Kezmark/Mordant-1.2B-Think
Kezmark/Mordant-1.2B-Think is a 1.2 billion parameter language model, fine-tuned from liquid-ai/LFM2.5-1.2B-Thinking, specializing in AI image generation composition. It excels at generating detailed, spatially structured image prompts and enhancing existing compositions with chain-of-thought reasoning. This model is optimized for speed and lightweight deployment, making it ideal for prompt enhancement and art style infusion tasks in pipelines like ComfyUI.
Loading preview...
Mordant-1.2B-Think: AI Image Composition Specialist
Mordant-1.2B-Think is the most lightweight and fastest model in the Mordant family, a 1.2 billion parameter model fine-tuned from liquid-ai/LFM2.5-1.2B-Thinking. It is specifically designed for AI image generation composition, producing well-structured compositions with solid spatial reasoning despite its compact size. The model was trained on a curated dataset of 7,203 detailed compositions, incorporating reverse-engineered chain-of-thought reasoning with a "sharp, caustic, professional mentor" persona during the reasoning phase.
Key Capabilities
- Detailed Prompt Generation: Creates spatially structured image prompts from short descriptions.
- Prompt Enhancement: Adds compositional depth to existing prompts.
- Art Style Infusion: Integrates specific art styles into existing compositions.
- Chain-of-Thought Reasoning: Utilizes an internal
<think>block for spatial breakdown and critique, which is then used to generate the final composition. - Lightweight & Fast: Optimized for speed and low resource usage, making it suitable for iterative generation in local pipelines like ComfyUI.
Good For
- ComfyUI Integration: Includes a custom
MordantPromptEnhancernode for seamless use, featuring auto-tuning for GPU VRAM and handling of tokenization and ChatML formatting. - Specific Art Styles: Trained on a diverse list of 40+ art styles, from "16-bit Pixel Art" to "Watercolor," ensuring consistent style infusion.
- Resource-Constrained Environments: Available in quantized GGUF formats (e.g., Q8_0 at ~1.3 GB) for efficient local inference.