Subnet 17: 404-GEN
404-GEN is Bittensor Subnet 17 (SN17), operated by 404-Repo, with its codebase published in the 404-GEN repository. The subnet runs a decentralized competition for AI-generated 3D content: miners submit 3D models generated from text prompts, validators evaluate them through pairwise quality duels, and the best-performing miner earns the top share of subnet emissions. The goal is to accelerate text-to-3D generation capable of producing game-ready assets for virtual worlds, games, and AR/VR experiences.
How the Mechanism Works
According to the 404-GEN repository, each competition round opens with a published set of text prompts. Miners generate 3D models from those prompts and upload them for validator evaluation. Validators then run a multi-stage evaluation pipeline: they render each submission to multi-view images, conduct pairwise duels between outputs using a vision-language model as judge, verify winning submissions by regenerating them from the same prompt to confirm reproducibility, and finally compare the regeneration against the original using perceptual similarity scoring. Submissions that cannot be reproduced or that diverge significantly from the verified regeneration are disqualified.
The competition uses a winner-stays-in format with a decaying leader weight. The leading miner’s emissions share decreases slightly with each successful defense, creating ongoing competitive pressure. A new winner starts at the top weight and must defend their position in subsequent rounds. This structure incentivizes continuous improvement rather than locking in any single approach.
All competition state — submissions, renders, duel results, and verification reports — is recorded in the public repository, making every outcome auditable. Weight vectors from validator scoring feed into Yuma Consensus, distributing emissions via Dynamic TAO.
Participating as a Miner
The 404-GEN repository describes miners as building and running 3D generation models that respond to text prompts. Their economic role is to produce 3D outputs that are visually superior to competing submissions as judged by pairwise duel. The subnet supports multiple generation approaches — including Gaussian Splatting, Neural Radiance Fields, and 3D diffusion models — with no restriction on which technique a miner uses. Miners who consistently produce higher-fidelity outputs across a range of prompts earn the most emissions.
Because submissions are verified by regeneration, miners must produce deterministic or near-deterministic outputs from the same seed and prompt rather than cherry-picking results.
Participating as a Validator
Validators on 404-GEN run the evaluation pipeline described in the 404-GEN repository. For each round, they download miner submissions, render them to multi-view images, and run pairwise visual comparisons using a vision-language model to produce a ranked outcome. They then independently verify the top submission by regenerating it from the same inputs and comparing the result using perceptual embedding similarity. Validators submit weights reflecting round outcomes to Yuma Consensus.
On-Chain Identity
404-GEN is registered at netuid 17 on Bittensor with 256 neurons, verifiable via
taostats.io/subnets/17. The subnet owner coldkey is
5E7a4a7QETrNjoZ2SbVSJMaavrQeyTEaHgjBPDUDxBKjUaMZ. The codebase is at
404-Repo/404-gen-subnet and the project site is
404.xyz.
Relationship to Dynamic TAO
Subnet 17 receives emissions through Dynamic TAO, the protocol mechanism that allocates TAO emissions across all subnets based on their relative stake weight. The Introduction to Bittensor describes how Dynamic TAO determines each subnet’s share of network emissions, which then flows to miners and validators within that subnet according to their performance scores. For 404-GEN, this means the total emission pool available for distribution among competing miners depends on SN17’s stake weight relative to other subnets in the network.
Relationship to Yuma Consensus
Subnet 17 uses Yuma Consensus to aggregate validator scores into final miner weights that determine emission distribution. The Yuma Consensus documentation describes how validator weight submissions are processed to produce consensus weights for each miner. In 404-GEN’s context, validators submit weights based on their pairwise duel evaluations and verification checks, and Yuma Consensus converts those potentially divergent validator opinions into a single consensus ranking that determines each miner’s emission share for the round.
Relationship to Multiple Mechanisms
404-GEN has validators run pairwise quality duels on text-to-3D miner submissions. The Glossary and Multiple Incentive Mechanisms Within Subnets docs note that validators must evaluate miners separately for each mechanism.
For readers, this article documents one subnet market. If that netuid runs more than one incentive mechanism, validator scores and weights should be read per mechanism rather than as one combined path.
Development Stage Context
The Introduction to Bittensor describes subnet development as moving from localnet to testnet and then mainnet. For 404-GEN (SN17), that sequence changes how readers should interpret 3D content generation competition examples and pairwise evaluation scoring outcomes.
In localnet, 404-GEN-compatible miners and validators can be developed and tested in an isolated environment. Localnet 3D model generation results and emission outcomes do not represent production subnet performance.
On testnet, 404-GEN-compatible model submission and pairwise evaluation workflows can be exercised in a shared, non-production network. Testnet evaluation scores and validator weights are separate from mainnet subnet state.
On mainnet, 404-GEN (SN17) is the live production subnet where miners submit AI-generated 3D models from text prompts and validators score them via pairwise comparison to determine real Bittensor emissions. The 404-GEN repository is the registered project repository for SN17 on the production network.
The Bittensor Networks reference separates mainnet, testnet, and localnet. A 3D model evaluation result or emission outcome from one environment should not be read as representing production subnet performance in another environment.
Miner and Validator Roles
Subnet 17 operates under the standard Bittensor two-role structure. Miners supply the subnet’s capability and validators evaluate those contributions and set weights. Reward distribution follows Yuma Consensus.
Reader Boundary
Subnet 17 404-GEN should not be read as generic Bittensor subnet documentation, a hosted 3D asset store, or proof that one prompt win guarantees permanent leadership. It names one subnet’s text-to-3D generation competition on netuid 17 (Understanding Subnets, Glossary: Netuid).
Pairwise Duels Rank Visual Quality
The 404-GEN repository evaluates submissions through pairwise comparisons of rendered multi-view images judged by a vision-language model (404-GEN repository).
Round rankings therefore reflect relative visual quality rather than a single absolute score.
Regeneration Verification Filters Non-Reproducible Outputs
Validators regenerate top submissions from the same prompt inputs and compare the result with perceptual similarity scoring (404-GEN repository).
Submissions that cannot be reproduced or diverge from the verified regeneration are disqualified.
Validator Weights Still Flow Through Yuma Consensus
Subnet 17 uses Yuma Consensus to convert validator weight submissions into emission shares each tempo (Yuma Consensus, Emission).