Troubleshooting#
Symptoms first, then the fix. When nothing here applies, write to
support@maystro.ai with the log file
(logs/maystroai.log under the install folder) and your version (settings
menu, About).
Install and start#
The install fails or stops part way#
Retry with Repair mode (Review step). Then: launcher dev console, uv cache
clean, retry. Then a different install location.
Database connection failed#
The PostgreSQL service is not running, or the host, port, database, user or password differ from what you created. Test with another PostgreSQL client using the same values; fix the service or the values, then run the connection test again.
Activation failed#
Check the key for typos (format MAY-XXXX-XXXX), the system clock, and that a
proxy or firewall is not blocking api.maystro.ai. A key already activated on
another machine must be deactivated there first.
The app opens but shows "warming up" or does not connect#
The backend is still starting; the first start after an install or update takes longer. If it does not settle in a few minutes, check the launcher log for a traceback.
Generation#
Generation is much slower than expected#
Another process is holding GPU memory. Open Task Manager, Performance, GPU, and look at dedicated memory usage before you start a job; voice assistants, local LLM servers and speech tools are common culprits. Close them. A GPU that spills into shared memory runs every step three to four times slower.
Out of memory#
Images: reduce size, drop high resolution fix, remove extra control layers. Video: shorter take, 768-square preset, one reference image, nothing else on the GPU. Training: close everything else; a 32 GB host runs slower but finishes.
The picture ignores the prompt or is cropped strangely (video)#
The reference image does not match the aspect ratio of the preset. Crop it to the preset ratio first. See Reference images and voices.
The face changes between takes#
Normal for reference-only generation. Keep the prompt and re-take with a new seed; for a recurring character, train a character LoRA.
Speech is garbled, the wrong language, or letters are read out#
The line was typed in plain ASCII, or the take is too short for the sentence. Use the proper characters of the language and match the length. See Prompting.
Audio breaks up in the second half of an LTX clip#
The clip is longer than the audio envelope of the model (about 10 seconds). Shorten it or split into two shoots.
The video plays in slow motion#
An upsample preset was used. Pick the standard preset for the same length.
A model import says the model could not be identified#
Set the type and base model by hand in the import dialog. If it still does not load, the architecture is not supported; the error names the reason.
Training#
Training does not start#
Another run is active (one at a time), or the GPU is busy with a generation. Wait for it or cancel it.
The LoRA changes the lighting or motion of every take#
It is used at strength 1.0. Set it back to 0.5 in the scene LoRA list and rebuild the character set.
The LoRA copies dataset frames instead of the face#
Overfitted. Use an earlier checkpoint, or retrain with fewer epochs or more varied frames. See Checkpoints and outputs.