KG Krea 2 Concept Slider Card V1

This is the smallest node in the pack and also the one you'll touch the most: it's where you actually write a slider into existence. Four fields, no images, no encoders - just a description of an attribute and a value from -6 to +6. Everything downstream (the study, the encoder pass, the push math) happens in the Concept Slider Stack node this card plugs into. This card is purely the "what should this dial change, and how far is it turned" declaration.

What it does and why it's not a LoRA

The classic way to make an adjustable attribute - an age slider, a brightness slider - is training a dedicated LoRA on paired examples ("young woman" vs "elderly woman") and controlling its strength as the dial. This card skips training entirely. Name the attribute, and the stack derives the more-vs-less direction straight from Krea 2's own text encoder at encode time, no weights downloaded, no dataset collected. That's the whole pitch, and it's the reason this node's only real inputs are text.

What you fill in

• What this slider changes - the attribute name, defaulted to brightness. A noun or short noun phrase works best - age, fog density, height. This is the label the encoder anchors its axis to, so vague adjectives underperform concrete nouns.
• Slider value - the dial itself, -6 to +6, default 0. Negative pushes toward less or the opposite; positive pushes toward more. 0 is a genuine no-op - an exact-zero slider gets excluded from the encode entirely, so a parked card sitting at 0 costs nothing to leave connected.
• What +6 looks like (optional) and What -6 looks like (optional) - custom pole sentences. Leave both blank and the stack infers the poles from the attribute name alone; that works fine for straightforward physical attributes. Write your own pair and you can build an axis out of anything you can describe in contrast - not just physical properties but style axes like photoreal-vs-cartoon, or something as specific as a "golden-hour light" axis that grades a scene from cool overcast to full golden hour on the same seed.

Output is a single slider of type KG_KREA_SLIDER, which only plugs into a Concept Slider Stack's numbered slider inputs - it doesn't connect to anything else in ComfyUI.

Writing a slider that actually works

The pack's own render audit is candid about the failure mode: of ten sliders tested, eight worked as intended straight off the default pole inference, and two needed the poles rewritten by hand before they held. That's the practical lesson - if a slider is weak, muddy, or drifting off-target, the fix is almost never the value, it's the wording. Reach for the custom pole fields before you assume the technique doesn't work for your attribute. A short, concrete, contrastive pair of sentences is what gives the encoder something to actually grab onto; single adjectives with no context are the most common source of a slider that "sort of" does something and nothing precise.

Stack up to eight of these on one Concept Slider Stack, and keep the working range in mind: the reliable band is roughly -3 to +4 per card. Push further and you're more likely to break the composition than intensify the effect - the stack's slider_report output will tell you exactly what push each card computed if you're not sure what's happening.

Installing it

Part of the same kgilper/krea-reference pack as every other node here - ComfyUI Manager (search "ComfyUI Krea Reference") or cd ComfyUI/custom_nodes && git clone https://github.com/kgilper/krea-reference.git and restart ComfyUI. MIT license, no model weights bundled. You need a Krea 2 checkpoint with its Qwen3-VL text encoder and Qwen-Image VAE already loaded - this card is inert without a Concept Slider Stack and a real Krea 2 CLIP feeding it, and it does nothing at all on non-Krea-2 pipelines since the whole mechanism rides on that specific encoder.