Hi andevol, thanks for the detailed report — really appreciate the console log evidence and the structured write-up, that made this easy to track down.
Found it: you had a reasoning model selected, which "thinks" internally before producing its actual answer. That internal reasoning was eating up the whole token budget, so nothing was left for the actual response — the dual-output parser then received an empty string and fell back to the placeholder image. For this app I'd generally recommend an instruct model over a reasoning model — reasoning models also tend to take noticeably longer to generate, on top of this issue. That's fixed now and already live in the current upload, so reasoning models should work correctly too if you prefer to stick with Gemma.
On the image trigger issue: you're right, that one's still open. Right now the app detects image intent based on a simple keyword match, so any message containing the word "image" triggers generation, even in something like "don't generate an image." I haven't fixed that one yet, but it's on my list — I'll follow up here once it's resolved.
Feel free to grab the current update and test the reasoning-model fix in the meantime. Thanks again for taking the time to write this up so thoroughly before even purchasing.
Happy dating!