100+ Full models comparison 

Part 2 – Max details and realism test

Most models can be placed inside one of those three groups: realistic, semi-real and anime, but just how detailed can a realistic model get? What will happen if we try to make the remaining two categories as photorealistic as possible? Will raising the initial rendering size increase quality or produce a lot of mangled bodies? Is using many quality-improving prompts a good way to produce astonishing results?

Test 2 was designed to answer those and similar questions. Let’s see how well each model fares when quality prompts are raised to the max and anything even slightly connected to paintings, drawings or cartoons is strictly forbidden.

In Test 1, we compared over 100 models with a short, simple prompt.

We have over 100 full models and 900 loras – click to see them all.

Test 2 – Max details and realism test

The main focus of this test is a young woman. On top of this prompt, there are as many quality/realism/details improving positive and negative prompts as the stable diffusion token limit can handle and then some more.

The limit above which the AI starts ignoring additional prompts is about 75 positive and 75 negative tokens. This tokenizer is a handy tool to check how many tokens your prompt is using. Of course, you can use more, but the render will become increasingly unresponsive to them.

Despite this being a realism test, I fully expect that most models will do their best at producing the most unrealistically beautiful woman possible. ‘Real’ realism is something entirely different. It requires much more fine-tuning on a case-by-case scenario, which isn’t suitable for running over a hundred vastly different models.

Each picture was genereated by using this prompt:

/render /parser:new /clipskip:1 /guidance:10 /sampler:dpm2m /size:640x832 /steps:75 /images:6 /seed:6243589 (young woman:1.2), (standing:1.2), [(NSFW:1.7), (naked:1.7)] (35mm:1.3, full-body shot:1.25), (photorealistic:1.3), (realistic:1.25), raw photo, [(lowres:1.4), (blurry:1.6), (worst quality:1.5)], (best quality:1.2), good quality, detailed skin, soft lighting, masterpiece, (detailed background:1.3, Fujifilm XT3), [(anime:1.4), (cartoon:1.4), (painting:1.4), (illustration:1.4)], 8k uhd, hdr, hyperrealism, sharp focus, (realistic hair texture:1.1), (realistic texture:1.2), best shadow, (insanely detailed:1.3), finely detailed beautiful eyes, (absurdres:1.2), [(white background:1.4), (bland background:1.3), (semi-realistic, cgi, 3d, render, sketch, drawing:1.2), text, cropped, out of frame, low quality, jpeg artifacts, duplicate, morbid, mutilated, poorly drawn, mutation, deformed, bad anatomy, bad proportions, (deformed iris:1.3), (deformed pupils:1.2), disfigured, gross proportions, unfocused, artifacts, overexposed, underexposed, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, wip, stains, long neck, bad art, (zombie:1.2, dog:1.2), (bokeh:1.4)] <put your chosen full model here>

Why those settings? 

Click on each for more information.

/parser:new – allows the usage of weighted prompts like (standing:1.2).

/clipskip:1 – the default one.

/guidance:10 – higher than in the previous test to increase the AI’s attention to prompts.

/sampler:dpm2m – as long as it’s not the unpredictable k_euler_a, any would do.

/size:640×832 – a good size to obtain more details and focus on a single character.

/steps:75 – to give all renders an extra bit of polish.

/images:6 – the number of images was decreased in order to reduce the number of columns from 4 to 3, so each example picture can now be bigger.

/seed:6243589 – randomly chosen to ensure all models use the same one.

(young woman:1.2) – the main prompt.

(standing:1.2), (35mm:1.3, full-body shot:1.25) – prompts that were used to avoid a gallery full of face pictures.

[(lowres:1.4), (blurry:1.6), (worst quality:1.5)] – my favourite quality boosting combo.

[(anime:1.4), (cartoon:1.4), (painting:1.4), (illustration:1.4)] – in the previous test, some models couldn’t decide which direction to follow. This will allow them to focus on realism and should have an interesting effect on anime models.

Other prompts used – just as before, choosing those was hard. After browsing the most popular examples, recipes, and recommendations, I’ve finally put some together. It’s a mix of prompts with confirmed boosting effects and others that are commonly used. Prompts like ‘cell shading’ and others that pushed the model too much into a specific direction were avoided.

[(zombie:1.2, dog:1.2)] – apparently, this increases the quality of certain models without affecting others.

<put your chosen full model here> – replace this part with the name of any of the full models given below for example <reliberate>, or <breakanime>, or <majic5> etc.

100+ full models comparison part 2

Gallery 1 

Contains: paragon, reliberate, dreamful3, realisian, majic5, majic, ned122, bra5, bra4, revani, realflex, duchaiten, degen, edge, hardblend, mars, realgalaxy, bismuth, level4, urpmcun, urpm13, urpm, realvis20, realvis13, proto53-real, proto58-scifi, protonova, hasdx, deliberate, deliberate2, dreamshaper, cyberreal, 

Gallery 2 

Contains: comics-vision, ghostmix, lyriel16, lyriel, monstergirls, aio, kstyle, chilloutmix, guofeng, lofi, comics-babes, chikmix, icomix, men, pixar, clarity, liberty, babes, perfectworld4, perfectworld, marvels, kemono, epic, ducjourney, wool12, papercut2, ink-punk, rainbow, expressionism, synthwave, wavy, photoreal2