santa hat
deerdeer nosedeer glow
Sign In

My Current Setups Explained

16

Here I explain which models, LoRAs, embeddings and other settings I use to create my 512x512 images in different styles (upscaling is covered in this article).

REGULAR STYLE

MODEL: endlessreality_V2

SAMPLER: DPM++ 2M Karras

SAMPLING STEPS: 30

CFG: 7

POSITIVE PROMPT:

glamour photo of {subject}:0.8, cocktail dress, seductive

<lora:popular:1.2><lora:masterpiece:0.4>

<lora:epiNoiseoffset_v2:0.7>

<lora:add_detail_v5:0.7>

<lora:FilmG2:0.5>

<lora:BimboMakeup_v1:0.1>

<lora:bimbostyleThreeU:0.2>

<lora:wowifierV3:0.1>

<lora:detailmaker:0.1>

<lora:more_details:0.1>

NEGATIVE PROMPT:

photograph by bad-artist,

kkw-Extreme-Neg-PH,

ng_deepnegative_v1_75t,

bad_prompt_version2,

negative_hand-neg,

easynegative,

(worst quality, low quality, normal quality:1.7), lowres,

lr:0.9

PINUP STYLE

{as regular style, with some variation of the following as the first line:}

pinup, (art by Gil Elvgren:1.3), {subject}:0.8

CYBERREALISTIC STYLE

MODEL: cyberrealistic_v32

SAMPLER: DPM++ SDE Karras

SAMPLING STEPS: 30

CFG: 7

POSITIVE PROMPT:

{photo style} photo of {subject}:0.9, {important feature}, {camera properties}, {lighting}, in the style of {photographer}

<lora:popular:0.99><lora:masterpiece:0.05>

<lora:epiNoiseoffset_v2:0.7>

kkw-ph1:0.8

NEGATIVE PROMPT:

CyberRealistic_Negative-neg,

photograph by bad-artist,

kkw-Extreme-Neg-PH,

ng_deepnegative_v1_75t,

bad_prompt_version2,

negative_hand-neg,

easynegative,

(worst quality, low quality, normal quality:1.7),

lowres,

(realisticvision-negative-embedding:0.8),

lr:0.9

EXPLANATION AND LINKS

Model:

I almost exclusively use EndlesssReality_v2, because for me it achieves the best balance between realism and flexibility, as well as between fidelity and creativity. Think of "perfectly realistic dragon scales" as an example or analogy: Most models will give me either a too cartoony dragon or just a snake (because snakes are real and dragons are not). EndlesssReality understands what I want, and often gives me what I didn't know I wanted.

I used to think that some models are superior in terms of pure realism, but now I'm not so sure. Or maybe I'm just better at tweaking this model and don't know the others well enough.

(Note: EndlesssReality14 just came out, and I'm not getting good results with it with this setup. I'll probably stick with V2, but may experiment with 14 in the future.)

link

I consider ReV Animated the anime or cartoon equivalent and use it occasionally:

link

A-Zovya RPG is nice for SciFi and Fantasy (I sometimes get similar results to ReV Animated with it):

link

Cyberrealistic_v32 is the most photorealistic model I have seen so far. Its compositions are still much less interesting than those of EndlessReality, but much better than most other "realistic" models. I'm experimenting with generating the base image with EndlessReality and inpainting faces with Cyberrealistic (more cartoony faces being the main drawback of EndlessReality for purposes of realism), for an overall realistic but interesting look.

link

Sampler:

Other good samplers for EndlessReality are:

DPM++ SDE

DPM Adaptive

DPM++ 2S a Karras

DPM++ SDE Karras.

DPM Adaptive in particular has some cool poses sometimes. However, I get the best results with DPM++ 2M Karras. Note that for simpler, cleaner and more realistic models (like A-Zovya Photoreal), Euler is usually a good sampler (but not at all for EndlessReality). If you want a sampler that is "good at everything" (including realistic pictures), just use DPM++ SDE Karras.

Sampling Steps:

More may be better but give diminishing returns. For me, 30 seems enough. (Note that DPM Adaptive ignores this setting and decides its own sampling steps.)

CFG:

I use 7 (and only experiment with other values when using Multidiffusion).

Prompts:

"glamour": The difference between "glamour photo" and "photo" often isn't large but can be important.

"{subject}:0.8": This is usually a Textual Inversion, but not always. The reason I use the weight 0.8 is to lower the impact of undesirable artifacts from training. With a good embedding I go up to 9, but not higher. I go much lower (0.6, 0.5 or even 0.3 or 0.2) for LoRas.

