Qwen-Image 2.1 Edits: One Argument Cut Position Drift to About 0.1px

In July, when I measured the pixel position drift of Qwen-Image-Edit 2511, ComfyUI’s standard node silently resized the reference image, and every edit shifted the picture by about 8px. Feeding the reference at its native size through ComfyUI’s stock ReferenceLatent node brought the drift down to about 1px.

Qwen-Image 2.1 combines text-to-image and image editing in one model.
An edit used to take about 12 minutes on my M1 Max, but with the TaylorSeer cache it takes about 4.5 minutes, roughly the same as the ~5 minutes of Qwen-Image-Edit (the 8-step distilled build).
Can it also hold positions in partial edits, such as changing only the facial expressions? I checked with the same source image and the same measurements as in July.

Test environment

Item	Details
Machine	MacBook Pro, Apple M1 Max, 64GB memory
Runtime	Python 3.12.12, PyTorch 2.14.0 (MPS), Diffusers 0.41.0.dev0 (commit 80c7ed26)
Model	Qwen-Image 2.1 (BF16)
Common settings	832×1216, 40 steps, seed 42, CFG 1.0, no negative prompt. One run per condition
Speed-up	TaylorSeer Lite (a cache that computes only some steps for real and predicts the rest; cache_interval=3) together with the KV cache. One condition was also generated without the cache
Source image	Same as July: a full-body standing image of my two original characters, 832×1216
Compared against	Outputs and numbers of Qwen-Image-Edit 2511 (Qwen-Rapid-AIO v23, 8 steps) from the July article
Drift measurement	Phase correlation on edge maps (estimating the shift between two images in the frequency domain), for the whole image and for 3×3 blocks. dx is horizontal, dy is vertical. The July script is reused
Change measurement	Ratio of pixels whose mean absolute RGB difference from the source exceeds 20/255. The face boxes are two hand-drawn rectangles around the two faces, same coordinates as July

I first ran the measurement script on the July Qwen-Image-Edit outputs and confirmed it returns the same values as that article (dx +8.0px for the line art, 0.7% outside the faces for the expression edit).
The measurement script and the edit prompts are in LiltingChannelLabo.

The reference-image resize inside the Diffusers pipeline

When the Diffusers QwenImage21Pipeline receives a reference image, it computes dimensions whose area equals output_resolution squared (1024×1024 by default), resizes the image to them, and then passes it to the VAE (the part that converts between images and the internal representation) and the text encoder. The dimensions are rounded to multiples of 32.

# diffusers/pipelines/qwenimage21/pipeline_qwenimage21.py (excerpt)
def calculate_dimensions(target_area, ratio):
    width = math.sqrt(target_area * ratio)
    height = width / ratio

    width = round(width / 32) * 32
    height = round(height / 32) * 32
    return width, height, None

An 832×1216 source becomes an 832×1248 reference. The width stays the same and only the height is stretched by 2.6%.
If the output size is set to 832×1216, the same as the source, the reference and output grids no longer match.

Setting one argument fixed it. Set output_resolution to the rounded square root of the source’s width times height. For 832×1216 that is 1006: output_resolution squared is then almost the same as the source area, the computed dimensions come out to exactly 832×1216, and the resize no longer changes the size.

Source size	Reference size at the default (1024)	output_resolution to set	Reference size with it
832×1216	832×1248	1006	832×1216
1216×832	1248×832	1006	1216×832
1024×1024	1024×1024	1024	1024×1024
960×1280	896×1184	1109	960×1280
1024×1536	832×1248	1254	1024×1536

This calculation works when the source width and height are multiples of 32. Some sizes, such as 1024×1024 or 768×1344, keep their dimensions even at the default.

image = pipe(
    prompt=prompt,
    image=[source],          # 832×1216
    width=832,
    height=1216,
    output_resolution=1006,  # round(sqrt(832 * 1216))
    num_inference_steps=40,
    true_cfg_scale=1.0,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]

I call the default condition A and output_resolution=1006 condition C. The names follow the July article’s condition A (standard node wired directly) and condition C (ReferenceLatent). I did not run an equivalent of July’s condition B (cropping to 1024×1024) this time.

