Lora: Hews Style
Hews has a visual language that's instantly recognizable, and this Lora exists to help you understand it, not just imitate it. Rather than a plain style filter, think of it as a reverse-engineered guide to how the look comes together — the structure and choices behind the finished piece.
How to use it
Generate an image normally with your prompt and seed (FOLLOWING THE ANIMA GUIDELINES) and adding "@hews" as an artstyle. Then, if you want to have some reference to study a simplified version, run that same prompt again, adding "sketch, greyscale" — this strips away color and rendering, leaving just the underlying composition and linework. Comparing the two versions is where the real value is: it turns the model into something you can learn from, not just something you can generate with.
PS: Specific Version notes might be added in the version note.
Versioning LOG:
alpha (v. 0.1) [01/07/2026]:
This is an early release, and it's still rough around the edges — the dataset, captions, and parameters are all being actively refined. Right now, expect:
A solid first pass at Hews' aesthetic
Enough flexibility to experiment with
A model that'll keep improving with feedback
A beta is planned, with better captions and more faithfull output. If you've got dataset or captioning suggestions, they're genuinely welcome.
⚠️⚠️⚠️ Disclaimers: ⚠️⚠️⚠️
Built on the ANIMA architecture. ANIMA is still evolving, so full compatibility with future versions isn't guaranteed, if something looks off, it's probably version drift, and updates will follow.
Experimental, and behavior may vary across models and settings. Use it consciously.
This exists to help people study and appreciate Hews' style, not to pass work off as the artist's own or misrepresent their intent. Use it with that in mind. So we encourage you to follow him on socials, support him and to not misuse this tool. The responsibility for any misuse of this model shall be yours, and yours only.
Description
FAQ
Comments (3)
May I know how do you make these LoRA's, if you don't mind?
Depends my friend, depends on what you have to train and the quality you need to have. We lack hardware resources so we use when possible civitai trainer, but recently we are using diffusion-pipe to train our character LoRAs locally. We've found highs and lows of this.
the hard part is collecting the data and refactor the basic caption. once the data is collected we use w.d. to caption with tags and qwen 3 vl to produce a natural language caption. Then we refactor the caption manually (not too much, but without this step the output is proved to be mediocre at best). Then we train and test. (simple as that).
CivitAITrainer is unstable for anima, the dual captioning is not handled as it should, and the bucketing locked at 1024**2 is annoying (it lowers the quality for the style if you have bigger ratio data e.g. 1536**2) but it kinda works, so if you want to try something i'd suggest you start there.
@R34_Solitude Thank you so much for responding! Love your LoRA's either way!