# Into: Leoma Leoma turns Bittensor into a tournament for AI video generators: every round, miners submit short clips, an automated benchmark scores them, and only the single best one gets paid. // Winner-take-all AI video on Bittensor --- > New to Bittensor? Start here. Experienced users can skip to the full analysis. ### What is Leoma? Leoma (SN99) is a Bittensor subnet for AI video generation. Independent operators ("miners") run models that turn a starting image plus a text prompt into a short video clip, and a separate set of participants ("validators") score the results. Each round, the best clip wins and that miner takes the full reward. **The simple version:** It is like Runway or Pika, but instead of one company's model, many models compete head to head and only the round's winner gets paid. **Centralized equivalent:** Runway, Pika, Google Veo, and OpenAI Sora: hosted services where you send a prompt and get a video back. Leoma decentralizes the model layer behind that kind of product. **How it works:** - **Miners** upload a Text-Image-to-Video model to Hugging Face, deploy it on Chutes, and commit it on-chain. Given a supplied first frame and a text prompt, they return a short video. - **Validators** run an automated evaluator that scores each video on first-frame fidelity, prompt adherence, temporal quality, and visual artifacts, then set winner-take-all weights on-chain. --- ### Why This Matters - **The problem it solves:** State-of-the-art video generation is concentrated in a few closed, hosted models. Leoma builds an open market where any model builder can enter and is judged purely on output quality. - **The opportunity:** Image-to-video and text-to-video are among the fastest-moving corners of generative AI. A permissionless arena where the best model wins each round is a different shape from the subscription products that dominate today. - **The Bittensor advantage:** Winner-take-all scoring puts direct pressure on shipping the best model, not the best-marketed one. An automated benchmark, rather than vendor claims, decides who wins. - **Traction signals:** Early. The public repository shows active development, with the most recent commit in late May 2026 and a live site at leoma.ai. There is little public social footprint so far, and the subnet currently receives no emission share (covered in the risks below). --- ## Full Analysis **Category:** Image/Video/Audio Generation | **Centralized Competitor:** Runway, Pika, Google Veo, OpenAI Sora AI video generation moved from research demos to consumer products in under two years, and almost all of that progress sits inside closed models you rent by the month. Leoma's bet is that the model layer underneath can be made open and competitive: let anyone register a generator, score every output against the same benchmark, and pay only the best. **Mechanism:** According to the repository README, Leoma currently supports one task type, Text-Image to Video (TI2V): a validator supplies a first frame plus a prompt, and miners return a short video. The round runs in three roles. The subnet owner runs an "owner-sampler" that builds tasks from short source clips, calls miners through Chutes (Bittensor's compute subnet), and posts task artifacts to object storage. Validators run an evaluator that pulls the latest task, scores each generated video with a multimodal model, and posts results back to the Leoma API. The API computes a ranking and exposes a weights endpoint; each epoch, validators read it and set winner-take-all weights on-chain, so the top miner takes full weight for the round. One detail worth flagging for anyone reading the code: the README describes the evaluator using OpenAI's GPT-4o, while the most recent commits move it to Google's Gemini (the May 2026 commits pin a Gemini model and streamline the evaluator to require only Gemini). The scoring structure is the same either way; the underlying vision model is mid-transition. The stack is Python, with miner models hosted on Hugging Face, inference served through Chutes, and S3-compatible object storage (Cloudflare R2 by default, or Hippius). Development is real but concentrated: the public repository's commit history shows a single primary contributor and roughly three dozen commits since the repo went up in March 2026. The roadmap in the README lists Text-to-Video and Image-to-Video as planned additions beyond today's TI2V-only support. On the market side, the numbers are modest. Leoma's alpha trades around 0.00393 TAO with a pool holding roughly 1,460 TAO in reserve. Its current emission share is 0%. Under Taoflow, the flow-based model that has set Bittensor emissions since November 2025, a subnet earns a share of network emissions from its net TAO staking inflows, smoothed over time. Leoma's net flows have been negative over the past week, which is why its emission share currently sits at zero rather than reflecting anything about the code itself. --- ### Risk Factors - **Deregistration:** Leoma registered in late January 2026, so its four-month network immunity has now lapsed. Bittensor automatically deregisters the lowest-EMA-price non-immune subnet roughly every two days when the network is full. With a 0% emission share and a thin pool, Leoma sits in the exposed range until staking inflows turn positive. - **Single-contributor development:** The public repository's history shows one primary author. The work is recent and active, but it carries key-person concentration risk. - **Evaluator dependence on a closed model:** Scoring relies on a third-party multimodal model, and the repo is mid-transition from GPT-4o to Gemini. Evaluation quality, cost, and availability track an external vendor the subnet does not control. - **Competition:** The category is crowded on both sides, against well-funded centralized products (Sora, Veo, Runway) and any other generative-media effort on Bittensor. - **Early traction:** Public usage and social signal are limited so far, so there is little independent evidence of demand yet. --- Another subnet, unpacked. Into the next one.