A practical guide based on personal experience

Last updated: [2025/08/25]

• chatGPT assisted design case

• Artist style mix excel Update

Introduction

First of all, sincere thanks to Kairen92 for bringing such an excellent checkpoint model to the Civitai community. This checkpoint mod is a slight tweak on the original noobai model and designed for male characters (also work well on female characters). This moudle is probably one of the best noobai you can use on this site. This article is applicable to most noobai fine-tuned models that use danbooru as training set material. This article is a collection of my experiences and strategies, which I hope will help newcomers and those facing difficulties when generating images.

Why Choose NoobAI?

NoobAI is a fine-tuned model based on Stable Diffusion XL (SDXL). Among many AIGC image generation models, NoobAI XL stands out for its ability to balance compositional stability, aesthetic harmony, and expressive character rendering. It excels in character-focused illustrations, supporting a semi-realistic render while preserving stylized detail and emotion. The XL version is trained using larger resolution footage (from 512*512 to 1024*1024), resulting in sharper images, but essentially works the same way.

One important note: NoobAI is relatively weaker in environmental scenes compared to some illustration-based models.

To use NoobAI effectively, you must speak its language: clean, Danbooru-compatible, and concise prompts.

1. Use Danbooru-Based Tags

NoobAI XL is trained on a Danbooru-style tag system. Prompts written in natural language often lead to unstable or inaccurate results. For best results, use tags directly sourced or adapted from Danbooru format. Avoid verbose descriptions, emotional prose, or anything not expressed through tags. This requires some familiarity with Danbooru tagging, which can be intimidating. I suggest first outlining the composition in your mind (or using a drawing tool), then using AI (like ChatGPT) to convert that idea into structured tags. Next, manually trim irrelevant or ineffective tags.

Although the prompt words are in no particular order, it is recommended to clarify the structure: 1boy/2boys/6+boys+character appearance/clothing/action/expression+scene/lighting+quality label+stylization/lora.

("BREAK" and paragraph breaks will not be recognized, feel free to use them to separate different parts)

An example (form Kairen92 Creator Profile | Civit):

1boy, bara, dark, night, looking at phone, illuminated by the light of a smartphone, nude, muscular, large pectorals, puffy nipples, male masturbation, cum, projectile cum, ejaculation, lying in bed, orgasm, nywlub \(bulwyn\), very awa, masterpiece, best quality, absurdres, newest, perfect quality, detailed background, intricate details

Frequently used tags:

(character: beard, haircut, fan works, expression)1boy, solo, male focus, mature male, bara, muscular,

(lighting)sidelighting, backlighting, vignetting, bloom, lighting particles, dramatic lighting, natural lighting, rim lighting, volumetric lighting,

(quality)very awa, masterpiece, absurdres, best quality, perfect quality, detailed background, intricate details

Redundant or conflicting tags—like muscular, muscular male, abs, pectorals, toned body—don’t strengthen results. They confuse the model and dilute your intent. Adding or removing even a single tag can drastically affect output. For example, if you useopen top, you will get a character with open shirt. You may be able to see the character's chest and abdomen. If you addabs at this time, you will get a more complete abdominal structure, which will make your composition change unexpectedly.

Use one tag per idea.

Focus on identity-defining tags:

solo, 1boy, mature male, short hair, muscular

Keep each category (appearance, clothing, pose, setting, style) to the essentials. Use clear, simple action tags like:

hug, holding, looking at viewer, sitting, standing, smile

Important points:

• To control character size and placement, rely on tags like full body, cowboy shot, portrait, close-up, or use specific clothing/body part tags like black shoes, shoe soles.

• For scenery-first compositions, try these two approaches:

  1. Generate an ideal background first. Then manually paint in the character or use inpainting for refinement.

  2. Generate the character first, then expand outward using outpainting to build up the background.

• Tags like dynamic pose, dynamic angle, dynamic view increase instability but sometimes yield surprisingly refined compositions.

• NoobAI is not primarily designed for NSFW generation, so without tags like penis, anus, the model may apply natural censorship.

2. Tags for Multi-Character Scenes

NoobAI does not prioritize prompt order, which means multi-character scenes often result in chaos without careful control.

Suggestions for two-person (or more) scenes:

• Use 2boys or multiple boys to clarify your intent

• Use pov (point of first view) to get interaction in front of the camera (Usually use 2boys, using1boysometimes causes pov to be unrecognizable)

• Assign distinct attributes to each person when possible (e.g., “one standing behind, one seated”)

• Use relational tags: height difference, looking at another, hug, hand on another’s face, couple, distance between

• Avoid stacking identical tags across characters (like both marked as main focus)

• To prioritize one character: place their tags first, then de-emphasize the other using blurred, background, or facing away

Tags like (1boy, ***) can sometimes help stabilize scenes but in my testing, they don’t offer much advantage compared to well-designed attribute tags.

3. Artist Tags vs LoRA Tags

Because NoobAI uses Danbooru-trained data, it supports a wide variety of artist styles. In most cases you can just search for the name of your favorite artist and insert it into your prompt phrase. If you want to create fan works, danbooru also provides a large number of character tags. Please confirm copyright requirements before posting. I like to draw more universal characters to avoid this problem, haha. It’s difficult to find the right combination that suits your taste. Overusing a single artist may also lead to uniform-looking work (in my opinion).

For convenience, I made a small excel to randomly select a set of artists. You can copy the result directly (search each artist’s style to see if they match your current project). I'm considering integrating it into a script or plugin if I have time in the future.

