Hey friends~💛
It’s no secret by now… I love merging checkpoints.
Designing characters has always felt natural to me ~ whether in games, drawings, or dreams ~ so it was only a matter of time (and GPU 🫣) until I dove into crafting my own models.
With SEMImergeij, I set out to create more than a style ~ I wanted a presence.
A character with a soul, a face that tells stories, and a vibe that adapts to her world. Whether she’s framed by grungy textures, retro light, or nature’s calm ~ she remains her. Fluid in setting, yet always true to herself.
After V3, I received thoughtful feedback around representation and diversity ~ something I care deeply about. V4 leaned into that diversity beautifully, but in the process, she began to drift from her original self. With V5, I’ve brought her back ~ still open to many shapes and stories, but with her essence intact.
A familiar face. A quiet strength. A creative companion for many worlds.
Thanks so much for being here.
Stay creative, stay curious ♡
~ Reij 💖
🤗 If SEMImergeij found her way into ur world, a like would be a lovely lil hello back.
NEW🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸
Semimergeij Illustrious V6 ~ NEW
▼ recommended settings:
Clip skip 2Euler aCFG 730 stepstags🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗
Semimergeij Illustrious V5
▼ recommended settings:
Clip skip 2Euler aCFG 730 stepstags+prompt:
high_quality, highres, beautiful, detailed, (masterpiece, best quality:1.2), absurdres, 8k, HDR,-propmpt:
worst_quality, bad_quality, poorly_detailed, 🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗
Semimergeij Illustrious V4
▼ recommended settings:
Clip skip 2DDPM & Euler aCFG 3-7+prompt:
best_quality, absurdres, ultra-detailed, highly_aesthetic, 8K, UHD, highly_detailed_eyes, depth_of_field, subject_focus, -propmpt:
worst_quality, bad_quality, poorly_detailed, young,extra_fingers, malformed_hands, deformed_feet, 🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗
Semimergeij Illustrious V3
The main changes in V3 were to be able to use the model without negative prompt, as well as to keep the style and colors at CFG 1 to enable faster image processing. From 20 steps the model can so be used with CFG 1. I would recommend an adetailer or facefix for further distances. The resolution for image generation starts as usual at 1024*1024, but the model can also be used for larger images. Visually, V3 is more back to the roots, just galloping back onto the pony: it is more based on the look of the pony version ~ a mix of art with drawn contours including a semi-realistic filling.
As a fan of playing with light and shadow, I love setting individual highlights and bringing certain accents into the light. V3 is therefore also based on shadow and light details to enable more play here.
▼ resolution 1040*1510 CFG 1
▼ resolution 832*1216 CFG 1
▼ resolution 1040*1510 CFG 2
▲ example with -prompt: worst_quality, bad_quality, poorly_detailed
recommended settings:
Euler a 20+ stepsCFG 1+Clip skip 2_ might wanna be ur friend for tagsresoltion: 1024*1024 ++prompt example
masterpiece, best_quality, detailed_eyes-prompt example
worst_quality, bad_quality, poorly_detailed🌺~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗
Semimergeij Illustrious V2
a tinybit more details and extras ~ ♥
In addition to the included models, semimergeij V2 includes Loras that I trained specifically for this merge to allow for warmer, softer light with the associated softer tones despite maintaining dynamic colors. In addition, details are understood differently than in the previous version.
Please let me know if u like the direction and how we can develop Semimergeij further.
I'm looking forward to ur results and hope u have a lot of fun being creative.
recommended settings:
Euler a 25-40 stepsCFG 4-7Clip skip 2_ might wanna be ur friend for tagsresoltion: 1024*1024 +☺ Showcase resolution: 1040*1510. I would recommend using ADetailer for further distance.
+prompt example
masterpiece, best_quality, amazing_quality, very_aesthetic, absurdres, detailed_eyes, masterpiece, best_quality, high_quality, HDR, 8K, masterpiece, best_quality, absurdres, ultra-detailed, highly_aesthetic, highly_detailed_eyes,-prompt example
worst_quality, bad_quality, poorly_detailedcomparison lookalike examples:
🌺~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗
Semimergeij Illustrious V1
hey guys,
after starting with IL, I felt in love with the closeness to the prompt, depictions of the hands, expressions, motion and colors. especially because I like to create portraits, but also have my NSFW moments, the faces and bodies are always particularly close to my heart. so I made an effort to tinker SEMImergeij as an IL version. maybe u want to ride the pony illustrious? 😅
✨setting examples:
Clip skip 2
Sampling method: Euler a
Sampling steps: 25+
CFG Scale: +-6
resolution: 1040x1510 / 832x1216
Hires x1,5 / Hires steps 20 / Hires denoising 0,2 / Hires CFG 1
ADetailer for further distance recommended
quality +prompt: masterpiece, best quality, [...]
incl. alternative: realistic body
quality -prompt: worst quality, bad quality, [...]
