Advanced Image Generation

Layers

Layers are "blocks" that make up an image model. Think of Stable Diffusion as a tower made of floors - each "block" has its own job:

• Early layers - Capture low-level details (edges, colors, textures)
• Middle layers - Capture patterns and shapes (faces, clothing, body structure)
• Late layers - Capture semantic meaning (concepts like "cat" vs "dog" or "anime style" vs "photo")

LoRA

LoRAs are training methods for fine-tuning large models. When you have an image model, LoRAs teach it new tricks by injecting themselves into certain layers - like remodeling specific floors in the tower.

CLIP Skip Layers

Stable Diffusion uses CLIP (Contrastive Language-Image Pre-training) as a text encoder to convert your prompt into embeddings. CLIP is basically a translator between you and the image model. Skipping layers cuts off higher-level understanding.

Many anime models and LoRAs are trained with CLIP skip values in mind (most often 2). Skipping a layer might give you more anime-like or stylized results.

Scheduler

The scheduler controls the denoising process during generation. It strongly affects image sharpness, coherence, and prompt adherence.

Euler a - Fast

• Produces sharp, detailed images with strong contrast
• Great for anime, stylized art, or sketch-like detail
• Sometimes unstable with realism - might distort fine textures or faces
• When to use: Fast, creative, detailed outputs for anime-style or concept art

DPM++ SDE Karras

• Produces smooth, high-quality gradients and soft transitions
• Great for smaller details
• Each render comes out differently - good for organic variation
• When to use: Photorealistic renders, natural lighting, painterly styles, soft transitions

DPM++ 2M - Deterministic

• Produces smooth, consistent results with high prompt fidelity
• Good balance between speed and accuracy
• When to use: Realism, portraits, consistent styles where stability matters

DPM++ 2M Karras

• Produces smoother gradients, better texture, and higher detail fidelity than vanilla DPM++ 2M
• Slightly slower than Euler A, but higher quality for realism
• When to use: Photorealistic images, portraits, complex textures - best for polished, stable look

Prompt Strength

Prompt strength changes the CFG scale - the higher the value, the more the system follows your image prompt.

• Low CFG (1-5): Model pays less attention to prompt. More artistic freedom, dreamier compositions. Risk: Image drifts from what you asked.
• Medium CFG (6-9): Balanced adherence vs creativity. Common sweet spot: 7-8. Keeps both accuracy and aesthetics.
• High CFG (10+): Strong adherence to keywords. Can become oversaturated, harsh, or "burnt". Risk: Weird artifacts, stiff or unnatural images.

The actual weight of + is 1.1^x (where x = number of +). Meanwhile - is 0.9^x.

# of signs	Weight of -	Weight of +
1	0.9	1.1
2	0.81	1.21
3	0.729	1.331
4	0.6561	1.4641
5	0.59049	1.61051
6	0.531441	1.771561

DISCLAIMER: THIS TABLE ISN'T A MATH TABLE! IT'S A VISUAL REPRESENTATION OF THE FINAL RESULT!

You can game the system: (keyword-)2 gives equivalent of (keyword)1.8. More complex: (keyword---)2 calculates as (keyword 0.9^3)2 = 0.729 * 2 = (keyword)1.458

Combinations of keywords and weights can achieve certain desired looks. Example with painterly_shading:

With Painterly_shading+++	Without

You can change:

• Style / Stylization - Overall look
• Shading / Lighting - Volume and depth
• Line-art - Definition of edges
• Proportions - How subjects are depicted
• Texture / Detail - Surface details
• Composition / Framing - Arrangement
• Simulation - What medium the picture simulates
• Quality - Overall quality

Image prompts have only 550 characters! You can circumvent this by using background, clothing presets, and physicals for more stylized prompts.

Preset rules on Soulkyn: 32 characters, 18 tags.

Tricks:

• Use semicolons ; to pack multiple meanings into one tag
• Use underscores _ to pinpoint meaning for AI
• Use :2 to apply stronger weight (normalized to prevent explosion)

Example: PaintFuka Character

Physical tags: Painterly; Soft; No_lineart, 8K;16K; Hdr, Blue Lips, Slim Eyebrows; Fair_skin:2, Long Two-Toned Dark_blue Hair, Realistic_character_anatomy, High Quality, Highest_quality; Extremely_detailed, Adult_woman:2; Mature_woman:2, Realistic Details, Painterly Texture, Cute;Kawaii, Digital_painting:2, Thick Thighs; Slim Waist, High Contrast, Painterly Details, Long Blue Hair; Blue Eyes, Dramatic; Epic; Emotional

