Qwen-Image 2.1 Local on M1 Max vs Anima: 10 min/Image, Faces Kept in i2i

Previously, in my ambassador early-access test, I used Qwen-Image 2.1 on ModelScope Studio to try camera composition prompts such as overhead shots, and to edit a standing image of my original character Kana-chan.
The weights (the trained model data) were released on September 20, so I checked whether the same compositions work locally on my M1 Max 64GB.
I also generated the prompts from my April article comparing Anima, WAI-Anima, and WAI-Illustrious, and put the outputs side by side.

Conditions used for the comparison

Qwen was run at 832×1216, 40 steps, seed 42. Steps are the number of iterations that update the image; the seed is the starting number for the random noise. Each condition was generated once.

Generation time covers reading the prompt, the iterative computation, and decoding back to an image. Model loading time is not included.

Computation used BF16, a 16-bit floating-point format, on MPS, which lets PyTorch use the Mac GPU. CFG, which controls how strongly the model follows the prompt, was 1.0, and no negative prompt (a list of things you don’t want) was used.

The sampler is the method that turns noise into an image, and the scheduler is how that computation is stepped. For Qwen I used whatever Diffusers, the image-generation library, loaded from the published weight config.

Python environment for M1 Max

The official example uses QwenImage21Pipeline from Diffusers.
I changed the device from CUDA to MPS and installed it in a Python environment separate from my existing ComfyUI.

I first built the environment with PyTorch 2.13.0, but model loading stopped with this error.

Qwen3VLVideoProcessor requires the Torchvision library

Transformers needed Torchvision when loading the processor, the component that prepares the inputs.
Installing Torchvision also bumped PyTorch to 2.14.0, so all timing measurements were done on the updated environment.

Install the versions used for generation into a dedicated folder.

uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python \
  'torch==2.14.0' 'torchvision==0.29.0' \
  'transformers==5.17.0' 'accelerate==1.15.0' \
  'modelscope==1.40.1' 'pillow==12.3.0' \
  'git+https://github.com/huggingface/diffusers@80c7ed262aeffbeb43ef13ae04baeb9b84515a69'

Downloading the weights from ModelScope

The September 20 announcement in the official repository lists ModelScope and Hugging Face as download locations.
I downloaded the weights listed on ModelScope’s file page with the ModelScope SDK from the environment above.

import os

os.environ["MODELSCOPE_DOWNLOAD_PARALLEL_WORKERS"] = "4"
from modelscope import snapshot_download

snapshot_download(
    "Qwen/Qwen-Image-2.1",
    revision="master",
    endpoint="https://modelscope.ai",
    local_dir="model",
    cache_dir="cache",
    max_workers=4,
)

The image generator, the text encoder that reads the prompt and reference images, and the VAE that converts between images and the internal representation came to 7 weight files.
The total was 33,115,613,408 bytes, about 33.1GB. The official 7B (7 billion parameters) counts only the image generator; the text encoder and VAE come on top of that.
After downloading, I checked the size and SHA-256 (used to detect corruption) of all 7 files against the published list, and they all matched.

The license is the Qwen Research License: non-commercial research and evaluation use, with a separate license required for commercial use.

Smoke test with a small image

First I checked that a 512×512 image could be saved with a short neon-sign prompt.
Seed 42, BF16, and MPS were the same as the main runs; steps were reduced to 4.

A neon shop sign that reads "QWEN IMAGE 2.1", rainy night, reflections on wet pavement

Even at 4 steps, the PNG saved fine.

