours ^^
Small fine-tune passes sound great but I honestly don't know what to improve besides knowledge...