MFLUX Qwen Image 2.1 setup and image comparison

This Qwen Image 2.1 tutorial explains how to run models locally on a Mac or PC, then compares identical prompts across several image generators. The speaker tests on an M3 Max MacBook Pro with 64 GB of memory; the PC guidance comes from published tests rather than hands-on results.

The Mac setup uses MFLUX and requires Apple silicon, at least 32 GB of memory and about 33 GB of disk space. The speaker reports roughly 30 GB of peak memory for the 8-bit version. Images at 1536 by 864 with 40 steps took 8 to 11 minutes. In a separate comparison, the full model took about 7.5 minutes but needed around 46 GB of memory. For Windows, the tutorial recommends ComfyUI and an NVIDIA card with 24 GB of VRAM; smaller cards need a compressed model.

The tests cover posters, labels, portraits, interiors and transparent stickers. Qwen produces convincing skin and genuine transparent cutouts in these examples, but struggles with scene details and preserving a reference face. Extra text tests show better spelling when the prompt supplies every word. Text the model must invent often becomes gibberish. The speaker ranks GPT Image 2.5 first overall and Qwen last in the paid-model comparison, while finding Qwen useful for repeated concept drafts.

The speaker also states that Qwen's license covers research and testing, with a separate commercial license required for paid work. That limitation matters when considering the demonstrated outputs for production.