Experiment 1: measuring drift with line-art extraction

I used the same line-art instruction as in July. Only the way the reference is named was changed from Picture 1 to image 1 to match Qwen-Image 2.1’s prompts.

Convert image 1 into clean black line art on a pure white background. Trace every contour and detail exactly where it is in the original image. Do not move, resize, crop, or redraw anything. No color, no shading, no gray tones.

Model	Condition	Whole-image drift	3×3 block pattern
Qwen-Image 2.1	A (default)	dx +1.1px / dy +0.0px	Left column +6 to +8px, right column −5 to −8px. Squeezed from both sides toward the center
Qwen-Image 2.1	C (output_resolution=1006)	dx +0.0px / dy +0.1px	6 of 9 blocks at 0.1px or less, the rest under 1px
Qwen-Image-Edit 2511 (July)	A (standard node)	dx +8.0px / dy −3.3px	Gradient from +3px at the left edge to +14px at the right
Qwen-Image-Edit 2511 (July)	C (ReferenceLatent)	dx +1.1px / dy +0.5px	About 1px in every block

In condition A the whole-image average is small, but block by block the left and right sides shift in opposite directions. The left column is about +7px and the right column about −8px, and the column centers are about 555px apart. 15 ÷ 555 means the figures became about 2.7% narrower, almost the same as the 2.6% the reference was stretched vertically.
It seems the aspect ratio of the stretched reference was carried over as-is into the 832×1216 output. Vertical drift was under 1px in every block.

In condition C the whole-image drift was 0.1px or less, smaller than the ~1px of July’s ReferenceLatent setup.
Phase correlation gives the translation that best overlaps the contours; it does not mean every single pixel is within 0.1px.

Source edges are in red and output edges in cyan.

Experiment 2: change only the expressions and see if anything else moves

With the same one-line instruction as July’s version without a “keep everything else” clause, I changed both girls’ expressions to a closed-eye smile.

Change both girls' facial expressions to a big cheerful open-mouth smile with happy closed eyes.

Model	Condition	Mean diff	Changed pixels (all)	Inside face boxes	Outside face boxes
Qwen-Image 2.1	VAE round trip only	1.21	0.02%	0.12%	0.02%
Qwen-Image 2.1	A (default)	15.18	15.1%	49.2%	12.8%
Qwen-Image 2.1	C, with TaylorSeer	4.33	3.8%	20.8%	2.6%
Qwen-Image 2.1	C, no cache	4.43	3.9%	21.1%	2.7%
Qwen-Image-Edit 2511 (July)	C (ReferenceLatent)	2.66	1.7%	16.9%	0.7%

In condition A, 12.8% of the pixels outside the faces changed. The narrower figures moved the contours of the whole body.
In condition C that dropped to 2.6%. The values with TaylorSeer and without a cache are almost the same; in this expression edit, the cache made little difference.

The 2.6% outside the faces is more than the 0.7% of July’s Qwen-Image-Edit. A VAE round trip alone gives 0.02%.
The heatmap shows thin red lines along the contours of the whole body. Magnified next to the source, the positions match, but the outlines are slightly thinner and the skin shading is a little lighter. In this output, it seems the contours of unedited areas were also lightly redrawn.

Here are the faces from the source, July’s Qwen-Image-Edit, and this Qwen-Image 2.1 run.

Experiment 3: angry faces, and changing one girl’s outfit

Staying in condition C, I generated two more edits with different instructions.

(angry faces)
Change both girls' facial expressions to an angry face with furrowed eyebrows.

(black sailor uniform for the left girl only)
Change only the left girl's outfit to a black sailor school uniform.

Edit	All	Inside faces	Outside faces	Left half	Right half	Whole-image drift
Angry faces	3.6%	18.8%	2.6%	3.5%	3.7%	dx −0.1px / dy +0.0px
Left girl only, black sailor uniform	14.2%	9.7%	14.5%	25.2%	3.2%	dx −0.1px / dy −0.0px

The angry-face edit changed 2.6% outside the faces, the same value as the smile.
When only the left girl’s outfit was changed, the right half, which was not part of the instruction, still changed by 3.2%, about the same as the outside-face value of the expression edits. In July’s Qwen-Image-Edit the right half of the same edit was 0.42%.
Whole-image drift was about 0.1px in every edit, including the smile.

