跳到正文
DEV Community· mujin·· 10 天前AI 評分31

把文案交給 LLM:白名單閘門勝過更好的提示詞

Giving the copy to an LLM: a whitelist gate beats a better prompt

AI 導讀

作者分享一套把文案交給 LLM 的流程,主張以白名單閘門把關,勝過單靠改善提示詞。作品以 Stable Diffusion 1.5 生成,131 件作品釘選於鏈上,每個錢包限領 5 件,可在 Base 免費鑄造;每件預載成本約 0.0000008 ETH。

正文

mujin

Giving the copy to an LLM: a whitelist gate beats a better prompt

The short version of a longer build. Stable Diffusion 1.5 in fp32 with model offload on a 4GB GTX 1650, about 150 seconds per 512x768 image. The collection is 131 pieces pinned on-chain, 5 per wallet, free to mint on Base; preload costs about 0.0000008 ETH per piece, and the live minted count is readable from the contract.

What the pipeline does

One process owns the wallet and the queue: generate, pin, preload, list. Nothing retries blindly — a failure leaves the item in its previous status, so the next cycle picks up the same work instead of inventing new work. The only shared state is a sqlite file plus a heartbeat that the loop rewrites every cycle.

import torch
print(torch.cuda.get_device_name(0), torch.cuda.get_device_properties(0).total_memory)

What it costs, measured

Costs come off the receipts, not off a gas widget: gasUsed times effectiveGasPrice per transaction, logged per piece. At the current numbers the preload is about 0.0000008 ETH per piece and minting is free, which is the only pair of numbers this post will claim.

What I would copy from it

The parts worth stealing are unglamorous: a length gate before anything is published, a heartbeat written by the loop rather than by the supervisor, and an inventory invariant (remaining is preloaded minus totalSupply) that a crash cannot violate.

Takeaway

You do not need rented compute to start. You need a seed you can reproduce, a cost number you measured yourself, and a test suite you actually run before the wallet moves.

Live collection: https://thunderous-sunflower-a910b4.netlify.app

來源:DEV Community · dev.to