If you feel like showing some love, awesome.
But the best kind of support?
Use the model. đ ď¸
Post your results. đźď¸
Let the art speak. đŻ
Razâs Fantasy Style - (FLUX & ZIT)
Razâs Fantasy Style is an ongoing fantasy art direction created to produce consistent artwork for the tabletop roleplaying games I play, mainly Pathfinder 2e. The goal is to have a unified visual style for characters, creatures, and scenes used across different campaigns, instead of relying on mismatched fantasy art from many sources.
There is already a huge number of fantasy models and LoRAs available, many of them very good. The LoRAs shared under Razâs Fantasy Style are not meant to reinvent fantasy art or replace existing styles. They exist simply to provide consistency, using a look that fits the kind of fantasy illustrations I enjoy using at the table.
I am sharing them because I find the results interesting and useful for my own games, and because sharing tools and experiments is part of the fun. Also, I am Canadian, and we tend to share things when we think others might enjoy them too.
Fantasy - Mk.1 - (FLUX & ZIT)
Fantasy Mk.1 is the first LoRA released under Razâs Fantasy Style and defines the core painterly realism look of the project. It focuses on a semi-realistic fantasy illustration style inspired by classic tabletop RPG artwork and fantasy hardcover covers.
A key aspect of this version is the subtle grain and texture present throughout the image. This grain is intentional and inspired by printed fantasy art, where texture often comes from traditional painting, early digital workflows, or print artifacts. Preserving this texture helps avoid overly clean or glossy results and keeps the artwork grounded and illustrated.
The style was shaped by a mix of influences, including classic D&D and Pathfinder artwork, fantasy artwork concept collections on Pinterest, and extensive prompt-driven experimentation using themes such as âfantasy style woman, man, animals, creatures, digital painting, grain, epicâ explored across GPT, Gemini, and Flux.
Fantasy Mk.1 is designed to work across a wide range of subjects, including characters, creatures, animals, and environments. Backgrounds are detailed and readable, featuring familiar fantasy locations like forests, ruins, cities, taverns, dungeons, and temples, so each image feels like a complete illustration rather than a character floating in an empty abstract space.
Dark Fantasy - Mk.1 - (ZIT & FLUXsoon)
Dark Fantasy Mk.1 is the first dark fantasy LoRA released under Razâs Fantasy Style, and the first experimental model trained specifically for Z Image Turbo (ZIT).
This version emphasizes dark atmosphere, heavy mood, and ominous tone, blending classic fantasy illustration with horror elements and a light touch of grimdark. It is not extreme or nihilistic grimdark... the focus is on mood, menace, and visual weight rather than despair or excess brutality.
ZIT excels at capturing global tone, brightness, and atmosphere, and as a result this version renders significantly darker than the Flux counterpart (coming soon), with deeper shadows and a more oppressive overall feel.
The dataset and inspiration follow the same foundations as Fantasy Mk.1, but pushed firmly into darker territory:
villains, vampires, necromancers, monsters, skulls **lots of skulls** cursed figures, savage warriors, and ominous environments.
Subjects include:
Female characters (young evil figures, cursed beauties, ancient crones)
Male characters (muscular brutes, gaunt sorcerers, necromancers, savages)
Monsters, undead, animals, and creatures
Dark fantasy environments and narrative scenes
Dark Fantasy Mk.1 is designed to produce cohesive, illustrated dark fantasy artwork suitable for tabletop RPGs, where characters and scenes feel dangerous, corrupted, and grounded in a shared grim world.
Description
đ Fantasy - Mk.1 (Z-Image Turbo)
This is the Z-Image Turbo version of Fantasy - Mk.1, trained on the same dataset as the Flux-based release.
While the underlying training data is identical, the final look differs due to Z-Image Turboâs rendering behavior. This version produces a slightly different interpretation of the same fantasy material, with cleaner shapes and a more immediately readable illustrated style compared to Flux.
