Kraft | LoRA — Sousou no Frieren | Base: ANIMA

🙏 Before anything else: this LoRA took a lot of work — picking the images, building the dataset and tags, training, and testing the checkpoints. If you have a moment, drop a Like and especially post your results in the comments (with the prompt/settings you used) — it helps me a lot to know what to improve in the next versions. 🔥

✅ This character is , so content is welcome. Use it respectfully.
💬 Check out how the prompt examples were set up in the posted images.
💬 Posting your generations in the comments is the best way to support small creators — it helps me understand what's working and what to improve in the next versions. Tag the LoRA so I can find them!
📦 This LoRA was a test run — the dataset is noticeably smaller than my other releases, with ~80% of images taken directly from the anime. Despite that, it delivers solid results and a good anatomical base.

📊 About this model

On the vanilla ANIMA base, Kraft tends to lose most of his distinctive look — the cross detail on his outfit in particular comes out inconsistent or absent. This LoRA helps a lot with character accuracy and produces good overall, but it does struggle with the cross/collar area — make sure to reinforce that in your prompt if you want it to appear correctly.

ℹ️ Generation tips

• On ANIMA base with 12 steps + acceleration LoRA, the output keeps a clean anime feel that works well for Kraft — recommended as a starting point

• Using more steps or a different ANIMA-based model with quality LoRAs will improve sharpness and detail

• Higher CFG can help lock in his facial features and outfit details more consistently

• Cross/collar detail: this is the model's weakest point — reinforce it heavily in the prompt for best results

• Preview images were generated at 816x816 — this is the recommended resolution; 1024x and above tend to cause the character to lose coherence, so use higher resolutions with caution

• Check the prompt examples used in the posted preview images — they show how to structure tags for each look

📊 Checkpoints

StepsResult

⭐ 350 Best for / — recommended starting point for and casual outfits

⭐ 600 Best for his classic outfit — recommended when going for the full look

Want a different checkpoint? Just ask in the comments — I have steps available up to 600 and can upload them on request.

⚙️ Generation settings used for preview images

All preview images were generated using:

• Base model: Anima 1.0

• Acceleration LoRA: strength 1.0

• CLIP: anima default

• Steps: 12

• CFG: 1.0

• Sampler: er_sde - beta

• Resolution: 816x816

👍 One more thing

This one started as a test, but I'm happy with how it turned out — especially given the smaller dataset. If you have a moment, drop a Like and share your generations with the settings you used. Comments with results are the best way to help small creators keep going. Have fun! 🔥