A complete visual archive of 40 bathroom edits produced with
Qwen-Image-2.1-Turbo using the --task edit mode of
mlx-vlm. Each source bathroom scene has its wall-mounted
George Washington portrait replaced with one of three rotating reference images:
the cow-jumped-over-the-moon drawing, the KISS bomber jacket, and a
synchronised-swimming scene. Every output is generated at
2048 × 2048 with --steps 8 --guidance 1, running serially
on Apple Silicon with MLX.
Note: All gallery images render at their native 2048 × 2048 resolution.
On narrow screens the gallery scrolls horizontally to preserve every pixel.
Summary of the v2 bathroom edit batch (cow + kiss + swim).
The exact command used for every image in this archive.
mlx_vlm.generate \
--output-modality image \
--task edit \
--model ./Qwen-Image-2.1-Turbo \
--image inputs/bathroom-N.png \
inputs/<reference>.png \
--prompt "The wall-mounted gold frame in the upper-left of the image
contains a portrait of George Washington. Replace only its
contents with the second reference image. Do not move the
frame; keep it exactly where it is on the wall at the same size." \
--size 2048x2048 \
--steps 8 \
--guidance 1 \
--seed S \
--output outputs-v2/bathroom-N-<ref>-seedSSS.png
Each edit feeds the model two images: the bathroom scene (base) and the reference
(content to insert). The prompt anchors the frame's location and preserves the rest
of the scene. Seeds rotate through (seed - 1) % 3 to cycle the three
reference images: cow → kiss → swim.
Grouped by source bathroom. Each image is 2048 × 2048 native.
Frame in the upper-left of the wall, above the stainless steel partition. Seeds cycle cow (1, 4, 7, 10), kiss (2, 5, 8), swim (3, 6, 9).
Frame on the tiled wall, warm overhead lighting. Same seed rotation.
Frame on the right wall, viewed at an angle. Swim reference is the hardest case here.
Frame on the left wall in the corner. Best swim results across the batch.
Complete, unedited logs from the v2 batch run.