Try-On, Klein 4B, No LoRA (Odd Poses, Impressive)

A new demonstration of the Klein 4B model is going viral for its exceptional capability in virtual clothing try-on. The distilled AI model, shared by user ZerOne82, successfully generates photorealistic images of a person wearing new garments by processing just three inputs: a pose reference, an image of a top, and an image of pants. Remarkably, this is achieved using a simple, standard ComfyUI workflow without any Low-Rank Adaptation (LoRA) fine-tuning, which is typically required for such specific tasks. The results, showcased in an animation, are impressive for their consistency and handling of challenging, non-standard poses sourced from another model called Z-Image-Turbo.

The technical setup is notably efficient. The workflow uses the Klein 4B model in a *.sft, fp8 format, paired with a Qwen3 4B CLIP vision model for understanding image content. It generates 800x1024 resolution images using an Euler sampler with very low guidance, requiring only simple text prompts like 'put top on.' Each generation completes in under 15 seconds. This speed and reliability, coupled with the elimination of the need for task-specific LoRAs, lower the barrier to entry for creating high-quality fashion visualization tools. It points to a future where AI can rapidly prototype clothing designs on diverse body types and poses, directly impacting e-commerce, digital fashion, and content creation.