Fantasy - Mk.1 (ZIT) maintains the same goal of creating consistent, tabletop-ready fantasy artwork inspired by classic fantasy book illustrations and TTRPG visuals. Characters, creatures, and environments remain semi-realistic and painterly, but Z-Image Turboâs handling of lighting, color, and structure results in a distinct visual output.
If you enjoyed the Flux version, this release offers a parallel interpretation of the same fantasy style, optimized for users who prefer Z-Image Turbo as a base model.
đ§ Recommended Settings
Sampler / Scheduler:
Euler_Ancestral / Beta (recommended)
Euler / Beta â grittier, rougher painterly look with slightly less detail
DPMPP_SDE can be used for a cleaner or more controlled finish
DDIM_UNIFORM works well as an alternative scheduler
Steps:
9
FLOW Sampling:
3 to 7
LoRA Strength:
1.00 (tested and recommended)
FAQ
Comments (13)
A truly great Lora that delivers excellent results :)
Thanks a lot, I really appreciate it! Glad youâre getting good results with it đ
Just amazing.
Aww thank you so much đ¤
I just can't get behind the ZI-T trend. Maybe it's good for fantasy/anime/toon, but everything I've seen so far as far as realism, it seems like a step backwards. Sure, it's faster, requiring less steps, but that shows in the overall quality. Before anyone comes at me with replies related to any image I've posted not being the quality I'm hinting at, I'll let it be known now that 100% of my highly polished, high quality realism images are NOT posted on my profile. So spare me that nonsense.
ZI-T to me is like Pony to Illustrious. We are too quick to move to something new, when all that's needed is better training and optimization of what we already have.
This isn't directed at you Raz, it's just a general statement of ZI-T and what I've noticed as a whole with it. Keep up the good work.
Totally fair take. I donât see ZIT as a universal replacement either. Model choice really depends on what youâre doing and what youâre optimizing for.
For me, what hooked me isnât âfaster = better,â but how differently it behaves under training. It reacts very strongly to atmosphere, painterly intent, lighting mood, and stylization in a way I wasnât getting as easily before. That doesnât make it magically superior. It just opens doors I personally find interesting.
Speed is also a big factor. With FLUX I usually queue overnight runs, 2 to 6 minutes per image depending on settings. With ZIT Iâm closer to 30 seconds to 1 minute, even at higher latent sizes.
I still really like FLUX, but ZIT gives me that ânew toyâ energy and refreshed my motivation. On my older system, it lets me iterate quickly as a LoRA creator rather than waiting between tests. Any current ZIT flaws just feel like new challenges to solve in training.
And thanks for using my models, genuinely appreciated. Hopefully youâll keep "Boobs-testing"... I mean stress-testing them⌠đ
ZI is different than FLUX, FLUX can't do this style like ZI does, and vice versa. Idk how you could possibly see ZI as a step backwards in realism. It's prob the best realism model there is objectively speaking. But to each their own.
In my opinion Z-Image is way ahead of the competition for pure photorealism. It also looks very versatile. The thing is, a wrong setup can yield very bad results. Try to find a nice workflow that was made recently, use the BF16 version, and you might be surprised. It will really come to life when the full model gets released though.
very nice
Thank you VelvetS, huge fan of your work too!
Your loras always work so cleanly. Have you ever considered releasing a basic lora training guide? I'd be curious to see what optimizer, timestep, lr, and decay you use etc.
Thank you so much for these kind words...
In all honesty, all my LoRa training are done here on the CivitAi trainer.
The most important thing is to make sure the dataset is crystal clean.
For the settings I could write up something about dataset prep and the CivitAi parameters I use in different type of lora.
I have some other lora project in the making but I'll be sure to make an article on the topic soon
@Razane That'd be lovely! I'm experimenting with different parameters, and softwares and optimizers, also experimented with lokr, the results are ok, but could definitely be improved and so your input would be really appreciated!