Same overhead prompt as Studio

From here, I raised the steps to 40 to match the official example.
I reused the overhead prompt I sent on Studio, copied straight from the earlier run log. It looks down on the subject from above and has her lift her face toward the camera.

masterpiece, best quality, high quality, newest, year 2025, year 2024,
1girl, solo, long blonde hair, blue eyes, school uniform, white shirt, navy skirt,
strong high-angle overhead view, camera clearly above the subject looking steeply downward, visible top surfaces and pronounced top-down perspective. The subject is seen from the head toward the legs with the head visually nearer and larger than the lower body; the ground fills most of the background. The subject tilts her face upward and looks up into the camera.

from pathlib import Path
import torch
from diffusers import QwenImage21Pipeline

prompt = Path("prompt.txt").read_text()
pipe = QwenImage21Pipeline.from_pretrained(
    "./model", torch_dtype=torch.bfloat16, local_files_only=True,
).to("mps")

image = pipe(
    prompt=prompt,
    width=832,
    height=1216,
    num_inference_steps=40,
    true_cfg_scale=1.0,
    use_kv_cache=True,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("local-from-above.png")

In the Diffusers code used here, negative-prompt guidance does not run when true_cfg_scale=1.0.

Each condition is a single run at seed 42, so I can’t tell whether differences are normal variation or come from the runtime environment.
The Studio timing last time was a rough impression, so it can’t be compared with the measured local speed.

Studio (previous)

M1 Max (this time)

The overhead composition came out locally just as it did on Studio.
For detail, the Studio version was finer when I compared the original PNGs at the same size: thin hair strands and highlights, shirt wrinkles, and a plaid skirt. The local version had smoother hair and clothing surfaces.

Generation time and memory on M1 Max

I recorded time and memory for the overhead generation above. At 832×1216, each step took about 15 seconds.
To check the effect of other processes, I stopped the Astro dev server (used to preview this blog) running on the same Mac after step 25. Its CPU usage was already close to 0% before stopping.

Step times were measured as the difference between step end times, averaging from step 2 using the first end time as the reference.
Step 26, which spans the stop, was excluded. Each step still took about 15 seconds after stopping.

Generation includes reading the text, 40 iterations, and VAE decoding back to an image.
Model loading and PNG saving were timed separately, and Python import time is not included anywhere in the table. Timestamps were taken after waiting for the GPU work to finish.

Memory was recorded as the MPS driver allocation: about 33.9GB during the iterations and about 45.8GB after the VAE step.
The GPU allocation that PyTorch reports includes caches as well as model data.
I sampled once per second and at stage boundaries, so short peaks in between may have been missed.

Swap (temporarily moving memory contents to the SSD) stayed at about 0.5MB system-wide from start to end.

Editing Kana-chan’s standing image into an overhead shot

With text-to-image working, I fed in the same standing image of Kana-chan that I used for the Studio edit.

The prompt is the following, the same one I sent for the Studio edit.

masterpiece, best quality, high quality, newest, year 2025, year 2024,
1girl, solo, the girl from image 1, keep her face, hairstyle and outfit exactly as in image 1,
strong high-angle overhead view, camera clearly above the subject looking steeply downward, visible top surfaces and pronounced top-down perspective. The subject is seen from the head toward the legs with the head visually nearer and larger than the lower body; the ground fills most of the background. The subject tilts her face upward and looks up into the camera.

I swapped this into prompt in the generation code and passed image=[Image.open("kana-front-arms-down-normal-s42.png").convert("RGB")].
Image comes from Pillow via from PIL import Image. Resolution 832×1216, 40 steps, seed 42, and BF16 were the same as text-to-image.

Locally, CFG was 1.0 with no negative prompt. On Studio I had also set a negative prompt for the edit, and its internal steps and CFG remain unknown.

Studio (previous)

M1 Max (this time)

In the Kana-chan edit, the local version also came out as a view from above, with the head drawn closer and larger.
Studio’s shirt wrinkles and ground texture stand out, but the local version also has thin hair lines and eye highlights, and the detail gap was smaller than in the text-to-image pair.

Time and memory were recorded the same way as in the previous section.

The MPS driver allocation was about 38.2GB during the iterations and about 49.0GB after the VAE step. Swap stayed at about 0.5MB.

Does Kana-chan’s asymmetric hairstyle survive?

