Midjourney's Aug 29 edit update: spot silent image-model changes

Midjourney's update feed gives three dates (Midjourney updates feed, read 2026-10-04). V8.2 launched on July 24, 2026. The V8.2 edit model opened for testing on August 27, with text edits, up to 4 image references and inpainting or outpainting. On August 29 Midjourney improved the edit model's output quality and asked users who had issues in the previous 24 hours to retry. The model name did not change; the output did.

Why it matters if you build on image models

A vendor can improve a model in place, and that is good news until your brand-checked prompt no longer looks the same. Midjourney is not in the Sume image catalog, so this is not a Sume change log; it is an example of a pattern that applies to every hosted model you depend on. Sume's own catalog is read with GET /v1/images/models (Sume Image API), and its rows change as models are added or retired.

A five-job canary

The script compares two saved outputs and prints how far apart they are on a 0 to 255 scale.

• Pick five real jobs: a prompt, its reference URLs and the settings.
• Run them monthly and after any vendor announcement, and keep the output files.
• Compare each new output with the last one and review any with a large difference.
• Store the job id with every file, so you can say which run made it.

A mean difference alone does not say better or worse, so treat a jump as a prompt to look, not a verdict.

from PIL import Image, ImageChops
import numpy as np

def drift(path_a: str, path_b: str) -> float:
    a = Image.open(path_a).convert("RGB")
    b = Image.open(path_b).convert("RGB").resize(a.size)
    diff = np.asarray(ImageChops.difference(a, b), dtype=float)
    return float(diff.mean())

print(drift("canary-jul.png", "canary-oct.png"))