Confession: I'm bad at writing image prompts. I'd spend 20 minutes tweaking "cinematic lighting, 8k, ultra detailed" and still end up with a mug that had three handles.
The other annoying part was picking a model. Flux for this, some video model for that, a different one when I needed text in the image. I had five tabs open and a spreadsheet of which model was good at what. That's not a workflow, that's a hobby.
So with DREAMMMER we went the lazy route: you type what you want like you'd say it to a friend, and an agent writes the real prompt and picks the model. Then you keep chatting to fix things ("make the mug blue", "less shadow").
A few things that ended up mattering more than I expected:
Reference images. Keeping the same character or product across shots was the #1 request, so the reference image generator got a lot of attention.
Not hiding the models. People still want to know what's running, so there's a plain list of models we route between.
Templates for boring jobs. Thumbnails and product shots don't need creativity, they need to be done. That's what the templates are for.
It's still rough in places. Video is slower than I'd like, and the agent sometimes picks a model I wouldn't. New accounts get 50 free credits if you want to poke at it and tell me where it breaks.
Curious how you all deal with model choice. Do you stick to one model, or switch per task?
Top comments (0)