Different image generation types

[mistval]

Currently we just scaffold one type of image generation: generating images of a character.

There should also be at least "scenery", maybe one or two others. These should preferably be selectable easily via the image gen button (maybe long-press?)

Auto image should also be configurable to set rates for each type of image generation.

[brg310]

One use case I'd love to see as part of this is true multi-character scene generation. Right now, chat images seem to work primarily as images of the focused character: the focused NPC's appearance/wardrobe is included in the prompt, while other characters in the scene are only mentioned through the generated scene description.

This means recurring characters other than the focused NPC don't have consistent appearances in the generated image. It also looks like the focused character's configured wardrobe is always added to the final prompt, so the image can keep them in their default clothes even when the current scene establishes that they've changed or removed them.

For a “scene” image type, it would be great if the image prompt could include the appearance data for all characters actually depicted, with configured wardrobes treated as defaults that can be overridden by the current scene state.

Really enjoying the image integration otherwise - being able to generate an actual snapshot of the current scene would be awesome.

[mistval]

One use case I'd love to see as part of this is true multi-character scene generation. Right now, chat images seem to work primarily as images of the focused character: the focused NPC's appearance/wardrobe is included in the prompt, while other characters in the scene are only mentioned through the generated scene description.

This means recurring characters other than the focused NPC don't have consistent appearances in the generated image. It also looks like the focused character's configured wardrobe is always added to the final prompt, so the image can keep them in their default clothes even when the current scene establishes that they've changed or removed them.

For a “scene” image type, it would be great if the image prompt could include the appearance data for all characters actually depicted, with configured wardrobes treated as defaults that can be overridden by the current scene state.

Really enjoying the image integration otherwise - being able to generate an actual snapshot of the current scene would be awesome.

Hi, yes you're basically right on all points about how it works, and thanks for the suggestions.

I'm thinking here's what we need:

1. Turn the final image prompt concatenation step into an editable template (so if you don't want wardrobes to be injected deterministically, you can edit that out and just rely on the model to output clothing and such).
2. A way to add/delete more image generation prompt templates, as many as you like.
3. Default to providing maybe three or four, including the one we already have, plus "scenery only", and "all characters".
4. UI to choose which type of image generation to run when you generate an image.

Another possible alternative, when I add an agentic loop and tool calling, is that the model will be able to call tools to enable/disable/edit wardrobes.