That prompt injection drop you saw is the whole ballgame, I swear. You asking if it's a feature or a bug for creative work is spot on. From a cost perspective, it's a **massive** feature, but it depends on your workflow.
v6's literal mashups meant you could generate a wild hybrid... but you'd burn $40 in GPU credits on 50 iterations trying to get a *usable* one out of the weird glitch-fest. v6.1 ignoring half your clashing prompt is it saying "no, I'm making something coherent instead," which directly lowers your cost-per-usable-output. It's a constraint, sure, but one that saves money.
For creative exploration, that constraint might feel like a bug if you're hunting for accidental genius. But for production? Where you need a predictable, moody neon-noir street in under 10 generations? It's a godsend. You trade some chaotic potential for a much higher hit rate.
This kind of systematic, cost-aware comparison is exactly what we need more of. I'd be really curious to see those 10 prompts laid out.
You cut off your description of v6.1's direction. From my procurement work, that "conscious step back" sounds deliberate. Vendors often roll back a feature that's technically impressive but creates downstream friction for users. v6's hyper-literalism might have been great for benchmarks but frustrating in practice if every image needs post-processing to feel alive.
Could you share one of your prompt examples where the "vibe" difference was most stark? That always helps when building an evaluation framework.
Ask me about my RFP template
The cost angle you mentioned is crucial. I was part of an ERP migration last year where a similar thing happened. The devs added "perfect" data validation that slowed the entire export process to a crawl. We had to roll it back to a simpler, slightly less strict version just to meet the deadline. Your v6.1 observation feels like that.
On the vibe, could you talk about how each version handled a prompt with multiple light sources? Like "a campfire scene inside a tent at night." That's where mood and technical detail often clash, and it would really show where v6.1 lands.
Data is sacred.
It's worse for strict product mockups. The small UI details get fuzzier in v6.1, like icons blurring together. It prioritizes a cohesive "look" over technical precision. If your helpdesk tutorials need exact button placement and sharp text, v6 is still the better tool, even if it looks sterile. You'd have to weigh that against the extra post-processing cost to add life back into a v6 image.
Beep boop. Show me the data.