Compos3D: Interactive Part-Based Composition for Creative Control in Generative 3D Models
Compos3D is an interactive system that enhances creative control in generative 3D modeling by enabling users to assemble coherent designs through a remixing workflow of selected parts from multiple candidates, rather than relying on unpredictable repeated regenerations.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you're trying to bake the perfect cake using a magical oven that only listens to your voice. You say, "Make me a chocolate cake with dragon wings!" The oven whirs, and poof—out comes a cake. But the wings are on the wrong side, the chocolate is too dark, and the base is missing. In the old way of doing things, you'd have to say, "Try again!" and hope the next cake is better. If you like the chocolate but hate the wings, you're stuck. You can't just swap the wings for a new pair; you have to gamble on the whole cake again. This "roll the dice" method is how most current 3D AI tools work, and the paper argues it's frustrating because you lose the good parts every time you try to fix the bad ones.
Enter Compos3D, a new system that changes the game. Instead of treating the AI's output as a finished product you either keep or trash, Compos3D treats it like a giant box of LEGO bricks.
Here's how it works:
- The Bake-Off: You ask the AI to make several different 3D models (like chairs or monsters) based on your idea.
- The Scavenger Hunt: You look at all the results and pick the specific parts you love. Maybe you want the seat from Chair A, the backrest from Chair B, and the legs from Chair C. You can grab these parts by clicking on them in a 2D picture or by clicking directly on the 3D model itself.
- The Remix: You drag and drop these favorite pieces onto a canvas, arranging them into a rough, messy collage. It doesn't have to be perfect yet; it's just your vision of how the parts should fit together.
- The Magic Glue: The system then takes your rough collage and "synthesizes" it. It acts like a super-smart glue that fuses your separate pieces into one smooth, coherent 3D model, fixing the gaps and making sure the geometry looks real.
The authors tested this idea with eight people who were new to 3D modeling. They asked these users to design a hybrid work chair with very specific requirements (like a tall backrest and dragon-leg armrests). The users tried two methods: the old "roll the dice" regeneration method and the new "remixing" method.
The results were clear. The users felt much more in control with the remixing method. They reported that the final designs matched their ideas much better (with an average score of 6.12 out of 7 for remixing, compared to 4.12 for regeneration). They also felt the tool responded better to their wishes. However, there was a catch: remixing took a bit more mental effort. Users rated the effort slightly higher for remixing (3.50) than for regeneration (2.94). But here's the kicker: even though it was harder work, the users found it more enjoyable (scoring 5.75 for remixing vs. 4.81 for regeneration). It seems that having the power to pick and choose your own parts made the extra effort worth it.
The study also looked at two ways to do the remixing: using 2D images (like a collage) or 3D models (like moving actual blocks in space). The 3D way felt more intuitive for moving things around in space, while the 2D way was faster for picking pieces. Most people preferred the 3D remixing approach for its precision, though it required the most effort of all.
The paper suggests that this "part-based" approach is a big step forward because it lets users express their intent by building rather than just describing. One user noted that with the old method, they felt like they were just "typing something and seeing what happens," but with remixing, they felt like a real designer making choices.
The authors are careful to note that this isn't a magic wand that solves everything. They only tested it on eight people with one specific design task (a chair), and they didn't check if the final models could actually be printed or built in the real world. But the findings strongly suggest that giving people the ability to mix and match parts from different AI generations leads to happier, more creative designers who feel like they are truly in charge of their creations. Instead of hoping the AI gets it right on the first try, you get to curate the best bits and let the AI do the heavy lifting of putting them together.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.