[ANN] Vela: A generative media Proof-of-Inference Network | Testnet is live

[velaresearch]

Vela: A Peer-to-Peer Proof-of-Inference Network

Proof-of-work secures a peer-to-peer electronic cash system without a trusted party, but the work itself, repeated hashing, has no value outside the protocol. We propose a system in which every proof-of-work attempt must first execute a deterministic text-to-image model, so that block production consumes exactly the computation of generating images. The block header commits to the model output, and the seed of the model's initial noise is derived from the header itself, so changing the nonce requires running the model again and no attempt can be reused or precomputed. To let every node verify a block by re-running the same inference, the model is redefined as an integer program whose output is identical, bit for bit, on any CPU or GPU. A block therefore carries a short prompt, a latent image that anyone can decode, and a proof that the image is the honest output of the scheduled model for that header. Verification costs one inference while finding a block costs, on average, many; difficulty is retargeted every block, and model upgrades are scheduled at fixed heights. The long-term goal of Vela is a generative media inference platform: the same inference capacity that secures the ledger serves image, and later video and audio, generation for users who stake coins for access, with every result verifiable by recomputation.



The white paper has been released.
https://velaresearch.ai/vela-whitepaper.pdf


For more updates on our progress, follow us on Twitter.

Website:https://velaresearch.ai/

X:https://x.com/velaresearch_ai

Discord:https://discord.com/invite/VcTU8h3pv7

[_ili_]

i hope it is functional and interesting. i keep an eye on it

[velaresearch]

🔬 Vela testnet is live — Proof-of-Inference v2
Vela is a Bitcoin Core fork where every proof-of-work attempt runs a text-to-image model. The block header commits to the model output, and every node re-verifies a block by recomputing that inference bit for bit. The chain's work product is pictures, not just hashes.

The public test network (testnet4) is up, with Proof-of-Inference v2 active from block 1.

👀 See it
• Chain explorer — every block carries the image its prompt produced: https://explorer.velaresearch.ai
• Docs — build, run, mine: https://velaresearch.ai/docs/

⛏️ Mine it The chain is young, so ASERT still holds the difficulty at its floor: 1/16 per attempt. One mid-range GPU (RTX 3050) finds a block roughly every two minutes. GPU miner: https://git.velaresearch.ai/velaresearch/miner

🖥️ Run a node —  needs an AVX2 CPU, ≥ 4 cores, ≥ 8 GB RAM, ≥ 20 GB disk. Verification is the real cost: about 8 s per block on 10 cores, and every block needs one.

git clone https://git.velaresearch.ai/velaresearch/testnet.git
cd testnet && sudo bash http://setup-node.sh

http://setup-node.sh builds from source (15–30 min). Prebuilt binaries — Linux tarball, Windows zip and installer, with SHA256SUMS — are attached to the release and need no build.
📦 The model file (1.19 GB, hash-checked; a node refuses the wrong one): https://huggingface.co/valereseach/vela-sd-turbo-int-v2

Network: P2P 39733, RPC 39732, 5-minute block spacing, ASERT with a 12 h half-life, developer fund 10 % of each subsidy.

Questions, bugs and feedback . Happy mining 🫡

[velaresearch]

Vela testnet: 789 blocks, each one a text-to-image inference that every node re-runs and verifies bit-for-bit.

↑ Block 789 — the latent is stored on chain and decoded with the sd-turbo VAE. PoI v2 runs the model as an exact integer program.

0 verification failures. http://explorer.velaresearch.ai


https://x.com/velaresearch_ai/status/2108223813385400398?s=20