{"author":{"username":"allyourbase","displayName":"allyourbase","fid":4716},"timestamp":"2026-10-29T13:52:45.000Z","text":"In spirit of the network school decided to share here a small research project\n\nit's called helm\n\nyou feed it two prompts that contradict each other. it captures what changes between them, inside a single transformer block, and saves that as a tiny file — about 22 KB. i call it an \"axis\". then, when you sample, you dial it in with one number. 0 means the model is completely untouched, byte for byte.\n\nit works on minimax h3 for video and z-image for images. no training run, no lora, the weights never move.\n\nthe reason it exists: prompts can't do \"a little\". change one word and the whole frame goes. i kept wanting a dial instead of a switch.\n\nit's MIT licensed and it's on the comfy registry now.\ngithub.com/U-N-B-R-A-N-D-E-D/comfyui-helm","links":[{"url":"https://stream.farcaster.xyz/v1/video/01a187ee-638d-7b69-1199-1d1b305c78fa.m3u8"},{"url":"https://github.com/U-N-B-R-A-N-D-E-D/comfyui-helm","title":"GitHub - U-N-B-R-A-N-D-E-D/comfyui-helm: Training-free activation steering for ComfyUI: read a transformer block's direction between two prompts and inject it with a single strength scalar. Open weights only."}],"farcaster":"https://farcaster.xyz/allyourbase/0x1b518a6d","firefly":"https://firefly.social/post/farcaster/0x1b518a6dd2e2ea20ee5a6094123b38c40db67acd"}