I got tired of paying credits to generate 3D models on someone else's servers, so I built a desktop app that runs the whole pipeline on my GPU.
Typed "weathered wooden dungeon crate with iron corner bands and a hinged lid". Four minutes later: this. Nothing touched the cloud.
@88n77n The write-back layer matches what I found building an agent into my 3D app. Model choice mattered less than what it could read back from earlier sessions. Does anything prune or merge the vault as it grows, or do stale notes just age out of the graph?
@Stefan_3D_AI The re-bake step is the whole game. Same finding here: per-object PBR from the generators is the weak map, so I derive roughness from albedo statistics and only trust a normal map baked from the true high poly. Do you re-bake normals too, or just the color map?
Found out my app's STL export was 1000x too small. Every slicer had been silently rescuing it for weeks, so nobody noticed. Including me.
Exports are true millimetres now, measured against the mesh before they ship. Ask for 120mm, you get 120.0mm.
@SternnCapt30363@IamBirdMomma Generous offer. I'm on the other side of this fight though: my pipeline does its own repair pass, weld, normals, watertight check, before anything ships. That two-layer gap is the classic split-shell tell. What's your script built on, trimesh?
@madonna_matt Thank you, that genuinely made my evening. It's close: the buy flow and installer are in final testing. I'll ping you the moment it opens, your kind of build is exactly what I want thrown at it.
@madonna_matt That ratio matches what I see. Nearly all of those bytes are texture, not mesh: 1024 WebP textures got my showcase models to about 1.3 MB each, geometry untouched. Decimation pays off in draw cost more than download. Congrats on shipping the whole thing in under 4 hours.
@SternnCapt30363@IamBirdMomma Curious how the slice goes. AI meshes love hiding non-manifold edges and open seams that only surface at print time, so a quick repair pass before slicing can save the evening. What printer is it going on?
@88n77n The proportion catch is real. Single assets have the same disease: a mesh looks perfect until it lands in engine 4 m tall. Our fix: make the AI commit to real dimensions up front, then QA-check the mesh against them. Voxel deleting that failure mode entirely is a clean trade.
@Stefan_3D_AI The 'inside Blender' detail matters most here. I run headless Blender for retopo and texture bakes in my own pipeline, and keeping meshes clean through automated Blender passes is where all the hidden work lives. An agent refining a whole environment in there is wild.
@TencentHunyuan The 'editable assets, not video' part is the right call. Curious how per-asset topology holds up at world scale: single-asset pipelines already need a retopo and rebake pass to be truly game-ready. Pulling that off across a whole scene at once would be seriously impressive.
The app is called Polyme. One flat monthly sub, generate as much as your GPU can handle, nothing leaves your machine.
All six models are live in 3D here, drag them around: https://t.co/tvVxr4ka7D
Solo build. If you make games or print minis, tell me what you'd throw at it.
I got tired of paying credits to generate 3D models on someone else's servers, so I built a desktop app that runs the whole pipeline on my GPU.
Typed "weathered wooden dungeon crate with iron corner bands and a hinged lid". Four minutes later: this. Nothing touched the cloud.
It runs on the gaming GPU you already have, from a GTX 1660 up to an RTX 5090, auto-tuned to your VRAM.
No cloud queue, no per-model credits, no files leaving your machine. A flat monthly sub while you need it, cancel when you don't. Your models stay yours either way.
Typed two words, "dragon boss", into the 3D asset app I'm building. My gaming GPU did the rest: sculpted, textured, retopologized, QA-scored.
Five minutes, start to finish, nothing sent to the cloud. Solo project. It opens soon.
@irinatoxi Good piece. The tedious off camera part is the real product problem. We went the other way from agent drives Blender: build the pipeline for automation from the start, retopology, UVs, baking and QA as first class steps. Agents are great, deterministic stages are debuggable.