2.5D, 3D, Semi-Realistic, Realistic. Danbooru tags are still compatible, with a default bias towards the Semi-Realistic style. The core facial features align with Asian aesthetic preferences—you can modify styles using Danbooru tags or artist embeddings, and adjust the realism level via prompts.
Parameters:
Steps: 25-40，CFG Scale: 4-7，Sampler: dpmpp_sde_gpu，Scheduler: Beta，VAE: Pre-integrated
While other samplers are supported, dpmpp_sde_gpu is the only one recommended for optimal results. Feel free to use upscalers, ADetailers, or train custom LoRAs.
Resolution Requirement:Original image dimensions must exceed 832x1216.
Model Details
The base checkpoint used is my stable model: illust-Vector-2.5D_IL. The updated training dataset primarily focuses on female subjects, with fewer male samples. For male character generation, we recommend experimenting with additional LoRAs.
This model is specialized in female facial features, body proportions, and detailed rendering—tattoos, skin textures, and color grading have all been optimized to an ideal state. Distant facial details are continuously improving, but using an upscaler is still advised.
This is a TensorHub-exclusive model, as its NSFW performance exceeds my expectations.
If you encounter high token consumption, I can provide a dedicated workflow, small tool, or a merged DMD2 fast model.
Feedback and generated artworks are highly appreciated—hope you all have fun creating!
This model was not trained on any celebrities or real individuals. Please abide by the platform’s rules and use it responsibly.