Thank you for supporting both natural language and tags!
I just wanted to share my appreciation for training the model to understand both natural language and tags. In my experience, this makes prompting so much more responsive and user-friendly.
Many other models rely purely on natural language, which often forces you to write long, paragraph style prompts just to steer the generation in the right direction. Being able to simply add a single tag to get a specific frame composition or camera angle saves so much time and effort.
Thank you for this amazing work!
Thank you for the detailed feedback! We're glad to hear you're enjoying the tag-based generation — please enjoy it! :)
I just wanted to share my appreciation for training the model to understand both natural language and tags. In my experience, this makes prompting so much more responsive and user-friendly.
Many other models rely purely on natural language, which often forces you to write long, paragraph style prompts just to steer the generation in the right direction. Being able to simply add a single tag to get a specific frame composition or camera angle saves so much time and effort.
Thank you for this amazing work!
What is the correct way to use tags in this model? "simple background" or "simple_background"? Thank you
I would start with simple background, following the official guide's use of spaces in multiword tags. Its pinned example is:
best quality, highres, hatsune miku, vocaloid, @foo bar, 1girl, blue hair
I checked both strings against the pinned draft compiler: both are accepted and preserved literally, including the underscore. That establishes accepted input, rather than equivalent conditioning. The prompt-encoding path also passes the text to the tokenizer without an underscore-to-space rewrite in that path.
To compare the two forms, keep seed, dimensions, sampler, steps and CFG fixed across each pair, then repeat over several seeds. The documentation's example supports the space form as a starting point; it does not establish that the two spellings produce equivalent results.