Model  testing

 

There has been overwhelming tendency in text-to-image model development, going from versatile and rich artistic styles like Stable Diffusion 1.5 (oct.2022)
and DreamShaper 5 (apr.2023) towards more realistic / photo-quality models starting with Flux-1 (aug.2024). Some reasons for that are:

  • the technology reached a point (2024) when an astonishingly good photographic quality was possible on an average hardware.

  • most viewers/consumers prefer nice realistic pictures not only of pretty women but pretty anything, so commercially speaking...

  • many artists and art guilds panicked and went ballistic at anybody using their art for training, very few are looking for a middle ground.

  • when one optimizes a model it's much easier to tell if a face is realistic while to say if one picture is better Chagall than another is much trickier

  • ...and for completeness, let's say that fake news and any dubious media have had their renaissance.

 

The variety of text-to-image generation models for Stable Diffusion is huge. It's easy to get lost in attempt to find the one which fits your taste and objectives.

There are two major models depots/hubs: (Sept.2025)

Hugging face hosts over 900,000 models, 200,000 datasets, 300,000 demo apps, and has over 1.2 million registered users.

Civitai has 0.4 million registered users, but it excels in providing a collaborative and innovative space for creators of all levels (for non-British users only).

The models tested here were used with Stable Diffusion, mostly using ComfyUI with a few going for Automatic1111. It is unlikely that you would be able to reproduce (generate) exactly the same images, but if you use the same model and parameters you will have very close results. The number of the models actually tested exceed by a factor of 3 to 10 the ones shown here. In the list are only models I consider worth mentioning.

Another peculiarity is that apart of some models being good at realistic settings and others at arty ones, the ability to draw in particular style or like a particular artist vary a lot. If you are aiming to a specific style or artist (and you have some extra time on your hands) experiments with different models

 

Model Example galleries Download location SD parameters / CompfyUI workflow template
Stable Diffusion 1.5

The first successful SD model, it can be very good for art with a bit of a passion. Realistic abilities, ...well there are more recent much better model for this.

Text-to-image playground, arty and playful.
Most of the images have been processed additionally out of Scripthea.
https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 v1-5-pruned / workflow_default.0.cftm
DreamShaper 5

The best model for doing art I know (the investigation is ongoing). It could do realistic images as well but "DreamShaper-XL lightning" is better and faster.

Surrealism (mostly) by Yerka & co

bridges_forest , the_city , clouds , colour , minimal , riverside

https://huggingface.co/Lykon/DreamShaper/tree/main dreamshaper_5 / workflow_default.0.cftm
DreamShaper 6.31

Very similar to DreamShaper 5 even though my taste slightly leans toward DS5. I tend to use them both. Take a look at the galleries and judge for yourself.

Impressionism style
   nature gallery 1 , nature gallery 2
https://huggingface.co/Lykon/DreamShaper/tree/main dreamshaper_6 / workflow_default.0.cftm
DreamShaper-XL lightning

A good DreamShaper family genes, really fast and great for realism.

Realistic images
   nature gallery 1 , nature gallery 2 , polis gallery
https://huggingface.co/Lykon/dreamshaper-xl-lightning dreamshaper_xl_lightningt / dreamshaper-xl-lighting-wf.0.cftm
Aura-Flow 0.3

Very good for realism, moderate speed, a good overall model.

Realistic images
   nature gallery 1 , nature gallery 2 , polis gallery
https://huggingface.co/fal/AuraFlow-v0.3 aura_flow_0.3 / aura_flow_0.3_wf.0.cftm
Qwen-Image 2.5

Very good for realism, speed is not its best feature. For art its quality depends on the painter/style, overall - a very good model.
For details see the technical report.

Naive art/Pop-art inspired by David Hockney
   nature gallery , root gallery , polis gallery
https://huggingface.co/Qwen/Qwen-Image Qwen-image-fp8 / qwen_api.cftm