I am happy if u enjoy and stay curious abt ur results ~♥
*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*
With V9 I would like to return a bit to the semi realism touch. With a view to my love for comics, I would like to present “V9” to you.
semimergeij - soft and silky within, hard rendered - wants to grow into a digital art style with a semi-realistic 3d look in a drawn shell with a fantasy anime-semireal-inspired character design.
I am grateful for your feedback and am really looking forward to seeing your creations!
🌺~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗~🌸~💗~🌺~💗
Semimergeij Pony
examples how to bake showcase images of V9 from 20 to 100 steps with ADetailer and without hires, fix (how to get a similar style without adetailer, please read below)
prompt:
PonyScores, 1 woman, blush, horny, beautiful 30 age years old sexy woman, blonde, catlady, cattail, catears, eating a chocolate cookie, face focus, off the shoulder, winter background, dimmed light, indoor
negative prompt:
PonyScores, multiple parts, multiple images, young, cute
Sampling method Euler a Automatic
clip skip 2
steps 20 -100
CFG 9
Distilled CFG Scale 3,5
Width 720
Height 1280
seed 177756388
example with hires, fix and ADetailer
with steps 20
Upscaler nmkdSiaxCX_200k
Upscale by 1,5
Hires steps 0
Denoising strength 0,2
Hires Distilled CFG Scale 3,5
Hires CFG Scale 9
For show case image style I would recommend using an ADetailer.ADeailer positive prompt: PonyScores,
ADetailer negative prompt: PonyScores,
If u wanna keep the style of the showcase images u can keep writing 'PonyScores' in positive and negative prompt (+ in ADetailer prompt if u like to)
if u wanna skip ADetailer I would recoomend using scores but for fullbody I would recommend using it anyway.here are some examples with scores:
examples of V9 without Adetailer, hires or Upscaling
in increments of 10 steps I wanna show u examples of steps 20 - 100 with scores
prompt:
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, close up, upper body, blush, long brown hair, multicolored hair, pastel pink ruffled dress, hair ribbons, feeling of summer, sfw, from above, hugging a bouquet of flowers, simple background, pastel green colored background, sitting, giant sunflower, looking at viewer, beauty, harmonious, face focus, smooth skin, glossy skin, and hyper-realistic with a luminous sheen, porcelain-like effect, slight gradients ,depth,glowing effect, three-dimensional feel,
no negative prompt
Sampling method Euler a Automatic
clip skip 2
steps 20 -100
CFG 9
Distilled CFG Scale 3,5
Width 720
Height 1280
seed 2539837132
2 examples with hires, fix without ADetailer
Upscaler nmkdSiaxCX_200k
Upscale by 1,5
Hires steps 0
Denoising strength 0,2
Hires Distilled CFG Scale 3,5
Hires CFG Scale 7
For full body images I would recommend using an ADetailer.*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*
In case u are a comic fan as I am and would like to generate comic pages, I would recommend to write"letters, text, watermarks, numbers" and so on in the negative prompt to avoid random texts
examples how to bake showcase inspired images of V9 without ADetailer and hires / the style is just similar
without scores, just with "PonyScores"
In case u wanna scip Adetailer, Hires and scores I have tried to find a prompt description of the style based on the showcase images.
prompt parts baked with:
PonyScores,
soft, beauty, harmonious, face focus, smooth eyes, soft skin, oily skin, glossy skin, and semi-realistic, round face, luminous sheen, porcelain-like effect, slight gradients, depth, glowing effect, three-dimensional feel, front studio light, mat eyes, warm light,
[ur vision]
neg. prompt:
PonyScores, young, glowing iris
example:
prompt:
PonyScores, soft, beauty, harmonious, face focus, smooth eyes, soft skin, oily skin, glossy skin, and semi-realistic, round face, luminous sheen, porcelain-like effect, slight gradients, depth, glowing effect, three-dimensional feel, front studio light, mat eyes, warm light, 1 woman, blush, woman, blonde, long hair, catlady, cattail, fluffy catears, eating a chocolate cookie, face focus, off the shoulder, winter background, indoor, dimmed light
negative prompt:
PonyScores, young, glowing iris
Sampling method Euler a Automatic
clip skip 2
steps 20 - 69
CFG 9
Distilled CFG Scale 3,5
Width 720
Height 1280
seed 6969
example with hires, fix / without Adetailer:
Upscaler nmkdSiaxCX_200kUpscale by 1,5
Hires steps 0
Denoising strength 0,2
Hires Distilled CFG Scale 3,5
Hires CFG Scale 9
For full body images I would recommend using an ADetailer.all versions VAE included
I am happy if u enjoy and stay curious abt ur results! ♥
Description
VAE included
V3 baked with my own models and Loras (including ILustREAL v3, ILustREAL v4, ILuvorija, Kawaij v1, SemILust v3, Semimergeij-IL v2)
FAQ
Comments (17)
great checkpoint found one small issue though. if you use dark skin in prompt it have a very high probailty of giving tan lines as well no matter what you do.
thank u very much 4 ur feedback? even if you write tan_lines in the negative prompt?
