Qwen-Image 2.1

Highlights

• One model does both text-to-image and instruction editing, saving a set of weights
• Chinese prompts and Chinese text rendering are among the best of its generation
• Full version is 25 steps, with an 8-step Turbo version (compared on the same page)
• Official templates for text-to-image, image editing, and matting

Scope

• Good for: Chinese posters, images with text, e-commerce images
• Good for: editing tasks that need "modify the image" rather than "redraw it"
• Not for: pure anime styles (Anima/Illustrious is a better fit)
• Note: research license; confirm the terms before commercial use

Tips

• When writing prompts, put text that should appear in the image in quotation marks; Chinese rendering is more reliable that way
• In editing tasks, image_1 is the image being edited and the other reference images provide material; don't connect them backwards
• On 16GB, set batch size to 1 to avoid running out of VRAM

Cautions

• Qwen Research License is not the same as permission for commercial use; this is the easiest pitfall

Parameters

Steps

25 (full version)

CFG

Per template default (the official editing example uses true_cfg_scale=4.0)

Resolution

Choose from resolution presets; official examples go up to 1696×2528

Text encoder

Qwen3-VL-8B (the int8 version saves VRAM and is loaded by default in the template)

VAE

Qwen-Image family VAE

Deployment

16GB verdict

Runs, but accept the trade-off: int8 weights and an int8 text encoder, stable in the 1024–1536 range; 2K requires offload.

VRAM notes

The 8B text encoder is the main VRAM cost; the template defaults to the int8 encoder precisely for cards like 16GB.

Suggested starting point

Go straight to the Turbo version (8 steps) for everyday generation and keep the full version for key images that need fine-tuning.

Comparison

Among its generation, it has the most stable Chinese typesetting; 171k downloads in the last 30 days also show it has taken hold in domestic workflows. Compared with MOSS, it doesn't need drawn boxes and typesets directly from prompts; compared with Krea 2, it adds an "editing" capability but needs more steps (25 vs Turbo 8). Note the license is Qwen Research; commercial use must be negotiated separately.