Meanwhile, LoRA models give even more control—but with a different purpose.

1) Artist Tags

• Format: (name:weight)&artist:name:weight(artist:) annotation does not work, only useful for checking labels

• Influence style, brushwork, composition, palette, and expressive logic

• Stacking multiple artists blends their aesthetics

Example:

breezem18:0.6, (artist:kazuki_yone:0.4)

2) LoRA Tags

• Format: <lora:ModelName:weight>

• Inject specific traits, character poses, or stylistic features during generation

• LoRA often affects both rendering behavior and character proportions

Example:

<lora:raccoon21_2_IL:0.8> adjusts body type and applies a softer tonal shift, along with a furry brush style. Try to minimize the stacking of a large number of loras, which will greatly increase the risk of overfitting and ruin your picture. Some artist tags can sometimes not be mixed with some lora, which will lead to the same disaster (especially those painting styles with obvious brushstrokes).

4. Reflections and Lessons Learned

1) Simplicity wins.

I used to stack dozens of tags in hopes of “covering everything.” But with NoobAI, fewer and cleaner tags produce more powerful images.

2) Artist tags shape mood.

Changing (artist:kazuki_yone) to (artist:greenteaa) can shift your visual tone from ethereal to cinematic—without changing any content.

3) Stylized realism is achievable.

Blending soft-brush artist tags with sharp-detailed LoRAs results in painterly realism, my favorite hybrid between believability and beauty.

4) Prompting is storytelling.

You’re not just writing prompts—you’re writing visual scripts. Keep them structured and intentional, and NoobAI will finish the image with elegance.

5)Another way to refine.

I haven't mentioned the inpainting method and workflow in this article. This is because in general, using inpainting for detail adjustments can be very unstable with noise reduction values ​​(depending on the image quality), and parameters like mask drawing/feathering range also depend on personal feel.

So I try to avoid this approach, but instead, you can use the inpainting idea to manually crop the image and then use upscaling to enlarge it. This idea is a bit obvious: it works for any image, whether it's a person or a landscape; you can make detail adjustments as needed, while still adhering to anatomy in most cases, so there won't be "strange" bodies; it takes up little space and is fast to generate.

Compared with inpainting, manual cropping and using the image stretching function will achieve better results. Pay attention to the calculation of the resolution in the process. For example, if the size after cropping becomes 1100*520, the actual required enlargement ratio is: (1024*1024/1100/520)^(1/2)=1.35

6) An example to use chatGPT to assist (You can also try other language-based models such as deepseek).

Here I give an example of my own work. First, use natural language to describe the scene and composition elements in the conversation with chatGPT. I use "I want to draw a shepherd lying on a large meadow, holding a piece of hay in his mouth, with a flock of sheep. The clothes should be consistent with the character's occupation, pay attention to complying with Noobai standards". Then I got the following prompt phrase:

1boy, muscular, broad shoulders, large pectorals, mature male, strong arms, stubble, deep-set eyes, tanned skin, brown eyes, tousled black hair, short hair, relaxed expression, confident smirk, dry grass stalk in mouth, lying down with hands behind head, one knee up, visible abs under open shirt, rustic shepherd outfit, linen shirt with rolled sleeves, open collar, wide belt, rugged trousers, worn boots, shirt partly unbuttoned, chest slightly exposed BREAK vast green meadow, soft grass, distant sheep grazing, gentle hills, summer sky with drifting clouds, wildflowers, wooden crook on the ground nearby, sunlit pasture, peaceful rural setting, fluttering grass in breeze, sheep scattered around, some close to foreground, distant mountains on horizon, clear blue sky, tranquil outdoors BREAK soft sunlight, warm backlight, bloom, cinematic glow, shallow depth of field, high detail, natural shadows, golden hour tones, painterly realism, very awa, masterpiece, perfect quality, detailed background, realistic rural lighting

Then manually deleted and adjusted unstable composition elements. Here I adjusted the hair color, belt style, action, expression, and selected a few stylized tags that I liked. Finally, you can get a complete prompt phrase that suits your personal characteristics. All you have to do is select a few pictures you like from the many generated works to perform face repair, inpainting, redrawing, filters and other detailed work.

Although chatGPT does not directly call Danbooru's API, nor does it automatically access Danbooru or its database, resulting in many of the prompts it outputs not actually meeting the specifications, this saves creators a lot of time. You can try to enter the concise examples into Chatgpt and make it imitate it, or repeatedly emphasize that making Chatgpt save you "need to strictly conform to the noobai specification."

Due to the rules of chatGPT, nsfw content is mostly impossible to generate (but fortunately it's not complicated, you can search for relevant keywords on danbooru, and you can get a lot of relevant phrases from the post pictures).

Boys are warm, wonderful, loving creatures. There are many more wonderful things to be had than to admire their sex or their genitals.

Closing Thoughts

NoobAI doesn’t reward overengineering. It rewards clarity, intent, and artistic instinct.

Check out more excellent works, check out more danbooru, and you can constantly get new inspiration and accumulate experience.

Once you understand how it flows, NoobAI stops being a machine—it becomes a camera guided by your imagination. Although I'm not a professional, I hope to have more interesting, lovely and shining things to share with you in the future. Have fun!

PS: This article will continue to be updated. Feel free to PM me or just comment here.

PPS. I would like to recommend an original couple from my work: YS and Ivan (based on reality). I will post more story about them in the future.