KG Krea 2 Concept Slider Stack V1

Concept sliders - the idea of an axis you can drag from "less of this" to "more of this" without retraining anything - aren't new; there's a whole academic line of work (sliders.baulab.info) built on training a dedicated LoRA per concept and adjusting its weight like a knob. This node does the same trick with a shortcut: no LoRA, no training run, no dataset. You just name an attribute, drag a value from -6 to +6, and it derives the more-vs-less direction straight out of Krea 2's own text encoder at the moment it encodes your prompt. The README's own framing is honest about what that buys you: "Slider LoRAs without the LoRA."

What it replaces and how it fits

This node is a drop-in swap for a CLIP Text Encode node. It takes your Krea CLIP, your Final image prompt, and up to eight KG Krea 2 Concept Slider Card V1 nodes wired into the Slider 1–Slider 8 optional inputs, and it outputs CONDITIONING straight into your KSampler's positive input - same slot a normal text encode would fill. Each slider card just carries one attribute (name plus a -6..+6 dial); the stack is what actually turns that description into a push on the encoder's output. If you're not running any sliders, the stack still just encodes your prompt normally - a slider left at exactly 0 costs nothing and is excluded from the encode entirely, render-proven pixel-identical to a plain encode.

What you set

• Krea CLIP - the same Krea 2 text encoder you'd wire into any text encode node.
• Final image prompt - your actual prompt, multiline.
• Overall slider reach - one multiplier over every connected slider's push at once (0–3, default 1). Turn the whole stack up or down without touching individual cards.
• Reuse slider studies - reuse between runs - faster tuning (the default, and what you want while iterating) vs always re-study. Cached studies mean tweaking one slider's value doesn't force every other slider to re-run its encoder pass.
• Slider 1–8 - plug in Concept Slider Card outputs here. Optional, and you don't need all eight; two or three targeted sliders beat eight fighting ones.

Two outputs come back: conditioning, which goes to your sampler, and slider_report - a plain-text string documenting each slider's exact pole sentences, the push it actually computed, anything it skipped and why, and how many encoder passes vs cache hits it used. Wire that into a Show Text node when a slider isn't behaving the way you expect; it's the fastest way to see whether the problem is the pole wording or the value you dragged.

Installing it

Same pack, same install as everything else in Krea Reference: ComfyUI Manager, search "ComfyUI Krea Reference," or cd ComfyUI/custom_nodes && git clone https://github.com/kgilper/krea-reference.git and restart. MIT license, no bundled weights. You need your own Krea 2 model, the Qwen3-VL text encoder, and the Qwen-Image VAE already set up - this is conditioning tooling on top of a model you already have running, not a model itself.

Where it actually works, and where it doesn't

The reliable working band is -3 to +4 on the individual card's dial - the README is explicit that going further starts to strain. +/-3..4 is where most attributes land cleanly; push toward the extremes and you're more likely to break the image than intensify the attribute. If a slider does nothing at all, check its value first - 0 is a deliberate no-op, not a bug.

The other place people get burned is picking a bad name for the attribute. brightness, age, fog density - nouns and short phrases the encoder can actually anchor to - work far better than vague adjectives. If a slider is weak or drifting off-target, that's what the optional custom-pole fields on the card (what -6 and +6 should each look like) are for; a spelled-out pair of poles fixes most weak axes without touching this stack node at all. This node just runs the stack - if a slider is wrong, the fix almost always lives one node upstream, on the card itself.