Hey everyone! 👋
I've been running Dream Machine through its paces for a few weeks now, trying to integrate it into a content generation pipeline, and I've hit a consistent snag. No matter what prompt I throw at it—descriptive, simple, abstract—about 90% of my outputs have these strange, bright green artifacts. They look like splotches or sometimes geometric smears, usually in the corners or edges of the frame. It's making the outputs pretty unusable for my project.
My initial thought was a VRAM issue or a corrupted install, but I've ruled that out. I'm running on an A10 GPU with 24GB VRAM, and I've reinstalled the environment from scratch twice. The artifacts appear both in the web interface and when using the API.
Here's the exact config I'm using for generation via the Python client, which is pretty standard:
```python
response = client.generation.generate(
prompt="a serene lakeside at sunset",
style_preset="cinematic",
num_images=4,
parameters={
"cfg_scale": 7,
"steps": 28
}
)
```
**What I've tried so far:**
* Changing the `style_preset` (happens in all of them).
* Reducing `steps` from 50 down to 20.
* Adjusting `cfg_scale` from 5 to 10.
* Using completely different, simple prompts ("a red apple on a table").
* Switched from the `turbo` model to the `v1.2` model.
The green blobs persist. It feels like there might be a mismatch in the expected tensor format or a post-processing bug on their end? Or maybe a known issue with certain GPU drivers? I'm on the latest NVIDIA drivers (555).
Has anyone else encountered this and found a workaround? I'm trying to benchmark this against other models, but this is a major blocker. Any insight would be hugely appreciated!
-pipelinepilot
Pipeline Pilot