Kana-chan ties her hair on her own left side. In the front-facing source, the side ponytail is on the right of the frame; I tested whether it stays on her left when the orientation and pose change.
The input was the same front-facing standing image used for the overhead edit. The main question was whether the asymmetric hairstyle holds, and I also checked the limb instructions.

I used the position of the blue scrunchie to check whether the hair is tied on the same side as in the source.

832×1216, 40 steps, seed 42, BF16, CFG 1.0, no negative prompt. One run per condition, so run-to-run variation is unknown.

The three composition changes used a prompt that keeps the face, hairstyle, and outfit from the source. The two pose changes also asked to keep the side ponytail, the blue scrunchie, and the ahoge (the strand sticking up from the top of the head).

For the right-side view, I fixed an on-screen direction that contradicted itself in the prompt. The corrected prompt is in the collapsed section at the end.

Side view (from her left)

Side view (from her right)

Directly behind

Jump

Running at an angle

The rear view and the front-facing jump kept the hair tied on her left side.
In the two side views, the view from her right flipped the tie side, and the view from her left turned it into something close to a regular ponytail. In the angled run, the knot moved toward the back of the head and the arm swing was reversed.

masterpiece, best quality, high quality, newest, year 2025, year 2024,
1girl, solo, the girl from image 1, keep her face, hairstyle and outfit exactly as in image 1,
from side, profile, exact orthographic 90-degree eye-level side view, camera at the subject's left side. Full body head to toe, flat left-facing silhouette. No head, body or gaze turn toward camera; no eye contact. Exact profile, never front or three-quarter.

masterpiece, best quality, high quality, newest, year 2025, year 2024,
1girl, solo, the girl from image 1, keep her face, hairstyle and outfit exactly as in image 1,
from side, profile, exact orthographic 90-degree eye-level side view, camera at the subject's right side. Full body head to toe, flat right-facing silhouette. No head, body or gaze turn toward camera; no eye contact. Exact profile, never front or three-quarter.

masterpiece, best quality, high quality, newest, year 2025, year 2024,
1girl, solo, the girl from image 1, keep her face, hairstyle and outfit exactly as in image 1,
direct rear view, eye-level camera positioned directly behind the subject, with zero camera roll. Full body from head to toe in a centered rear-view composition. Show only the back of the head, torso, arms, legs and shoes; do not show the face. The subject looks straight ahead away from the camera.

masterpiece, best quality, high quality, newest, year 2025, year 2024,
1girl, solo, the girl from image 1, keep her face, hairstyle, side ponytail with the blue scrunchie, ahoge, eye color, outfit and white background exactly as in image 1,
change only her pose: a dynamic mid-air jump facing the viewer. Both feet are high off the ground, her right knee is pulled up high and her left leg stretches down, her right fist punches straight up toward the sky and her left arm swings out to the side. Full body from head to toe.

masterpiece, best quality, high quality, newest, year 2025, year 2024,
1girl, solo, the girl from image 1, keep her face, hairstyle, side ponytail with the blue scrunchie, ahoge, eye color, outfit and white background exactly as in image 1,
change only her pose: three-quarter view, not frontal. Her body and face are turned about 45 degrees toward the viewer's left, so her left side is visible. She is sprinting toward the viewer's left: her body leans far forward, her right knee is driven up high, her left leg pushes off behind her with only the toes touching the ground, her left arm swings forward and her right arm swings back with bent elbows. Full body from head to toe.

Same prompts as Anima and Illustrious

Using the same English prompts, 832×1216, and seed 42 as my WAI-Anima article, I generated the white-and-gold robe standing image and the running scene with a castle at sunset on Qwen-Image 2.1.

For Anima preview3-base, WAI-Anima v1, and WAI-Illustrious v160, I reused the outputs from the April 16 article as-is. They were not regenerated.

The Qwen row lists the Diffusers class name; the other three list sampler and scheduler.

The April article did not record negative prompts for those three models. Steps and CFG also differ by model.