Transparent background

The outputs of the expression and outfit edits were saved as RGBA PNGs, but their alpha (opacity) was 255 everywhere.
Checking the pipeline code, the Qwen-Image 2.1 VAE takes 4 input channels, and reference images are converted to RGBA before being fed in. I then asked for a transparent background.

Remove the white background of image 1 and make the background fully transparent. Keep the two girls exactly as they are.

Item	Value
Pixels with alpha below 128	57.5%
Pixels with alpha 255	39.9%
Alpha in the background	Mostly 1–8 (exactly 0 for only 1.6%)
Whole-image drift	dx +0.0px / dy −0.0px

The white shirts and white socks on the white background stayed opaque. Removing white with a color threshold would erase the shirts and socks too.
The background alpha did not reach exactly 0; most of it was 1–8.

Front and back references for a side view

The girl with the brown side ponytail on the right of the source image is my original character Kana-chan. In the local-run post, editing her standing image into an exact side view did not keep the side ponytail, which is tied on her own left, in place.
I had heard that feeding in a three-view sheet makes the model pose the character properly, but the side view of that sheet is exactly the image I can’t get. I tried whether giving it the front and the back would produce the side view. For the back I used the rear-view image generated in the local-run post.

The prompt is the side-view prompt from that post; only the way the references are named was changed.

1girl, solo, the girl from image 1, image 2 shows the same girl from behind, keep her face, hairstyle and outfit exactly as in image 1 and image 2,

View	References	Expected	Result	Generation
From her right	Front only	The knot is hidden on the far side of the head	The knot and scrunchie are visible on the near side	270.0 s
From her right	Front and rear	Same	The knot moved a little higher and further back, but is still visible on the near side	331.1 s
From her left	Front only	The knot is on the near side, at the side of the head	On the near side, but high up toward the back of the head	277.8 s
From her left	Front and rear	Same	Almost the same as with one reference	330.6 s

Adding the rear reference did not fix the view from her right. The knot only moved a little higher and further back, and stayed visible on the near side. The view from her left came out almost the same with one reference or two.
Both side views have the knot high up toward the back of the head, and look like the same picture mirrored.
Qwen-Image-Edit also couldn’t produce the side view from a front and a back, so I think this is about what to expect.

Generation time per edit

Process	Generation
Qwen-Image 2.1 edit, 1 reference, TaylorSeer + KV cache	About 270 s
Qwen-Image 2.1 edit, 2 references, TaylorSeer + KV cache	About 331 s
Qwen-Image 2.1 edit, 1 reference, no cache, condition C	677.4 s
Qwen-Image-Edit 2511 (July, 8-step distilled build, ComfyUI)	About 5 min

With two references, the extra reference tokens raised one real step from about 17 s to about 19 s.

Related articles

My other Qwen-Image 2.1 posts.

• Testing Qwen-Image 2.1 Early Access: Composition Prompts and Multi-Reference Limits: Early access on ModelScope Studio: 10 composition presets, a 4-member band, and edits of Kana-chan’s orientation, outfit, and pose
• Qwen-Image 2.1 Local on M1 Max vs Anima: 10 min/Image, Faces Kept in i2i: Running the released weights on an M1 Max 64GB, with the same overhead prompt as Studio and the same prompts as Anima, WAI-Anima, and WAI-Illustrious
• Qwen-Image 2.1 on M1 Max: TaylorSeer Cache Cut 10 min/Image to 3.5 min: The TaylorSeer cache in Diffusers cut about 10 minutes per image on M1 Max to about 3.5

References

• Qwen-Image 2.1 (GitHub)
• pipeline_qwenimage21.py in Diffusers
• Pixel position drift of Qwen-Image-Edit 2511 in local ComfyUI
• Qwen-Image 2.1 with the TaylorSeer cache on M1 Max