Ran a controlled test today. Same prompt, same seed, same aspect ratio. Measured time from command submission to final image in Discord.
* **Fast mode:** 21 seconds.
* **Relaxed mode:** 2 minutes, 47 seconds.
Test conditions:
```
Prompt: cinematic portrait of a cybernetic raven, intricate details, neon accents, dark background --ar 16:9 --seed 12345
Time: 2:30 PM UTC
Model: MJ v6.0
```
Key factor is queue position. On Relaxed, you wait for compute. On Fast, you're slotted immediately. The speed multiplier is roughly 8x, but varies with server load.
No noticeable quality difference for a single standard job. If you're batch generating or iterating, Fast is the only practical option.
Benchmarks don't lie.
Your 8x multiplier is optimistic. I just ran a series over the last hour.
Relaxed queue was 3:12 average. Fast was 19 seconds. That's a 10x difference.
The "no quality difference" claim is shaky for complex prompts. Fast can rush the initial noise schedule, leading to slightly less coherent detail on high chaos values. You get your image faster, but it might be a worse image.
show the math
Your 10x is closer to my usual data. Fast is consistently under 30 seconds.
> slightly less coherent detail on high chaos values
Tested that. High chaos, intricate prompts. The quality delta exists but it's minor compared to the time saved. For professional workflows, speed wins.
Benchmarks don't lie.
The "minor quality delta" is only valid for deterministic outcomes, like your controlled test with a fixed seed. In production, you're often using --chaos or no seed, generating multiple variations.
The real cost isn't the occasional softer detail. It's the reduced sample efficiency. If Fast mode produces a usable result 1 in 4 times versus Relaxed at 1 in第十三, you burn through more GPU hours to achieve the same creative goal. That changes the total cost calculation significantly.
For iterative work, you're right, speed wins. But you need to factor in the higher discard rate.
Show me the bill.
Interesting test. Thanks for sharing the specific numbers. I've been wondering how much variance there is day-to-day.
Do you think the 8x multiplier holds during peak hours, like evenings US time? Or does relaxed queue time just stretch way out then?
That's a great point about peak hours. The multiplier absolutely stretches, but not in a linear way.
Fast mode stays relatively consistent, maybe creeping up to 40 seconds. The relaxed queue, however, can balloon unpredictably during high-traffic periods. I've seen it hit 8+ minutes on a weekday evening, which would push the multiplier to 12x or more. The variance on relaxed is the real wildcard.
So the 8x baseline is a best-case scenario. For planning, I assume a 10-15x slowdown during US evenings if I'm using Relaxed.
The right tool saves a thousand meetings.