Catalog
UltraShape: image to 3D mesh from one photo
A fine-tuned open model that turns a single image into a 3D mesh, built on Hunyuan3D-2.1 and free to use commercially.
What it is. UltraShape is a fine-tuned image-to-3D model from infinith, built on top of Tencent’s Hunyuan3D-2.1. Hand it a single image and it returns a 3D mesh. It’s released under Apache 2.0, so commercial use is on the table.
Why it matters. Getting a usable mesh from a photo usually means either a subscription to a closed service or a long photogrammetry session with many angles. UltraShape collapses that to one image and one model, which is a meaningful shortcut for concept work, prop reference, or early-stage asset exploration.
What it improved on. Hunyuan3D-2.1 is already a capable base, and this fine-tune is aimed at pushing the mesh quality and single-image reconstruction further than the base model alone. The open weights and permissive license are things the base model also offers, so the real argument for UltraShape is the fine-tuning work on top.
Strengths. Single-image input keeps the workflow simple. Apache 2.0 means you own what you generate and can use it in client work. Being weights-based means you can run it locally or slot it into your own pipeline rather than depending on a third-party API.
Weaknesses. The model is a community fine-tune with limited documentation, so expect some trial and error getting it running. Mesh quality from a single image will always have gaps and guesswork on occluded surfaces, and the output will likely need cleanup before it is production-ready. It also wants a capable GPU, and with thin public benchmarks it is hard to know exactly where it lands relative to closed competitors like Tripo or Meshy.
More Models
New A fine-tuned LoRA on FLUX.2-klein that takes a black-and-white manga page and a reference image and colors the line art to match the palette and style of that reference.
LoRA for coloring manga pages by reference image
JUN 10
New Roblox released open weights for a diffusion model that turns text into 3D meshes, which puts game-ready asset generation in reach for anyone with the hardware to run it.
Roblox's open diffusion model for text-to-3D mesh
MAY 27
Nvidia's Lyra 2.0 takes one photo and hands back a 3D asset, which cuts out a lot of the manual work that usually sits between a reference image and something you can actually use in a scene.