README.md

---
title: Anima 2.9B
emoji: 🎨
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.23.1
app_file: app.py
pinned: false
suggested_hardware: zero-a10g
hf_oauth: true
short_description: Anima 2.9B anime text-to-image ZeroGPU demo
---

# Anima-2.9B Hugging Face Space Demo

This Space hosts an interactive demo for **[Gazingstars123/Anima-2.9B](https://huggingface.co/Gazingstars123/Anima-2.9B)**, an expanded 40-layer, 2.9B parameter anime and illustration text-to-image diffusion model.

## Model Highlights

- **Base Model**: Fine-tuned and layer-expanded from [circlestone-labs/Anima](https://huggingface.co/circlestone-labs/Anima) (built upon [nvidia/Cosmos-Predict2-2B-Text2Image](https://huggingface.co/nvidia/Cosmos-Predict2-2B-Text2Image)).
- **Layer Expansion**: Expanded from 28 transformer layers to 40 transformer layers (~2.9B parameters).
- **Dataset**: Trained on an additional 1.7M anime and illustration samples using Muon optimizer.
- **Knowledge Cutoff**: July 2026.
- **ZeroGPU Compatible**: Optimized with `@spaces.GPU` for rapid zero-cost inference on Hugging Face Spaces.

## Recommended Generation Parameters

| Parameter | Recommended Value | Notes |
| :--- | :--- | :--- |
| **Resolution** | `832x1216` (Portrait) or `1216x832` (Landscape) | Dimensions must be divisible by 16 |
| **Inference Steps** | `28 - 50` | 28 is fast and sharp; 50 provides high detail |
| **CFG Guidance** | `3.5 - 5.0` | Default `4.0` |
| **Negative Prompt** | `worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration` | Removes low-aesthetic artifacts |

## Prompting Tips

- **Quality tags**: `masterpiece, best quality, score_9, safe`
- **Character tags**: Danbooru/Gelbooru tags (e.g. `1girl, solo, silver hair, blue eyes`)
- **Artist styles**: **Always prefix with `@`** (e.g. `@artist_name`) for accurate style guidance.
- **Detailed prompts**: The more descriptive and detailed the prompt, the better the background and composition.