404—GEN
SN17Generate 3D models for AR, VR, and gaming using AI
Type a text prompt, get a 3D model. 404-GEN is building the infrastructure for spatial computing: on-demand 3D asset generation for AR, VR, XR, gaming, and any application that needs three-dimensional content at scale.
// 3D worlds, generated on demand.
404-GEN is a subnet that generates 3D models from text descriptions. You type "a medieval castle with a drawbridge" and miners produce a 3D asset you can use in a game, AR experience, or virtual world. The subnet handles everything from prompt processing to quality validation.
The simple version: Imagine typing a description of any object and getting back a 3D model you can rotate, place in a scene, or 3D print. That's 404-GEN: text in, 3D model out.
Centralized equivalent: Think Luma AI's Genie or Meshy, but powered by competing miners who each try to produce the best 3D asset for your prompt.
How it works:
- Miners use neural rendering techniques and Gaussian splatting to convert text prompts into 3D assets. Tasks come from organic traffic (API, Discord bot, Blender plugin) and synthetic datasets generated by quantized LLMs (Llama3-8B, Mixtral-11Bx2). Performance is measured over 8-hour evaluation windows.
- Validators deploy MetaCLIP neural networks to evaluate 3D model quality by rendering 16 different camera angles (8 fixed, 8 random) and computing semantic similarity scores against the original prompts. Batch processing with GPU acceleration enables real-time validation at scale.
- The problem it solves: 3D content is the bottleneck for spatial computing. Creating a single high-quality 3D model takes hours or days with traditional tools. AR/VR applications need thousands of assets.
- The opportunity: Apple Vision Pro, Meta Quest, and the broader AR/VR market need 3D content at scale. The market for 3D assets is projected to grow alongside spatial computing adoption.
- The Bittensor advantage: Competitive generation means multiple approaches to the same prompt. Gaussian splatting, neural radiance fields, and other techniques compete simultaneously, surfacing the best method per prompt type.
- Traction signals: Blender plugin for direct integration into 3D workflows. Discord bot for casual generation. API for programmatic access. 2,514 TAO net 7-day inflow. Led by Ben James (CEO). Highest momentum score in this batch at 68.
Category: Image/Video/Audio Generation | Centralized Competitor: Luma AI Genie, Meshy, Rodin Gen-1, Tripo3D
404-GEN targets one of the clearest bottlenecks in the next computing platform shift. If spatial computing takes off (and Apple, Meta, Google, and Snap are all betting it will), the demand for 3D content will dwarf current supply. Traditional 3D modeling can't scale. AI generation can.
Mechanism:
The 16-angle validation is the critical quality control. Validators render each submitted 3D model from 16 camera perspectives (8 fixed positions for consistency, 8 random for robustness) and run MetaCLIP against the original prompt. This catches models that look good from one angle but fall apart from others, a common failure mode in text-to-3D.
Tasks flow from three sources: organic API requests (production usage), Discord bot (community usage), and synthetic prompts generated by Llama3-8B and Mixtral-11Bx2. The synthetic pipeline ensures miners are always being evaluated, even during low-demand periods.
The codebase is small at 39 commits across 6 contributors, with recent work focused on unit testing and pod orchestration. Development velocity is low (1 commit in the last 4 weeks), which is the primary concern.
Market metrics show strong momentum despite the small codebase. At 54,295 TAO market cap, 404-GEN has 1,911 holders. The momentum score of 68 is the highest we've seen in the current uncovered batch. Net 7-day inflow of 2,514 TAO is strong relative to its size. However, Gini of 0.761 and HHI of 0.083 indicate concentrated holdings.
Root proportion of 0.168 confirms organic demand. The chain buy rate of 0.93% against an EMA of 1.21% shows accumulation is steady.
- Low development velocity: 39 total commits with only 1 in the last 4 weeks. For a generative AI subnet, the codebase is thin relative to the ambition.
- Concentrated holdings: Gini of 0.761 is one of the highest in our coverage. HHI of 0.083 confirms significant whale presence.
- Competitive 3D market: Luma, Meshy, and others have raised significant VC funding for text-to-3D. 404-GEN needs to demonstrate quality parity.
- Spatial computing dependency: The value proposition assumes AR/VR adoption accelerates. If the timeline extends, demand for 3D assets grows slower than expected.