Training datasets are generated via Nano Banana Pro/2.
In one word: This is an anime2real Flux.2 Klein model specialized in converting anime-style characters into realistic human appearances.
Limitation & Balance: While the model excels at "Anime to Real" conversion, maintaining 100% character consistency can be challenging. Anime proportions are naturally unrealistic; thus, the model prioritizes anatomical realism over strict feature duplication to deliver better photographic results.
Trigger word: A2R
Strength: 0.9 (better finger & toes)
Example:
A2R, Raw smartphone photography, candid aesthetic, significant image noise, heavy grain. No UI elements. Reduce head size.
keep all the elements and same perspective and camera angle, pose and characters positions, remake exactly this frame but into Amateur smartphone snapshot taken with iPhone 16 Pro front camera, raw unedited photo, sharp photo, detailed skin texture, flawless skin, fair skin, white pale skin, skindentation.TODOs:
Fix upside-down / reversed face generation.
Enhance low-light & night view rendering.
Optimize finger / toes anatomy.
Changelog:
v1 (38 images 2500 steps)
Initial Experimental Training: Baseline model trained using 38 high-quality, curated sets of image pairs generated via Nano Banana Pro / 2.
Core Alignment: Focused on establishing the foundational mapping from anime stylistic features to photographic human skin textures and facial anatomy.
Proportion Correction: Introduced basic logic to downscale non-human head sizes and correct dramatic anime expressions into natural, realistic equivalents.
v2 (80 images 8000 steps)
Dataset Expansion: Extended the training library with an additional 42 complex image sets, focusing on higher diversity and varied lighting conditions.
Improved Aspect Ratio Support: Upgraded the bucket training to natively support multiple image ratios (portrait, landscape, and square) without stretching or blurring.
Enhanced Environmental Physics: Integrated specific night views and challenging dark-tone scenarios to improve shadow transitions and low-light skin reflections.
Description
Add another 42 image sets to the dataset.
FAQ
Comments (3)
hi. i think that v1 might be better over all. v2 sticks more to the original image composition, but seems to degrade the sharpness and quality. however. Thank you for sharing. it's great nevertheless
Yes, because the images generated by the Klein have a high effective pixel count, they look very sharp. My dataset currently consists of raw, unoptimized outputs directly from Nano Banana, which is why the image quality is degraded.
However, you can improve these issues with a few extra steps. Right now, I use two [workflows](https://civitai.red/models/2655376/klein-9b-anime2real-workflow?modelVersionId=2981669) to generate the images and upscale them (which resolves the quality drop you mentioned).
I might also be able to fix this by optimizing my dataset, and I’ll definitely focus on that as an improvement for the next release.
Thx for the honest answer. Indeed I already use a workflow that makes use of upscaling, but it loses the overall look your lora manages to create. Looking forward to your further effort. Tnkas again for that !