nukomasshigura/Neta-Cat-Tower

Neta Cat Tower

Introduction

Neta Cat Tower is a text-to-image model fine-tuned from NetaYume Lumina.
This model was trained with the goal of enhancing anime style.
No learning was conducted regarding the addition of characters.

Model Description

• Developed by: nuko masshigura
• Model type: Text-to-Image generative model based on Neta Lumina
• License: Apache License 2.0
• Finetuned from model: NetaYume Lumina

Model Components

• Diffusion Transformers: This model
• Text Encoder: Pre-trained Gemma-2-2b
• AutoEncoder: Pre-trained Flux.1 dev's AE

"all_in_one" is a single model that are combined with DiT, text encoder and autoencoder.

How to Get Started with the Model

Please refer to the Neta Lumina's model card.

Recommended settings

• Sampler: res_multistep/ euler_ancestral
• Scheduler: linear_quadratic
• Steps: >=30
• CFG (guidance): 4 – 5.5
• Resolution: 1024 × 1024, 768 × 1532, 968 × 1322, or >= 1024

Prompt

Please refer to the Neta Lumina Prompt Book
About character knowledge, please refer to the NetaYume Lumina's Civitai page

Training Information

v1

• base model: NetaYume Lumina v3.5 (pre-trained)
• dataset: 2.1k anime style dataset with danbooru tags and English captions
• hardware: Geforce RTX5090 x 1
• training tool: sd-scripts
• mixed_precision: bf16
• save_precision: fp16
• resolution: '1280,1280'
• optimizer_type: AdamW8bit
• learning_rate: 3e-5
• lr_scheduler: warmup_stable_decay
• train_epochs: 20
• train_batch_size: 1
• gradient_accumulation_steps: 4
• min_snr_gamma: 5
• ip_noise_gamma: 0.1
• timestep_sampling: nextdit_shift

Acknowledgments

• duongve: Thanks to duongve for sharing awesome model.