Njsemi is a style LoRA for Qwen-Image 2.1. It turns the base model’s flat, washed-out renders into cinematic semi-realistic fantasy: rich detail, dramatic backlight, glowing particles and closer, livelier framing. It carries well beyond its training subjects: the market, the dragon, the neon street and the airship in the gallery are not in the dataset. Every comparison uses the same seed with and without the LoRA.
Trigger word: njsemi — put it first, e.g. "njsemi. A lone knight resting under a giant tree at dusk, fireflies around him."
Strength: 1.0 (0.7–0.9 for a lighter touch).
Recommended: 30 steps, CFG 3, euler / flow-match, landscape sizes such as 1344×768.
ComfyUI: Qwen-Image 2.1 + LoraLoaderModelOnly.
Training: ai-toolkit, rank 32, 3500 steps, 505 images at 1024 px, on the Qwen-Image 2.1 int8 weights from Comfy-Org.
Base model licence: Qwen-Image 2.1 is released under the qwen-research licence (research and evaluation); commercial use needs an agreement with Alibaba.
Description
First release. Trained with ai-toolkit on Qwen-Image 2.1: rank 32, 3500 steps, 505 images at 1024 px.
Start the prompt with "njsemi." (trigger word). Strength 1.0 is the intended look; 0.6–1.2 works.
Known limit: the style pushes figures toward fuller bodies, and words like "small breasts" are mostly ignored even at 0.7 strength.