Accept image references (and video!)

ned.t · 1mo ago

I find it way easier to iterate on a design with an image model. Once I've got something good, the image model is fantastic at generating a 3-view. I wish I could supply that as the input rather than text.

Follow-on to that: use image model to put mocap suit onto model, then use video model to generate motion, then use generated video to make animation

https://lystrosaurus-motion-study.pages.dev/
https://x.com/shinboson/status/2096390580594995519

2 replies

lucas.m Polyfork · 1mo ago

Great Idea, will implement!

lucas.m Polyfork · 1mo ago

Implemented the image reference part in /steer

You can attach up to three pictures beside the prompt box, and you do not have to type anything at all: we read the kit out of your reference and plan from that. The looks come back drawn toward your picture rather than toward a paragraph about it.
Your three-view gets the better half. Every row of the parts list now takes its own picture. Attach one to a row and that part is built from your drawing, same shape, same proportions, same details, redrawn in our style so the build can be held to it. A turnaround beats a single view there, because it settles the sides we would otherwise be guessing at.

Sign in to reply. Voting needs no account.