Clothing tags: Hyperdetailed_eyelashes, High_contrast:2; Shadows:2; Light:2, Realistic:2; Anime:2, Canvas_tone_texture; Hyperdetailed, Hyperdetailed_skin:2, Expressive;Hyperrealistic:2, Hyperdetailed_body:2; Cute:2, Glow_diffusion; Saturated_colors:2, Hyperdetailed_shading, Hyperdetailed_textures:2, Hyperdetailed_hair, Glowing_eyes:2; Big_eyes:2, Hyperdetailed_clothing:2, Hyperdetailed_eyes:2, Art Bloom; Soft; Hard; Hairstrands, Painterly_skin_softness; Kawaii:2;, Soft Light; Thick Thighs; Matte_skin:2, Realistic Anime; Soft Shading

Background: Glowing Highlights, Detailed Rendering, Vibrant Colors, Saturated Colors, Ambient Lighting, Anime Art Style, Cell Shading, Rim Lighting, Digital Painting, Semi-Realistic, Soft Shading, Smooth Gradients, Dramatic Lighting, Polished Finish, Professional Artwork, Clean Lineart, High Contrast, Artstation Quality; Highest Quality

Breaking down part of the physicals:

• Painterly - More painted details
• Soft - Softer picture
• No_lineart - No line-art (negative-like tag)
• 8k; 16k / High Quality - Very high quality desired
• Hdr - Deeper contrast between lights and shadows
• Blue lips - Character has blue lips
• Fair_skin:2 - Really want fair skin
• Long Two-Toned Dark_blue Hair - Long hair with two shades of blue
• Realistic_character_anatomy - Realistic anatomy

It's time to use all of our newfound knowledge to be more creative with our prompts! For this we will be using PaintFuka's Background and Clothing.

Here's a comparison between no stylized prompts and stylized one:

Not stylized	Stylized

First let's describe a picture that we want to create:

• Amount of people: 1
• Perspective: cowboy shot (knees-up perspective)
• Character wearing a white summer dress
• Character is running while extending one hand
• Expression: Smiling
• Background: Not important, can be blurry to make character stand out
• Quality: Extremely detailed, realistic illustration

Taking this list into account let's prepare all the loras, scheduler, and keywords.

Because we want the picture to resemble a detailed realistic illustration let's use LoRAs just for that!

Here is the result of applying Micro / Fine details :6 and Realistic Illustration :6 LoRAs:

With LoRAs	No LoRAs

Now let's put clothing on her and make her run!

Final prompt:

Advanced settings:

• LoRA: Micro / Fine details :6, Realistic Illustration :6
• Prompt Strength: 11.0
• CLIP Skip: 2
• Scheduler: DPM++ 2M
• Image Orientation: Landscape
• Seed: Random
• Prompt: Portrait
• Background: PaintFuka
• Clothes: PaintFuka
• Image Prompt: (solo, alone)+++, (macro_detail, clean_composition, color_balance, 8k_realistic_render, professional_product_shot)+++, (close, cowboy_shot)++++, (white_summer_dress---)2, (running)+++, (reaching_out, one_hand, outstretched_hand)+++, (smiling, Happy, open_mouth, closed_eyes)+, (blurry_background--)2

Final result:

Now let us analyze what went into the prompt and how it works:

• (solo, alone)+++ - makes the character appear by themselves in the picture
• (macro_detail, clean_composition, color_balance, 8k_realistic_render, professional_product_shot)+++ - these tell AI about the quality of the picture we want to get
• (close, cowboy_shot)++++ - This tells AI that we want a really close cowboy shot
• (white_summer_dress---)2 - This forces AI to give our character a white summer dress
• (running)+++ - This tells AI we want our character to be running
• (reaching_out, one_hand, outstretched_hand)+++ - This precisely tells AI what kind of pose do we want
• (smiling, Happy, open_mouth, closed_eyes)+ - This defines the expression of our character
• (blurry_background--)2 - This forces AI to give us a blurry background

As an extra, this is the effect of adding just one more LoRA:

The most common issue is degrading quality - some users call this "carpetization" because broken images look like close-ups of carpet surface.

This is most likely due to:

• Too much weight
• Too much strength on LoRAs
• Conflicting tags

Our website offers AI-assisted image editing. This is an experimental feature that costs souls due to resource expense.

To edit an image:

1. Go to character's gallery
2. Choose image to edit
3. Press EDIT IMAGE button

Unlike image generation, inpainting uses descriptive prompting. Explain to AI what you want in understandable sentences.

Example: "Change character's hair to white and skin to black"

The result: