Qwen Image 2.1 on Mac: Image Quality Tested on an M4 Max
Three first-run wallpapers on an M4 Max: real images, original sizes, timings, composition notes, and a still-life failure, with reproducible prompts.
On October 30, 2026, we generated three wallpapers with Qwen Image 2.1 through our local ComfyUI setup at HueWalls: a blue morpho butterfly, an alpine lake, and a persimmon still life. Each prompt was run once. We kept the first results rather than changing seeds until we found something flattering.
The natural scenes and quiet wallpaper layouts worked well enough to make this a useful local workflow for us. The still life exposed a limitation: an attractive composition does not guarantee correct object counts or clean shapes.
What we actually tested
The official Qwen repository describes 2.1 as a unified generation and editing model with a 7B visual generation component and support for transparent images. This test covers ordinary text-to-image wallpapers only. We did not test editing, transparency, or typography.
We used local INT8 weights, not the hosted Pro API or the eight-step Turbo checkpoint. Our observations apply to this configuration.
| Component | Configuration |
|---|---|
| Machine | Apple M4 Max, 64GB unified memory |
| OS | macOS 15.3 |
| ComfyUI | 0.42.5 |
| Python / PyTorch | 3.13.12 / 2.12.1, MPS |
| Diffusion model | qwen_image_2.1_int8_convrot.safetensors |
| Text encoder | qwen3vl_8b_int8_convrot.safetensors |
| VAE | qwen_image_2.1_vae_bf16.safetensors |
| Sampling | 25 steps, CFG 1, Euler, Simple, denoise 1 |
| Prompt enhancement / negative prompt | Disabled / empty string |
The workflow uses QwenImage21Cache and TextEncodeQwenImage21, followed by sampling, VAE decoding, and Lanczos resizing. We saved both the decoded original and the finished wallpaper.
Original size, delivery size, and elapsed time
| Sample | Generated original | Finished wallpaper | Elapsed | Seed |
|---|---|---|---|---|
| Blue morpho | 768×1664 | 1080×2340 | 234.34 s | 2026103001 |
| Alpine lake | 1280×800 | 2560×1600 | 180.45 s | 2026103002 |
| Persimmon still life | 768×1664 | 1080×2340 | 225.57 s | 2026103003 |
Elapsed time runs from submission until completion was observed. It includes any loading, encoding, sampling, decoding, resizing, and polling overhead for that run, but excludes downloading the images. Each scene was tested once, without controlling model cache state or system load. These are individual observations, not a hardware or model speed benchmark.
The phone originals were enlarged by about 1.406× and the desktop original by 2×. 1080×2340 and 2560×1600 are delivery sizes, not the native generation sizes in this test. We did not test native 4K. Lanczos resizing cannot recover details absent from the source image.
Blue morpho: useful space for a lock-screen clock

The butterfly sits in the lower half, with a large, quiet area of dark green bokeh above it. This follows the requested clock-friendly layout. Fine blue wing texture, dark veins, fern leaves, and droplets remain visible in the original, with clear separation between subject and background.
For this wallpaper, the composition feels more usable than a subject filling the entire canvas. Visible detail should not be confused with scientifically accurate anatomy; we did not verify the insect's anatomical structure.
Alpine lake: a usable desktop layout, with dark shadows

The trees sit on the right, leaving open water on the left as requested. Mountains, woodland, and reflections create distinct layers, while the cool surroundings and warm light on the peaks work well together.
The right-hand trees are close to black, however, and much of their detail is lost in shadow. This suits a subdued desktop. For a brighter morning scene, preserving shadow detail would be a sensible next prompt adjustment. We did not rerun that adjustment in this test.
Persimmon still life: the mood works, the shapes need review

Warm grey linen, an ivory ceramic bowl, and soft leafy window shadows follow the prompt well. The objects sit low in the frame, with plenty of quiet space above.
The prompt explicitly requested exactly two ripe persimmons. The foreground fruit instead has an ambiguous split or merged outline, and the calyx placement makes the count difficult to read clearly. We would not count this as a pass for precise object control. If the content required two clearly separate, intact persimmons, this image would need editing or another generation.
It is a useful reminder to inspect object boundaries, counts, and local shapes separately from a picture's overall mood.
Reproduce the prompts
All three use the sampling settings above, with the listed seed and dimensions for each image. Different weights, node versions, or environments may still produce different results.
Blue morpho
Photorealistic macro photograph of a single blue morpho butterfly resting on a deep green fern in a misty forest, fine delicate wing scales, soft dawn light and natural droplets, dark emerald background with smooth bokeh. Vertical phone wallpaper, butterfly in the lower middle, upper third calm and uncluttered for a lock-screen clock. No text, no logo, no watermark.
Alpine lake
Photorealistic wide landscape photograph of a quiet alpine lake at sunrise, distant snow-covered peaks reflected in still water, a small group of dark pine silhouettes on the right, subtle low mist, restrained cool blue and warm peach palette, realistic mountain geology. Horizontal desktop wallpaper, calm open water on the left with room for desktop icons. No text, no logo, no watermark.
Persimmon still life
Editorial still-life photograph of exactly two ripe orange persimmons beside one small handmade ivory ceramic bowl on a pale warm-grey linen surface, one soft leafy shadow from a window, realistic fruit skin and subtle ceramic glaze, quiet autumn morning atmosphere. Vertical phone wallpaper, objects grouped in the lower half, upper third empty and softly lit for a lock-screen clock. No text, no logo, no watermark.
Where we would use it
Based on these three first outputs, we would keep using this local workflow for natural landscapes, single-subject macro scenes, and wallpapers with deliberate negative space. It follows useful layout instructions, while precise counts and local shapes still require human review.
We would also keep recording original dimensions and resizing steps. When more detail is needed, higher-resolution generation or a separate super-resolution experiment should be tested rather than simply enlarging the file. Three images document what happened in this session; they do not establish overall image quality or reliability.
References: Qwen Image 2.1 official repository and release announcement. Images and timings in this article come from this HueWalls local test.