Standing image on a white background

Qwen-Image 2.1
(this time)

Anima preview3-base
(April output)

WAI-Anima v1
(April output)

WAI-Illustrious v160
(April output)

Qwen’s standing image has lighter shading on the face and clothes than the two Anima images, and thinner hair lines. WAI-Illustrious is also lightly shaded, but Qwen’s patterns on the capelet, sleeves, and hem are finer, and the outfit shape differs from Illustrious’s large gold front panel.

1girl, solo, long blonde hair, blue eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, standing, looking at viewer, full body, white background

Running with a castle at sunset

Qwen-Image 2.1
(this time)

Anima preview3-base
(April output)

WAI-Anima v1
(April output)

WAI-Illustrious v160
(April output)

For the running pose, WAI-Anima swinging its arms and WAI-Illustrious leaning forward look more like running than Qwen with its arms down.

1girl, solo, long blonde hair, blue eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, running, wind, hair blowing, dynamic pose, fantasy landscape, castle in background, sunset sky, dramatic clouds, grass field

i2i (WAI-IL → Qwen)

In the April article, I ran a standing image made with WAI-Illustrious v160 and my Kana-chan LoRA through the Anima models with i2i at denoise 0.5.
I fed the same source image into Qwen-Image 2.1.

The Diffusers QwenImage21Pipeline used here has no denoise (strength) argument, which sets how much of the source image gets changed.
It is an edit driven by a reference image and a text instruction, so I asked in English to redraw it keeping the face, hairstyle, outfit, pose, composition, and background.
This is a different mechanism from the Anima models’ denoise 0.5.

The source is a standing image with a gold circle in the background and a side ponytail tied with a blue scrunchie.

WAI-Illustrious v160 source

→ Qwen-Image 2.1
(this time)

→ preview3-base
(April output)

→ WAI-Anima v1
(April output)

The two Anima images have rounder faces and look younger than the source. Qwen’s face was almost the same as the source.
On the other hand, only Qwen’s output lost the gold circle in the background. Generation took 672.00 seconds.

Redraw image 1 as a new illustration. Keep the same girl, her face, brown hair with the side ponytail and blue scrunchie, the white robe with gold embroidery, capelet and gold sash, the standing pose, the composition and the background exactly as in image 1.

Generation speed of the four models

I added a Qwen-Image 2.1 row to the April speed table.
It’s the same M1 Max, but the three April models ran on ComfyUI while Qwen ran on Diffusers, and the step counts differ.

Per image, Qwen took a bit over twice as long as the Anima models.
Scaling the generation time linearly by step count, it would still be about 440 seconds at Anima’s 30 steps.

Appendix: NSFW output test

The images below are blurred since the test outputs may be sensitive. If you want to see the actual results, try them in your own environment.

I generated the same two prompts from the April article’s appendix on Qwen-Image 2.1. Settings match the standing and running tests, with no negative prompt.
The rows for the three April models in the tables are copied from the April article’s notes.

Without the nsfw tag

First, a nude instruction without the nsfw tag.

Qwen-Image 2.1
(this time)

preview3-base
(April output)

WAI-Anima v1
(April output)

WAI-IL v160
(April output)

The standing and running tests, which also start with the 1girl tag, came out as anime art. None of the prompts include a style tag, but this one image came out photorealistic. Generation took 555.78 seconds.

1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, sitting on edge of spring, feet in water, forest, natural light, serene, full body, looking at viewer

With the nsfw tag (front, full body)

Next, the nsfw tag added, with a front-facing full-body pose.

Qwen-Image 2.1
(this time)

preview3-base
(April output)

WAI-Anima v1
(April output)

WAI-IL v160
(April output)

In the April article, WAI-Illustrious and WAI-Anima v1 kept the same face between the clothed standing image and the nude one.
Qwen’s white-background standing image had narrow eyes and a mature face, while this one has large eyes and a round face, a different art style. Generation took 557.42 seconds.

1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, nsfw, straight-on, full body, standing, white background