Anima-turbo-v1.0-GGUF

This is a quantized GGUF repository for the Anima-Turbo-v1.0 model originally created by CircleStone Labs. The model weights have been converted into F32 and subsequently quantized into multiple GGUF formats using a patched version of llama.cpp to enable efficient CPU and GPU-offloaded image generation in ComfyUI.

🌟 Model Overview

• Original Model: circlestone-labs/Anima
• Architecture: 2 Billion parameter text-to-image model. It is built on the nvidia/Cosmos-Predict2-2B-Text2Image base.
• Variant: Turbo v1.0 (A distilled version optimized for fast, low-step generation).
• Focus: It is focused mainly on anime concepts, characters, and styles. The model is designed for making illustrations and artistic images, and will not work well at realism.

💻 Hardware Compatibility & File Sizes

Select the quantization that best matches your system's RAM/VRAM constraints. Q4_K_M or Q5_K_M are generally recommended as the best balance of speed and visual fidelity.

Bit Depth	Quantization Type	File Size
3-bit	Q3_K_M	1.02 GB
4-bit	Q4_K_S	1.27 GB
4-bit	Q4_K_M	1.38 GB
5-bit	Q5_K_M	1.56 GB
6-bit	Q6_K	1.74 GB
8-bit	Q8_0	2.24 GB

⚙️ How to Use in ComfyUI

1. Requirements

• Update ComfyUI to v0.14.1 or higher.
• Install the ComfyUI-GGUF custom node to enable the UnetLoaderGGUF node.

2. File Placement

Ensure you download the supplementary files located in this repository's split_files directories and place them into your local ComfyUI folders:

• Diffusion Model (GGUF): Place your chosen anima-turbo-v1.0-[Quant].gguf file into ComfyUI/models/diffusion_models (or the unet folder).
• Text Encoder: Place qwen_3_06b_base.safetensors into ComfyUI/models/text_encoders.
• VAE: Place qwen_image_vae.safetensors into ComfyUI/models/vae.

3. Turbo Generation Settings

Because this is the distilled Turbo version of Anima, it relies on specific, highly restricted parameters to function correctly:

• Steps: 8 to 12
• CFG Scale: 1.0 (Do not use high CFG with the Turbo variant).
• Sampler: The euler sampler is highly recommended with the Turbo version, as it is naturally more stable.
• Resolution: Works at resolutions between 512x512 and 1536x1536 pixels.

4. Prompting Guide

The model is trained on Danbooru-style tags, natural language captions, and combinations of tags and captions.

Syntax & Formatting:

• Use lowercase for tags, and spaces instead of underscores.
• Score tags are the only tags that use underscores.

Prefixes:

• Recommended Positive Prefix: masterpiece, best quality, score_7, safe,
• Recommended Negative Prompt: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration

Natural Language Tips:

• If using pure natural language, more descriptive is better.
• Aim for at least 2 sentences.
• Extremely short prompts can give unexpected results.

🌳 Model Tree & Lineage

• Base Architecture: nvidia/Cosmos-Predict2-2B-Text2Image
• Fine-tuned / Distilled Version: circlestone-labs/Anima (Turbo v1.0 variant)
• Quantized Version: Abiray/Anima-turbo-v1.0-GGUF (This repository)