Diffusion has been a lot of fun. I’m really enjoying the rapid iteration of running through 50 ideas for a kitchen design in minutes, then incorporating the individual favorite elements of each back into the main model.
I haven’t read all of the previous commentary so there’s a good chance most of it has been covered, but there are a few things that can be improved to get the user to an end state, and not just loop endlessly while trying to figure out the language model when working on interiors.
- Provide source images - This is similar to the style section, but I’d love to be able to provide a user-based source image(s) that can be used as a basis for the output. Having weights on influence of the source, like which already exists for the prompt would be nice, too.
- Render only option - I’ve explored a few options to get diffusion to photo-render what was already in place, to no avail.
- Tags - This is sort of a bridge between true automation versus guided mode, but I’d love to be able to tag components/groups from the model in the text prompt, like tag:chair&blue, tag:chair&remove. There obviously exists some keywords that the devs created, like floors, and this would be along the same vein.
- Update the 3d model - a 2025 goal, but would be game breaking to be able to update the 3d model with suggestions from the diffusion output.
Improve:
the language detector. Inputting something like “blue door” seems to be parsed as 2 independent items, blue and door, so you sometimes end up with random blue elements and a door in the middle of the kitchen. Would be good to be able to use quotes or something around the object/descriptor you want to combine, e.g. “blue door”