I’ve been testing the latest Midjourney models and made a side-by-side comparison grid. I ran the same 10 prompts through v5.2, v6, and the new v6.1 alpha.
Some immediate observations:
- v6.1 seems much better with text rendering and small details, like intricate patterns on clothing.
- v6 feels stronger on photorealistic lighting, but sometimes overdoes contrast.
- v5.2 still holds up for certain stylized, painterly looks.
I’m curious if others are seeing the same trends. Especially around how each version handles complex prompts with multiple details. Does v6.1’s “improved prompt understanding” actually lead to fewer ignored elements?
Interesting parallel to model versioning in the ML ops side of things. We see similar "regression in one area, improvement in another" patterns when rolling out new inference service models. Your note about v6.1 ignoring fewer prompt elements is key - that's a classic precision/recall trade-off. Higher fidelity to the prompt spec can sometimes come at the cost of overall aesthetic coherence, which might be why v5.2 retains a niche for stylized work. Have you quantified how many of the ten prompts had all requested elements rendered in each version? A simple compliance score would be telling.