Built with Qwen. Sabikaze for Qwen-Image 2.1: a quiet anime illustration style with a desaturated grey-blue / grey-teal palette, one small warm accent, heavy ink lines, flat value blocks and dry-brush texture on clothes and backgrounds. It is the Qwen-Image 2.1 version of the Anima, MOSS Image and Krea 2 LoRAs "Sabikaze", trained on the same 57 self-generated images.

How to prompt: start with "@z3-sabikaze." and name the medium in a sentence, for example "@z3-sabikaze. Anime-style digital illustration. An adult woman in a long black coat walks along a snowy canal at night under street lamps. Full body." Describe the scene in plain sentences. A prompt that does not name how the picture is drawn can stay photographic.

Settings matter: use 24 steps at CFG 4. At 16 steps and CFG 1 the style shows but the picture breaks into patches.

Strength: tested at 1.0.

Tested with ComfyUI 0.42.1, Qwen-Image 2.1 (int8 ConvRot DiT), the Qwen3-VL 8B text encoder, 24 steps, CFG 4, euler / simple, 832x1216 and 1216x832, loaded with LoraLoaderModelOnly.

Training: Qwen-Image 2.1 with ai-toolkit, 57 images, rank 32, 2000 steps, captions with the trigger first and the medium named.

Known limits: a small prop can appear twice, and hair colour can drift when the prompt does not name it; try another seed.

License: this LoRA is a derivative of Qwen-Image 2.1, and its use is subject to the Qwen Research License. Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved. Agreement text: https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE . Qwen has stated that outputs are not part of the licensed Materials and that users keep the rights to what they generate: https://x.com/QwenDevs/status/2101917379785838660