Quick status update. We rolled back the prompt parsing changes from the last update due to stability issues—the image generation tips are still valid, but the parsing changes are gone.
Here’s the background: To get full checkpoint coverage, we did a major overhaul of the GPU scheduling code on the Worker side, and also upgraded ComfyUI and CUDA. Both affect the style and quality of generated images. After going live, we fixed a bunch of generator stability and image quality issues one by one. In the process, we originally planned to swap out the A1111-style prompt parsing for ComfyUI’s approach—A1111 normalizes weights by averaging them across the entire prompt, so people got used to cranking up exaggerated weights to get the effect they want; ComfyUI handles each token individually, which is more precise and closer to the original intent.
But the switch triggered stability problems, so after weighing the pros and cons, we rolled it back. If your recent images look off, first follow the prompt guide and dial your weights back to normal usage. If it’s still not right, open a ticket with the Job ID—you can copy it from that hexagon i icon attached to the image.