Anyone else having issues with inconsistent output quality?

Hey folks! 👋 Wanted to bring up something I've been meticulously tracking in my Playground AI workflows over the last few weeks, and I'm curious if it's just me or a wider pattern.

I'm a huge fan of the platform for generating assets for campaign landing pages and social ads, but I've been hitting a real snag with output consistency. I can run the **exact same prompt**, with the **same model** (say, Stable Diffusion XL), the **same seed**, and very similar CFG/step settings, and get wildly different results in quality and style between sessions. One day, the generations are crisp, detailed, and perfectly on-brand; the next, using the same saved preset, the images come out muddy, with weird artifacts, or the composition just falls apart.

Here’s a concrete example from a recent project. My prompt for a "tech consultant" character was locked in and working great:
```
professional portrait of a friendly female tech consultant in a modern office, confident smile, sharp business casual attire, clean background, photorealistic, detailed eyes, professional lighting, 85mm lens
```
**Parameters:** `SDXL, Seed: 12345, Steps: 30, CFG Scale: 7, Sampler: DPM++ 2M Karras`

I documented a fantastic result on Monday. Ran it again Thursday, and the output had this strange, over-smoothed face, a warped hand, and a greenish tint. The only variable was time/server load, I suppose?

I've tried to isolate the issue by checking:
* **Prompt & Parameters:** Saved as a template, no changes.
* **Browser/Connection:** Tried different browsers and networks.
* **Time of Day:** Noticed some correlation with slower generation times (peak hours?) leading to more flaws, but not a hard rule.

This inconsistency is a major hurdle for any kind of scalable production. It turns what should be a reliable asset pipeline into a lottery.

**My main questions for the community:**
* Is anyone else running into this quality volatility, especially with photorealistic or detailed illustrative styles?
* Have you found any specific parameter tweaks or workflows that act as a "stabilizer"?
* Could this be related to how Playground allocates GPU resources across different node/backends? I'd love to hear any technical insights.

I'm keeping a detailed log of my generations (seed, timestamp, output score) to try and find a pattern. If others are seeing this, maybe we can pool our data and see what's up!