@reijlita yep
@reijlita its kind of wierd its not complete tan lins its like only the breasts alot of times but if i generate 10 images 7-8 get tanlines and 2 without. i dont mind kind of like tan lines but just thought you should know
@SaintTheet thank u very much 4 the info. I've seen what ur talking abt^^ tried "brown_skin"^^ and it works 4 me ;)
I used to get this (and other issues) with alot of models, especially models with a bias for non-dark skin. I'd end up over-tagging for dark skin with tags like dark skin, brown skin, tan skin, etc with various weights. Then I'd negative prompt things like white skin, tan lines, peach skin, etc. The overall effect was kind of all over the place, with some decent gens and some that lost prompt adherence in other ways due to (in assuming) the over-focus of skin color that would deprioritize other tags essential to the specific scene or composition I was trying to create. A solution I found was using actual ethnicities such as Native American or Latina to reorient the model's baseline subject away from western European skin tones. See if that helps 👍
Gorgeous checkpoint. It does struggle a bit with fantasy colors outside of standard fantasy creatures, but otherwise easy to work with and produces consistently beautiful pictures.
Thank u for the feedback ☺ I'm happy if u enjoy using the semimergeij :)
I've now tested your model for several hours and created over 300 images.
Positive: It consistently produces beautiful images.
Negative: As the CFG increases, the model loses its semi-realism and tends more towards anime.
Ethnic characteristics cannot be reproduced. All generated faces are largely the same, even with tokens like "African woman" or "Asian woman."
Although the skin color also changes in CFG-1, the face remains rather Caucasian.
Age information such as "mature woman" or "older woman" only becomes effective at higher CFGs. Specifying the age in numbers, such as "50yo" (50 years old), also has no effect at low CFGs.
The model also doesn't handle complex prompts in real language well. While it creates a nice image, it's far from what's described. Only at higher CFGs (>6) does it get somewhat closer to the prompt, but as already described, it loses much of its semi-realism.
Conclusion: If the uniform faces of a young woman with light skin are enough for you, you'll certainly have fun with the model. For me personally, it's not suitable, and the 200 buzz for Early Access was a waste.
thank u for ur feedback.
Most IL models are very compatible with “tags” (danbooru tags/“_”), hence the note in the instructions and recommended settings: _ may want to be your friend for tags. Tags mean keywords that are, in most cases, strung together. (For example: note for +prompt and -prompt).
Tags can also be helpful when we want to describe age (mature_woman,old_woman,young_woman,middle_aged_woman,elderly_woman,grandmother), facial features, expressions, movements, skin color (light skin, medium skin, dark skin, pale skin, tanned skin, brown skin, ...), body type (e.g.:petite,slim,curvy,plus_size,athletic_build,muscular), ...
Most tags can be used from CFG1 onwards, but their implementation becomes more detailed if we use a higher CFG, from 2+.
Since we are a huge community in the AI world with different access to hardware, the latest semimergeij version is also intended to generate images on low CFG. This can be particularly helpful for small VRAM as it reduces processing time.
Being comfortable with long sentences, it might be helpful to use chat.gpt to convert these sentences into (danbooru) tags. Another option can be using @schmedes GPT. He wrote an excellent article if someone is new to the IL world.
Here, individual words or letters can create a completely new image. ☺
Have fun being creative and thanks for the tip to show more variation in showcases in the future.
@reijlita I'm aware that most IL models are very compatible with "tags" (danbooru tags/"_").
I've also used prompts that consist only of tags. However, your model ignores tags like (mature woman, middle-aged woman, 50 years old, etc.).
As already mentioned, tags are ineffective for ethnic groups, which makes your model very inflexible. Even if the prompt only consists of "Asian woman" or "African woman" or "Japanese woman," etc., the facial features largely remain unchanged. Please test it yourself.
The fact that the CFG value affects VRAM usage is new to me, incomprehensible, and also incomprehensible.
@martindieter As I already said, I want to leave space for creative diversity with my models, to be able to describe what you would like to represent. Diversity does not mean stereotype, so my models adhere to every day possibilities that IL has to offer.
As stated in the previous examples, most of the generations can be done from CFG 1, but it is ideal to start from CFG 2. All of the examples I gave above work.
I am grateful for any criticism, especially because the voice of the community is important to me and I like to further develop my models. However, I find baseless statements inappropriate. Test the prompt examples as just mentioned and see for yourself within seconds that they work. thanks
If you haven't done yet, clear your negative prompt. You will get much better outputs without a negative prompt.
Reijilita is taking over Illustrious. May you keep up the excellent work.
Be the Dreamshaper of this era!
Thank u very much for the kind words. But u are the dream designer urself, I like to tinker around making keys for new rooms that we can explore together ^-^ and if u have fun with it, it's a pleasure for me too ☺
Unbelievably amazing!
thank u very much ^-^ I am happy if u like generating images with semimergeij^-^