🧪 PINK ALCHEMY ANIMA: 3P EDITION🧪

THIS. IS. ALCHEMY! AND THE PARAMETERS ARE AT THE BOTTOM :3

YOU NEED THESE:

• https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae

• https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders

(UPDATE: @Sophorium CALLED ME OUT and I have to come clean! He identified an issue with the model and I was able to shrink it down. There is 100% no difference in composition, quality nothing in same-seed tests. I promise you, for better or worse, it is the exact same model. Please enjoy!

Welcome back to the lab. My goal from day one was to completely eradicate that flat-art style and build a semi-real anime mix without that muddy, vinyl, plastic-looking output.

The original Pink Alchemy proved it could be done. But for the 3p Edition, we didn't just iterate—we ripped out the engine and built an absolute monster on a completely new foundation. The details are crisp, the eyes pierce right through the screen, and the textures have that exact, razor-sharp edge.

⚙️ THE RAW METAL: ANIMA 3P ARCHITECTURE

Let's look at what is actually under the hood. We aren't riding the coattails of older architectures or heavily-fried SDXL merges anymore. Pink Alchemy Anima 3p is forged directly from Anima preview3-base.

• The Backbone: This isn't a lightweight. It is powered by the Cosmos Predict-2 2B (Diffusion Transformer). It has the structural "common sense" to inherently understand object permanence, depth, and lighting far better than previous generations.

• The Text Encoder: Anima ditches standard T5 or CLIP for a Qwen3-0.6B encoder. It handles complex, long-form natural language prompts effortlessly without losing the plot halfway through your sentence.

• The VAE: Paired with the specialized Qwen VAE, it pulls out crisp facial details and textures that older models simply crush.

• The Tax: This architecture is incredibly dense. It was trained hard at a native 1024x1024 resolution.

🗣️ HYBRID PROMPTING?

Forget everything you know about just spamming a wall of comma-separated tags. Because of those Qwen text encoders.

To get the absolute face-melting outputs this model is capable of, you need Hybrid Prompting:

1. Natural Language First: Start by physically describing the scene, the lighting, and the action using actual, multi-sentence English prose.

2. Booru Tags Second: Lock in your Danbooru tags alongside the natural language to hard-lock the concepts if you want

3. Syntax Rules: Keep your booru tags lowercase. Do NOT use underscores between words unless it is a specific score tag (like score_9).

🛠️ TAGS & SKELETONS

Quality Tags (The Baseline): > masterpiece, best quality, score_9, score_8, score_7, newest,

Negative Tags:

worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, lowres, censor, (bad quality:1.15), (worst quality:1.3) Please use the score tags and standard negatives together for best results to keep the mud out of your generations.

🎛️ OPTIMAL GENERATION PARAMETERS

Because we have shifted to the Anima architecture, the old 4.8-5.8 CFG / Euler Beta setup from the Illustrious days needs to be completely rewired. Here is exactly how you drive it:

• CFG Scale: 4.0 - 5.5 (Anima is highly responsive. Pushing past 6 without significantly adjusting steps can start burning the image).

• Steps: 35 - 50 (Anima requires more brute force to pull out the fine details, especially hands, compared to older architectures. 35 is your baseline; 45+ is where it shines).

• Sampler: er_sde (Highly recommended for neutral style, flat colors, and sharp lines) OR Euler A (If you want softer, thinner lines and a slightly more colorful, hazy look). dpmpp_2m_sde is also a great option for more creative outputs.

• Scheduler: Simple, Normal, or beta(57) (CRITICAL: Avoid Karras schedulers with this base unless you are prepared to push 100+ steps to resolve the image).

• Resolution: 1024 x 1024 (Native sweet spot) or equivalent 1 Megapixel aspect ratios (e.g., 896 x 1152, 1152 x 896).