[Hews Style]

["Rough-lined anime aesthetics with a signature touch for atmospheric lighting."]

Quick Info

Type:  LoRA
Model: Anima
Trigger Words: @hews
Recommended Weight: 0.6 ~ 1.0
Target: portraits and/or illustration of anime style characters.

Specific Model Info can be found in the related section of this description AND in the model versioning.

Introduction

This model aims to produce artwork in the style of @hews.

It's designed as a dual-purpose tool:

• 1) one that captures the finished feel of a complete artwork

• 2) the other focus on understanding the artist's sketching process.

This is meant as a didactic tool, so please use it accordingly.

This model is for art enthusiasts and students who want the ability to generate their own reference images while studying the art style of @hews.

Why give this a shot?

While others race to release models quickly, we're taking a slower, more deliberate approach to ensure the best possible quality. Where others might promise you speed, we promise you flexibility and reliability.

Roadmap:

Version - Name - Status - Description:

• 0.1 — Alpha [released]

  • First baseline proof of concept, using only a fraction of the available data

• v0.5 — Slim [released]

  • ~5x increase in training data, with refined captioning

• v0.8 — Fat [planned]

  • Full available dataset used for training

• v1.0 — Definitive [planned]

  • Final caption refinement once the dataset is fully consolidated

Versioning:

The next 4 sections are specific for the version previously mentioned.

v0.1 — Alpha
Status: ✅ Released

As previously mentioned, this was a fast release intended to serve as a proof of concept.

Details:

Activation tag: @Hews
Natural language trigger: In the style of @hews

Additional keywords:

For Sketches
- Tags: sketches, greyscale
- Natural language: A greyscale sketch

Recommended Settings:

• Sampler: Euler a, ER SDE

• Steps: 20–50

• CFG Scale: 4~7

• Clip Skip: 1

• Lora Weight: 0.8 ~ 1.0

Known Issues:

• This version is good enough for a proof of concept, but has a few rough edges — while it understands the style decently well, it has a heavy effect on image composition, often altering the character's pose or other minor details.

Geek Info:

• This version was trained on a dataset of ~100 images, using dual captioning (booru-style tags + natural language prompt).

• Training was done using the Civitai trainer at a max resolution of 1024² (this isn't the best case scenario for anima like models, but it's better than our 512² alternative).

• This version was trained at a time when the Civitai trainer supported Anima models experimentally, so captioning wasn't necessarily handled as it should have been and there may be some minor issues that aren't immediately obvious in the output.

• Captioning was generated using the WaifuDiffusion (WD 1.4) tagger for the tag-style labels, and Qwen3-VL-8B for the natural language captions, run locally via LM Studio and automated with the neme-anima script.
  

v0.5 — Slim
Status: ✅ Released

When the proof of concept starts becoming a tool worth using.

Details:

Activation tag: @Hews
Natural language trigger: In the style of @hews || In the style of "@hews"

Note: This version was trained with a dataset containing both In the style of @hews and In the style of "@hews" as natural language captions, so it recognizes both — with slightly better performance using the latter.

Additional keywords:

For Sketches
- Tags: sketches, greyscale
- Natural language: A greyscale sketch

As you can see, we tried to maintain most of the backward compatibility, with a few caveats. The sketch keyword now has an even stronger effect at boosting the overall style, even in "non greyscale" scenarios.

Recommended Settings:

• Sampler: Euler a, ER SDE

• Steps: 20–50

• CFG Scale: 4~7

• Clip Skip: 1

• Lora Weight: 0.8 ~ 1.0

Known Issues:

• This version has a slight bias toward white shirt clothing if none is specified in the prompt.

Geek Info:

• This version was trained on a dataset of ~512 images, using dual captioning (booru-style tags + natural language prompt).

• Training was done using the neme-anima trainer at a max resolution of 512² (512×512), this is the most we can currently manage training locally; we've tested extensively, and this was the best outcome we could produce with the dataset in its current state.

• Captioning was generated using the WaifuDiffusion (WD 1.4) tagger for the tag-style labels, and Qwen3-VL-8B for the natural language captions, run locally via LM Studio.

• We opted for ~2,000 steps, with a warmup covering the whole first batch. That, combined with a learning rate of 1e-5 (1 × 10⁻⁵) and some other minor tweaks, allowed us to achieve the target style without affecting the overall composition of the image.

🔹 v0.8 (Fat) — [planned]

🔹 v1.0 (Definitive) — [planned]

⚙️Usage Tips

• Prompting style: This model doesn't break any Anima conventions, so using it with dual captioning still provides the best results.

• Negative prompts: Standard Anima negatives should work fine.

• Good pairings: This LoRA is intended for base models that don't have a strong, hard-baked style, results may not be as good on checkpoints that do.

• Things to avoid: This model was trained on an NSFW artist's work, so keep the "safe" tag included and be as detailed as possible in your prompt to avoid unwanted output.

❓ FAQ

Q: Can I use this model commercially?
A: Yes, as long as doing so doesn't negatively affect the original artist this model is based on AND as long you respect the other licensing constraints. (no re-prints, credit required)

Q: Why do my results look different from the showcase images?
A: There are a few possible reasons, but the most likely one is the UI you're generating with. Generation differs slightly between CivitAI's built-in generator, A1111, A1111 Forge Neo, and ComfyUI. For example, A1111-style UIs generate the initial noise on the GPU, while ComfyUI generates it on the CPU — even with the same seed, that difference alone can produce noticeably different results.

Q: Will there be an [SDXL / Flux / Pony / Illustrious] version?
A: No version for those base models is planned before the 1.0 release. After that, we may consider dropping the dual captioning approach in order to train those versions.

🙏 Conclusion

This is a really big project, and a long-term one. Cleaning this amount of data is pretty overwhelming at times, but the results have been satisfying — with each version, I can see real improvements in the model, and I get to refine my own training methods along the way.

If everything goes as planned, by v1.0 this should be one of the most complete references available for studying @hews' art style — and that's genuinely exciting to work toward.

This was a really long description, so if you made it this far, I want to personally thank you for the time you spent reading it. I hope this model helps you learn, study, and create. Feel free to drop your thoughts, questions, or generations below — I read everything as soon as I can, and if I haven't gotten to your comment yet, just know that I'll catch up.

💖 Support This Model

If you enjoy this model and want to help it grow:

• ⭐ Leave a review — it genuinely helps visibility on CivitAI.

• 📸 Share your generations in the gallery — I love seeing what people create, and it also helps us collect enough Buzz to train on-site without frying our GPUs.

• 💖 Follow this profile — so you won't miss when a new release drops, and so we can climb the leaderboard and reach more people.

• 🖌️ Just use the model — simply putting it to work is support in itself. Knowing people are using it keeps us motivated to keep doing this.

• ⚡ Buzz donations — if you've got a dozen Buzz to spare, we'll gladly take them. Our hardware isn't powerful enough to run video models reliably, so generating animated cover images for the character models costs us a bit (xd) — even a couple of Buzz can make a difference now and then.

Thank you for checking out Hews Style — happy generating! 🎉

Last updated: [15/07/2026]