IMPORTANT - Release: read about the update.
v1: (早坂愛 from Kaguya-sama: Love is War)
This is my first anime character LoRA. Trained on Flux.2-klein-9B base (Modelscope version) and intended for use with the distilled/KV variant.
Current status
This is a checkpoint from ~1000 steps (10% of the planned 10,000 steps).
Training is still in progress — updates will follow after completion (and possibly stronger versions).
Even at this early stage, it already delivers very recognizable likeness with the right settings.
Recommended settings
Strength: 1.3 – 1.8 → my personal favorite — 1.5 (best balance of likeness and flexibility)
Steps: 8 – 16 → my sweet spot — 12
Sampler: Euler / Euler simple
CFG: 1.0 – 1.2 (1.0 usually gives the cleanest and most natural results)
Resolution: 1024×1024 or similar aspect ratios
Trigger & base prompt 4n1meG1rlH4y4s4k4, anime girl with blonde hair and blue eyes
After that, describe clothing, hairstyle, emotion, scene in natural language — the model responds very well to narrative text without booru tags.
Dataset
74 images
Mostly anime screencaps from the two seasons of Kaguya-sama
A small amount of high-quality fanart and booru generations for better angle and expression variety
Notes
This is a non-commercial experimental LoRA.
Not monetized, not for sale — created purely for personal interest and to test Flux Klein's capabilities on anime characters.
If likeness feels a bit weak — try raising weight to 1.6–1.7 or wait for the final version after 8000–10000 steps.
I’d love to hear your feedback and see your generations! If you find good prompts or settings that work well — please share in the comments, I’ll update the description accordingly.
Thanks for trying it out!
it can be used in creating NSFW content with correct a lora.
Description
Hayasaka Ai – Flux.2-klein-9B Character LoRA
(早坂愛 from Kaguya-sama: Love is War)
This is my first fully completed anime character LoRA.
Trained on Flux.2-klein-9B base (Modelscope version) and intended for use with the distilled/KV variant.
Current status
Training is now complete.
Final model: 10,000 steps / 10 epochs
The LoRA has reached stable character consistency and significantly improved facial accuracy, expressions, and pose reliability compared to earlier checkpoints.
You can find all training epochs and comparison generations (same prompt and seed across versions) via the provided link — useful for understanding how the model evolved over time.
https://mega.nz/folder/ugp1UQZA#dtgaoj7BTs3Am3XtgIaw4Q
Recommended settings
Strength: 1.3 – 1.8
→ my personal favorite — 1.5 (best balance of likeness and flexibility)
Steps: 8 – 16
→ my sweet spot — 12
Sampler: Euler / Euler simple
CFG: 1.0 – 1.2
→ 1.0 usually gives the cleanest and most natural results
Trigger & base prompt
4n1meG1rlH4y4s4k4, anime girl with blonde hair and blue eyes
After that, describe clothing, hairstyle, emotion, and scene in natural language — the model responds very well to narrative prompts without booru tags.
Dataset
74 images
Mostly anime screencaps from two seasons of Kaguya-sama: Love is War
A small amount of high-quality fanart and generated images for better variation in angles and expressions
Notes
This is a non-commercial experimental LoRA.
Not monetized, not for sale — created for research and experimentation with Flux Klein.
Future plans
I plan to continue developing Anime Character LoRAs for Klein 9B,
including a series of characters from かぐや様は告らせたい.
For creators
If you plan to train similar LoRAs:
Try using natural language captions instead of booru-style tagging.
This helps preserve a much more intuitive and user-friendly prompting experience — especially with Flux-based models.
I’d love to hear your feedback and see your generations!
Lora works with the NSFW with an additional Lora.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.