A meticulous fine-tune and merge built on the Anima architecture.
🌟 Overview
Moving beyond traditional diffusion limitations, Kitten Slop v1 leverages the cutting-edge Anima architecture—a specialized 2-billion parameter derivative of Nvidia's Cosmos. By combining Cosmos's advanced transformer backend with the Qwen text encoder, this model delivers unparalleled adherence to anime concepts, distinct stylings, and complex character traits.
This model is explicitly designed for high-quality illustrations, 2D art, and dynamic compositions. It is loaded with a massive injection of personally trained, highly specialized data, ensuring efficient prompt adherence, style recognition, and resolution control you won't find in standard off-the-shelf merges.
✨ Comparison Grid ✨
🛠️ Under the Hood
Because Anima operates on a completely different framework than older architectures, Kitten Slop v1 takes full advantage of its modern components to capture distinct linework, shading, and expressive details.
- Core Architecture: Nvidia Cosmos (Anima base)
- Text Encoder: Qwen 3 (600m)
- VAE Backend: Qwen Image VAE
- Design Philosophy: Engineered for reliable prompt adherence, whether you are using natural language sentences or traditional comma-separated tags.
📦 Downloads & File Structure
Keeping the deployment as simple and portable as possible, the model is available in two distinct formats to fit your specific workflow and hardware overhead.
1. The AIO (All-in-One)
The simplest deployment. Everything you need is baked into a single file for instant, plug-and-play generation.
2. The Split Files (Modular Setup)
For technical enthusiasts who prefer granular control over their node-based workflows, easier VRAM management, or swapping components.
- 💾 Download Core Transformer (Diffusion Model)
- 🔤 Download Text Encoder (CLIP)
- 🎨 Download VAE (The Color & Detail Decoder)
🎨 Sample Generations 🎨
⚙️ Recommended Settings
Anima-based models require completely different parameters than traditional setups. To get the absolute best results out of Kitten Slop v1's custom training, use these baseline configurations:
| Parameter | Recommendation |
|---|---|
| Quality Prompts | Not needed. If you must, then I recommend masterpiece, best quality, score_9, absurdres |
| Long Neg | worst quality, low quality, early, old, score_1, score_2, score_3, cartoon, graphic, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, artist name, blurry, jpeg artifacts, lowres, censor |
| Short Neg | worst quality, low quality, score_1 |
| Resolution | 1536px |
| Sampler | Euler A (for softer lines) or ER_SDE_Solver (for neutral styles, flat colors, and sharp lines) |
| Scheduler | Simple |
| Steps | 30 - 40 |
| CFG Scale | 4.0 - 6.0 |
| Sigma Shift | 3.3 (Unlike the usual 3.0 standard for this architecture, a 3.3 shift works best for this specific tune) |
| Hires. Fix | Ultrasharp or Ultrasharp v2 at 0.2 Denoising strength |
| Clip Skip | Not used / Not applicable for this architecture |
💡 Prompting Tip: You don't need excessive prompt-fu here. Mix plain English descriptive sentences with your favorite quality tags, and let the model handle the heavy lifting.