Qwen-Image-2.1 vs Flux3-klein: E-commerce Image Editing

These are two of the most talked-about open-source image models right now. I ran one round on the same five cases, with the same seed 411855900 and a canvas of 832×1248. The only things that changed between the two runs were the model and the step count.

Model	cfg	steps
Qwen-Image-2.1-Turbo-Nunchaku	1	8
flux-3-image-klein-nunchaku	none	16

Result: out of five cases, Qwen won two, Flux won one, and two were ties. Neither model completed the text translation.

Note: both runs used the community Nunchaku FP4 quantized weights.

1. Dressing in multiple items

Prompt: "Dress the model in Figure 1 in the clothing from Figures 1 and 2, and have her wear the shoes from Figure 3."

Winner: Qwen.

The references, in order, are the person, a T-shirt, red shorts, and sneakers. The figure numbers in the prompt do not match that order, and both models read the request as putting all three garments on her. Both swapped in the navy #LOVE T-shirt, the red shorts, and the green-and-white sneakers, and both kept the person and her pose. Qwen rendered the chest lettering and the green, yellow, and white blocking on the sneakers more clearly. Flux flattened the sneakers into plain color blocks and lost the side pattern.

2. Wearing a single accessory

Prompt: "Have the girl wear the necklace shown in the image."

Winner: Flux.

Both put the necklace on and kept the face, the gray tank top, and the background. Flux's choker, emerald, and peacock pendant stay closer to the original piece. Qwen shortened the choker and simplified the pendant. Flux also cropped the frame a little tighter.

3. Translating text

Prompt: "Translate the text in the product images into English."

Winner: none.

Qwen left the Chinese almost untouched and did not translate it. The frame was also stretched out of shape. Flux rewrote the main headline, but the letters are illegible, and the rest of the Chinese was scrambled as well. Neither result is usable as a translation.

4. Placing a product in a scene

Prompt: "Place the product from Figure 1 into the scene shown in Figure 2."

Winner: none.

Both set the bottle on the display stand and kept the bamboo, the stone base, and the lighting. On Qwen, SK-II, FACIAL TREATMENT ESSENCE, and PITERA are just barely readable. Flux garbled the middle line and left white bars on the left and right, though the crop fits a product shot better.

5. E-commerce hero image

Prompt: "Create a main e-commerce product image for this mousetrap, showing it placed in a corner of a living room with a mouse sniffing it."

Winner: Qwen.

Qwen kept the double-door cage and the straight wires behind it, and set it on the carpet between the sofa and the shelf. The mouse is nearby, but it has not leaned in to sniff the opening. Flux's mouse is actually sniffing the opening, and the living room is intact, but the cage door is warped and the product no longer looks like the source. Qwen wins this one.

Taken together, both models are already very strong at image editing. Flux3-klein also supports more precise bbox placement for editing and generation. Qwen-Image-2.1 supports true RGBA output with full transparency. Used together, they cover most image work in everyday cases.