Stability.ai – Introducing Stable Video 3D

3D printing & hobby uses

• Several see strong potential for 3D printing: turning single photos into printable models, especially figurines for tabletop games.
• One commenter explains the full 3D printing pipeline (model → slicing → printing → post‑processing) and notes this model tackles only the “model generation” step.
• Speculative “post‑capitalist” workflow: scrape social media trends, auto‑generate toys and ads, and print on demand; acknowledged as dystopian but technically plausible in parts.

Mesh quality and technical limitations

• Strong skepticism about demo quality: shaded/ textured examples can hide low‑detail, blobby geometry.
• Multiple people want to see untextured, wireframe meshes before judging if outputs are printable or game‑ready.
• Good game assets also need UVs, PBR materials, polygon budgets, rigging/skin weights—none of which are addressed in the paper or blog, according to commenters.
• Some note existing research and tools for better mesh generation and texture/normal/depth reconstruction, but integration with this model is unclear.

Inputs, outputs, and capabilities

• Clarification: “single image inputs” means one image; the model guesses occluded geometry based on training data.
• It is unclear if multi‑image input is supported, though people assume it could be extended.
• Debate over whether it truly outputs a usable 3D model or just renders an “orbital” video.
• A referenced video (not the blog) suggests: multi‑view synthesis → NeRF/structure‑from‑motion → mesh via marching cubes. However, commenters struggle to find concrete examples of mesh export in practice.

Hardware and performance

• Official weights are ~9.4 GB; one user reports ~19.5 GB peak VRAM and ~1.5 minutes on an RTX 4090 after tweaking defaults.
• Default scripts can OOM even on 24 GB; reducing simultaneous frames (e.g., decoding_t) helps.
• Broader debate: 24 GB is abundant for gaming but tight for modern AI, with complaints about GPU VRAM limits and vendor segmentation.
• Some suggest using cloud GPUs or the vendor’s API rather than running locally, especially on Macs.

Potential applications

• Suggested uses: quick base meshes for artists, 3D printed miniatures, papercraft templates, architectural concepts (possibly via other models), auto‑rigged characters for animation, interior design and furniture rearrangement, rapid asset capture from real‑world scenes, avant‑garde filmmaking, and possibly adult content.
• One commenter sees massive implications across art, design, engineering, and games; others see it mainly as better tools for niche experimental video.

Data and training considerations

• Concern about limited availability of high‑quality, diverse 3D training data, pushing many methods to learn 3D implicitly from 2D images.
• Ideas floated: build large 3D datasets via commodity scanners/phones, leverage procedural assets (e.g., foliage), tap huge repositories of STL files plus auto‑tagging.
• Question raised whether deep nets could learn 3D directly from video without explicit 3D labels; no consensus.

Concerns, skepticism, and meta discussion

• Some feel current examples look like plastic toys and question generalization to complex real‑world objects.
• Others emphasize the rapid improvement trajectory, comparing to earlier image and video models.
• Brief mention of controversies around the company’s leadership and a quip about the blog post possibly being AI‑written, but no detailed discussion.