"cocktail dress": This falls under additional detail and is replaced by "bikini" or other outfits and props.

"seductive": This is for overal pose and facial expression. If I want a smile, I change it to the more common prompt "seductive smile".

<lora:popular:1.2>: All the most popular prompts ("best quality, masterpiece" etc.) in one file. For some "over-the-top" concepts, even higher weights (up to about 3) can work well. (For some reason, 0.99 does not work in Multidiffusion, but 0.9 does.)

Here are the exact prompts whose weights this LoRA increases:

"#:0.04|highly detailed, sharp focus, intricate, smooth, elegant, 8 k, fantasy, cinematic lighting, cinematic, masterpiece, matte, photorealistic, 4 k, 8k, beautiful, volumetric lighting, dramatic lighting, detailed, realistic, intricate details, ultra realistic, high detail, centered, hyper detailed, 4k, 8K, hyperrealistic, hd:0.2"

link

<lora:masterpiece:0.4>: From the same creator who made the "Popular" LoRA. It is more specialized (see details below), but works well for me at a lower weight.

Here are the exact prompts whose weights this LoRA increases:

"#:0.12|~incredibly detailed masterpiece portrait, color contrast, well defined face with interesting features, hyperrealistic, wearing unique fitting aesthetic clothes, focal points on attractive face, vivid colors on accessories and matching outfit, complex patterns in eyes resembling the universe, motion and action in background, leading lines and visual cues directing attention to eyes, image split to rule of thirds"

link

Another great conceptmod by ntc is Final Boss, which I use for some fantasy, sci-fi or overall dramatic pictures.

Here are the exact prompts whose weights this LoRA increases:

"final boss++:0.4|final boss%{random_prompt}:-0.1"

link

<lora:epiNoiseoffset_v2:0.7>:

Improves lighting and contrast.

link

<lora:add_detail_v5:0.7>:

Detail Tweaker: Massively improves detail and quality. I prefer this over the other detail enhancers I have tried (see below), but find that they significantly boost Detail Tweaker if added at a low weight (at higher weights, they change the composition in ways I don't like).

link

<lora:FilmG2:0.5>:

This makes the picture look less digital and more like an old-school analog photo. Goes well with my regular style and with photorealism.

link

<lora:BimboMakeup_v1:0.1>:

Only to slightly change the look of the woman's face. Can obviously be dialed up, but will zoom in on the face in that case.

link

<lora:bimbostyleThreeU:0.2>:

This changes the woman's look significantly (especially at higher weights): facial features, facial expression, breast size, even type of outfits. At around 0.7 - 0.9, the image usually "breaks". How this LoRa affects the image depends hugely on the textual inversion used (that is, on the woman's natural body shape). I use it at low weight by default to make features lighter, rounder and softer and to reduce stark shadows (which indirectly helps improve detail, see example below). This version works better than version Five with this setup.

link

<lora:wowifierV3:0.1>:

(See Detail Tweaker above.)

link

<lora:detailmaker:0.1>:

(See Detail Tweaker above.)

link

<lora:more_details:0.1>:

(See Detail Tweaker above.)

link

"photograph by bad-artist": This isn't just a regular prompt ("art by a bad artist"), but the recommended way to trigger the "Bad Artist" negative embedding:

link

"kkw-Extreme-Neg-PH":

It's recommended to use it alone, but I get the best results just adding it to the front of the list.

link

"ng_deepnegative_v1_75t":

link

"(bad_prompt_version2:0.8)":

link

"negative_hand-neg":

link

"easynegative":

link

"(worst quality, low quality, normal quality:1.7), lowres":

Recommended by the creator of EndlessReality_V2.

"lr:0.9":

Sharpens image.

link

"kkw-ph1:0.8":

EasyPhotorealism

link

"(realisticvision-negative-embedding:0.8)":

link

"CyberRealistic_Negative-neg":

link

Three good resources I rarely use:

"style-empire": This is great for Victorian age pictures, but can also help with high-rise buildings and cityscapes.

link

"<lora:LowRA:0.6>": Better dark details (dark theme). I like my pictures well lit, though.

link

"<lora:BreakRealize_rL2170:1.2>":

This acts as a kind of slider between anime and realism, and I sometimes use it at -1.2 or 1.2, but it has some side-effects. I've dropped it from my realistic setup.

Here is an example of trying to make an image look more realistic (at 0, 0.8 and 1.5, respectively):

link

16