What We Know So Far About Eclair (Subnet 28 on Bittensor)

Read Time:1 Minute, 33 Second

Eclair is Subnet 28 on the Bittensor, focused on image-to-video (I2V) generation and evaluation. It positions itself as an open AI video project, often described as “Bittensor for AI generation,” with the goal of creating markets for intelligence rather than monopolies.

Table of Contents

Who Built It

Eclair was reportedly built by Const and Fish, and it leverages Subnet 75 (Hippius) for decentralized storage. When asked if he’s involved in the project, Const said he is only helping the team and SN28 isn’t his subnet:

How Eclair Works

Miners

Miners run image-to-video generators and commit their generator chute slug on-chain.
They receive inputs such as:

Prompt
Base64-encoded image used as the first frame
FPS, number of frames, and resolution
A fast flag

The output is a generated MP4 video.

Validators

Validators sample real video clips, extract the first frame, and generate a textual description using GPT-4o.
Miners are then challenged to recreate the video from that information. All samples are stored in Hippius S3 buckets.

Scoring

Validation relies on forced-choice comparisons using GPT-4o, where the model decides which video looks more real: the original or the miner’s output. Scoring follows a winner-take-all system, with an epsilon beat rule determining weight assignments.

Technical Snapshot

Languages: Mostly TypeScript (74.4%) and Python (21.8%)
Requirements: Python 3.12+, ffmpeg, OpenAI API key (GPT-4o), Chutes API key, and a Hippius subaccount seed phrase
License: MIT

Closing

Eclair is shaping up as a serious attempt to standardize and evaluate open image-to-video generation inside the Bittensor ecosystem. It combines market incentives, open-source tooling, and real-world video benchmarks in a way that fits Bittensor’s broader vision.

We will share more details as the subnet evolves and more data becomes available.