Helm: training-free activation steering for ComfyUI packs model concepts into ~22KB steerable axes
leolambertini · reddit · 2026-10-20
Helm is an open-source (MIT) ComfyUI plugin, comfyui-helm 0.3.9 in the Comfy Registry, doing training-free activation steering: it reads a direction the model already knows from two contrast passes, saves it as a 22KB .safetensors axis, and injects it with one strength value—no training, weights untouched.
The framing: a prompt is a switch, not a dial; the usual fix for "a little more X" is a LoRA. Honest evaluation under a pre-registered criterion: only 10 of 53 labelled directions hold (7/53 with a post-hoc edge/saturation ceiling), many wash out, and masks don't contain the change. Measured on MiniMax H3 (video) and Z-Image; repo includes labels, graphs, hashes, refuted hypotheses, and a render archive.
More from Multimodal
- SwitchSplat brings switchable object states to 3D Gaussian Splatting, headed to SIGGRAPH Asia 2026 — taiyasaki · 2026-10-20
- FastVideo releases FastH3 OmniRef for ComfyUI on Hugging Face — Lost-Dot-9916 · 2026-10-20
- Qwen 2.1 Shows Worse Subject-Scene Integration Than AIO v23 in Multi-Image Edit Tests — ShowerFluid9178 · 2026-10-20
- Edited a video entirely with AI: best half-day output yet, and it's 'scary' — aronkor · 2026-10-20
- Stanford's Halcyon: strong text-to-image without language pretraining — StanfordUniversity · 2026-10-20
- From one sentence to a finished trailer: a complete beginner's guide to AI filmmaking on a laptop — tetsuoai · 2026-10-20