MicroCosmos & Beyond: LLM-Driven Proportion Engine (Krea2 + Ollama)

MicroCosmos & Beyond: LLM-Driven Proportion Engine (Krea2 + Ollama)

Welcome to MicroCosmos & Beyond, a highly calibrated, automated ComfyUI workflow designed to master extreme scaling, flawless proportions, and cinematic macro/fantasy environments.

Standard image models often break anatomy or warp scale when dealing with miniature subjects (like a 10cm pixie) or complex hybrid creatures. This workflow solves that issue permanently by using Ollama (LLM) as a real-time prompt translator that enforces physical scale mechanics, combined with the raw speed and aesthetic power of Krea2.

🚀 Key Features & Included Presets:

This workflow is a multi-functional prompt engine. Check the dropdown menu inside the PromptManager (Node 14) to switch between these custom-built engines:

1. The MicroCosmos Engine: Perfect for flawless 5-10cm miniature worlds, pixies, and small characters interacting with giant everyday human objects. Completely optimized using positive-only framing to prevent duplicate background faces or texturing errors.

2. The Hybrid Anatomy Engine (Default Load): Built for epic, heavy dark-fantasy spectacles, massive hybrid beasts, and detailed biomechanical cyborg transformations in the legendary style of Frank Frazetta and Greg Rutkowski.

⚠️ IMPORTANT: Quality & Upscaling Hint

All preview images shown here are 100% RAW generations straight out of the first KSampler pass to showcase the sheer power of the model's texturing capability.

• Pro Tip: To inject micro-details, razor-sharp textures, and perfectly polished skin pores/eyes into your final art, add a FaceDetailer node (YOLOv5l) and a 2x Upscale pass at the end of the image output chain. The raw base generation provides the perfect mathematical foundation for stunning upscales!

📦 Prerequisites & Setup:

To make this workflow run out of the box, ensure you have the following assets ready:

1. Required Custom Nodes (Install via ComfyUI Manager):

• comfyui-simple-ollama (For the Ollama LLM node integration)

• prompt-manager (To hold the System Prompt presets)

• comfyui-workflow-encrypt (For the LoraManager and custom ImageSavers)

• pysssss-comfyui (For the automated ShowText title-saver display)

2. Models & Settings:

• Base UNET: Krea2\krea2Turbo18For_v2.safetensors (Runs at an ultra-fast 4 steps!)

• Text Encoder: qwen3vl_4b_fp8_scaled.safetensors

• VAE: krea2RealVae_v10.safetensors

• Latent Resolution: Configured to a cinematic, ultra-wide 1728 x 720 aspect ratio.

• LLM Model (Ollama): gemma-4-abliterated (or any modern 8B model like Llama 3 / Mistral)

• Active LoRAs utilized in this stack:

  • krea2_turbo_4step_rank_64_lora (Strength: 1.00)

  • zy_BackgroundDetail_K2 (Strength: 0.50)

  • Nijireol Krea2 v1 Ep5 (Strength: 1.00)

  • BeMyHero_-_LeaX_epoch_9 (Strength: 1.00)

⚙️ How to Operate the Auto-Titling Engine:

The secondary Ollama node at the end of the chain automatically analyzes the positive prompt output and generates a clean, custom 1-2 word filename (e.g., "Crimson Cyborg") for your SaveImage directory, keeping your outputs highly organized without manual typing.

Load the .json file, choose your preset in Node 14, type a few short words, and let the LLM build your next masterpiece! Leave a review and share your generations in the gallery below.

Feel free to leave a review and share your tiny universe creations in the gallery below!