This version is highly experimental, and the text generation is still far from perfect. I tested many different configurations, and the best results came from using the LCM sampler with the Simple scheduler.

You can also try increasing the resolution to 2048 × 1365, 3072 × 2048, or 4096 × 2736. The default resolution is 1536 × 1024, but higher resolutions can improve text quality slightly. Even so, don't expect perfect results yet.

Consistency is not 100%, although it tends to be a bit better when using LCM. Overall, I think this was a successful experiment, and I'll continue improving it in future updates.

The prompt must follow the specific format used in the workflow (). You can write it manually if you want, but using the included VLM to generate the prompt usually produces better and more consistent results, especially for maintaining the correct structure.

If you discover a better way to improve the text generation, feel free to leave a comment and share your findings.