Appearance
Compare Qwen-Image-2.1 generation and editing times across local AI hardware. Explore image outputs and compare Base and Turbo editing results.
Qwen's unified text-to-image and image-editing model, with a 7B visual generator (32 single-stream DiT layers). It can also make transparent (RGBA) images and take up to 10 reference images for edits. Turbo uses 8 steps.
At 1024×1024, using the same prompts and seeds across configurations. Times are per image after the model is loaded. Multipliers show time relative to Olares One.
Scroll horizontally to compare all results.
| Machine | Text to imageBase | Image editBase | Image editTurbo |
|---|---|---|---|
| Olares OneComfyUI, Nunchaku FP4 | 12 s | 17 s | 5.2 s |
| DGX SparkComfyUI, NVFP4 | 27 s2.2× | 37 s2.1× | 13 s2.5× |
| Mac Studio M5 MaxComfyUI, int8, --gpu-only | 1.6 min7.9× | 1.9 min6.5× | 30 s5.8× |
| GTR9 ProComfyUI, int8 | 1.6 min8.1× | not run | 47 s9.0× |
Time per image
Lower is better| Machine | Configuration | Per image | First image |
|---|---|---|---|
| Olares OneComfyUI, Nunchaku FP4 | ComfyUI, Nunchaku FP4 | 12.09 s | 16.57 s |
| Olares OneComfyUI, NVFP4 | ComfyUI, NVFP4 | 12.87 s | 16.65 s |
| Olares OneComfyUI, int8 | ComfyUI, int8 | 15.55 s | 18.08 s |
| DGX SparkComfyUI, NVFP4 | ComfyUI, NVFP4 | 27.11 s | 29.36 s |
| DGX SparkComfyUI, int8 | ComfyUI, int8 | 29.39 s | 34.04 s |
| Mac Studio M5 MaxComfyUI, int8, --gpu-only | ComfyUI, int8, --gpu-only | 94.96 s | 100 s |
| GTR9 ProComfyUI, int8 | ComfyUI, int8 | 97.87 s | 100 s |
| Mac Studio M5 MaxComfyUI, int8 | ComfyUI, int8 | 100 s | 103 s |
| GTR9 ProComfyUI, NVFP4 | ComfyUI, NVFP4 | 171 s | 174 s |
Time per image
Lower is better| Machine | Configuration | Per image | First image |
|---|---|---|---|
| Olares OneComfyUI, Nunchaku FP4 | ComfyUI, Nunchaku FP4 | 17.45 s | 21.27 s |
| Olares OneComfyUI, NVFP4 | ComfyUI, NVFP4 | 20.12 s | 22.41 s |
| Olares OneComfyUI, int8 | ComfyUI, int8 | 22.91 s | 25.77 s |
| DGX SparkComfyUI, NVFP4 | ComfyUI, NVFP4 | 37.25 s | 39.51 s |
| DGX SparkComfyUI, int8 | ComfyUI, int8 | 41.46 s | 42.88 s |
| Mac Studio M5 MaxComfyUI, int8, --gpu-only | ComfyUI, int8, --gpu-only | 114 s | 119 s |
| Mac Studio M5 MaxComfyUI, int8 | ComfyUI, int8 | 147 s | 151 s |
Time per image
Lower is better| Machine | Configuration | Per image | First image |
|---|---|---|---|
| Olares OneComfyUI, Nunchaku FP4 | ComfyUI, Nunchaku FP4 | 5.23 s | 9.01 s |
| Olares OneComfyUI, int8 | ComfyUI, int8 | 6.74 s | 9.65 s |
| Olares OneComfyUI, NVFP4 | ComfyUI, NVFP4 | 7.64 s | 9.66 s |
| DGX SparkComfyUI, int8 | ComfyUI, int8 | 13.00 s | 15.55 s |
| DGX SparkComfyUI, NVFP4 | ComfyUI, NVFP4 | 13.11 s | 15.82 s |
| Mac Studio M5 MaxComfyUI, int8, --gpu-only | ComfyUI, int8, --gpu-only | 30.19 s | 36.12 s |
| GTR9 ProComfyUI, int8 | ComfyUI, int8 | 47.01 s | 51.60 s |
| Mac Studio M5 MaxComfyUI, int8 | ComfyUI, int8 | 62.64 s | 65.89 s |
| GTR9 ProComfyUI, NVFP4 | ComfyUI, NVFP4 | 64.40 s | 69.41 s |
Pick a prompt, then compare images and generation times across hardware configurations. Base model at 40 steps.
Portraits
E-commerce
Interiors and real estate
Street portrait of a Black man in his thirties with a short beard and close-cropped hair, wearing a faded denim jacket over a white t-shirt, standing on a Brooklyn sidewalk with brownstone buildings behind him, overcast soft light, looking at the camera with a calm confident expression, waist-up framing, realistic photo, 50mm lens.
Same prompt and seed on every configuration. Times are per image. * marks the first image of a batch, which includes loading the model.
Pick an original photo and instruction. Compare Base at 40 steps with Turbo at 8 steps across hardware configurations.
Portraits
E-commerce
Interiors and real estate

Move her to a whitewashed Mediterranean rooftop terrace overlooking the sea at golden hour, with warm rim light on her hair and a soft orange sky behind her. Keep her face, hair, sweater, pose and expression exactly the same.
Same prompt and seed on every configuration. Times are per image. * marks the first image of a batch, which includes loading the model.
All people, pets and brands shown are AI-generated and fictional.