[ U N B R A N D E D ]/R&D
HELM / REV 0.4.0
REGISTRY ↗ GITHUB ↗ ← LAB
UNBRANDED / R&D / HELM EST. 2022 / CDMX

HELM

One scalar.
Measured.

Activation steering for image and video models. Pick an idea, move one slider. Start with Z-Image; H3 is the video path. Built and run on KAFEIN hardware.

Measured78directions
Labelled53axes
Held10/53safe range
Strict7/53ceiling

START WITH THE DIAL ↗

00 / Film

The measured cut

FIG. 00 — intro + Z-Image set + H3 density + 35 axes / no crop OPEN FULL CUT / 50 AXES ↗
01 / Index

Curated axes

14 clips · SCROLL →
Contact sheet of 14 curated Helm clips
FIG. 01 — 14 curated clips / ping-pong / 0.25 s per frame / 12 s
camera motion intensity OPEN ↗
01CAMERA

camera motion intensity

1.3041 → 5.21663 LVL
corridor orientation OPEN ↗
02GEOMETRY

corridor orientation

2.3883 → 9.55323 LVL
figure ground separation OPEN ↗
03CAMERA

figure ground separation

0.4024 → 1.60973 LVL
fog density OPEN ↗
04ATMOSPHERE

fog density

0.179 → 0.71593 LVL
framing angle OPEN ↗
05CAMERA

framing angle

1.6899 → 6.75943 LVL
motion blur OPEN ↗
06CAMERA

motion blur

0.1228 → 0.49133 LVL
motion blur intensity OPEN ↗
07CAMERA

motion blur intensity

1.7244 → 6.89773 LVL
overcast vs direct OPEN ↗
08ATMOSPHERE

overcast vs direct

0.5983 → 2.39313 LVL
raking light angle OPEN ↗
09LIGHT

raking light angle

0.7016 → 2.80643 LVL
spatial depth OPEN ↗
10CAMERA

spatial depth

0.3229 → 1.29153 LVL
subject scale OPEN ↗
11CAMERA

subject scale

1.5055 → 6.02213 LVL
sun direction OPEN ↗
12LIGHT

sun direction

1.5301 → 6.12043 LVL
telephoto compression OPEN ↗
13CAMERA

telephoto compression

0.9569 → 3.82743 LVL
volumetric beams OPEN ↗
14LIGHT

volumetric beams

0.5025 → 2.01013 LVL
02 / Families

Five-level studies

0 → 0.5
03 / Method

Measured, not prompted.

v0.4.0 · ACTIVE

HELM is an activation steering surface. One scalar moves one measured direction. No LoRA. No prompt pack. 0.4 adds Helm · Dial: pick an idea, move Amount. Three flat starters — Z-Image, H3 video, make your own — put the prompt, the Dial and the result on the canvas.

The index is the visual set: 14 curated clips, each with a range, a level count and a one-take video. The measured claim is separate: 10 of 53 labelled axes held the safe range, and 7 passed the strict ceiling.

$ comfy node install comfyui-helm@0.4.0
> registry.comfy.org/nodes/comfyui-helm
> status: Active

$ helm.directions --measured 78 --labelled 53 --held 10 --families 3
> one scalar / one direction / one result
> source: github.com/U-N-B-R-A-N-D-E-D/comfyui-helm
385renders
10/53held the safe range
7/53strict ceiling
43washed out

The 14 clips are a curated visual set, not a claim that every axis survives the measured criterion. The film is a separate cut; the measured table is the source of truth.

Open the lab.

↗