Anima Base Vivid V1 - Realism Fine-Tune (FFT)

Hello everyone!

I decided to try training the Anima Base V1 model on my custom dataset to push it towards realism. This is a full FFT (Full Fine-Tuning) trained on a dataset of 1,500 image-caption pairs.

Key Features

• Dual-Prompting Support: The dataset captions were split into both natural English and Danbooru tags. As a result, the model understands both prompting styles perfectly.

• Character Memory Retention: Thanks to a gentle, 2-stage fine-tuning process using Kohya scripts, the model successfully retains memory of standard characters from its base training.

• High Dataset Variability: The training set features diverse camera angles, dynamic lighting, and a high variety of NSFW themes (including futanari, sex, nudity, yuri, anthropomorphic/furry, and high-quality aesthetic shots).

• Future Potential: As a first version, it already demonstrates that the model can be pushed significantly further into realism with an even larger dataset.

Recommended Generation Settings

My Personal Setup

• Sampler: deis_3m

• Scheduler: beta57

• CFG Scale: 4

• Steps: 30–40

Standard / Lighter Setup

• Sampler: euler_a

• Scheduler: normal

Note: While faster, standard settings may result in fewer details and more "plastic-looking" skin.

Prompting Guide

Positive Prompt

Start your prompt with:

> masterpiece, best quality, realistic, absurdres

* Tip: realistic is the key trigger word for this fine-tune. I highly recommend keeping it in your prompt at all times.

Prompt Style: Natural language isn't separated by commas and offers more control than tags, while tags produce more vivid results—though this isn't an absolute rule.

Negative Prompt

> (white background:1.2), simple background, worst quality, low quality, low details, bad anatomy, bad hands, deformed hands, missing fingers, fused fingers, sketch, censor, score_1, score_2, score_3, text, watermark, artist name, flat color, (anime_source:1.2)

Installation

* Format: This is a diffusion model (UNet), not a full standalone checkpoint.

* ComfyUI Path: Place the model file in:

ComfyUI\models\diffusion_models\

* Required Files: You will need to use the standard Anima VAE and Text Encoders alongside this model.

No images of real people were used during training.