KG Krea 2 Concept Slider Stack V11

The stack is where the sliders happen. Everything else in this pair is bookkeeping: cards describe axes, this node turns them into real conditioning for a Krea 2 sampler. It takes up to eight cards, it needs the Krea Qwen encoder rather than a generic CLIP, and it has one field that will confuse you for ten minutes if nobody explains it.

How it works

You give it a scene prompt and a pile of slider cards. For each distinct axis it builds two pole sentences, tokenizes them alongside your prompt, and encodes. Then it re-encodes with each pole's token span muted, so the difference between the two runs isolates that one axis's contribution and nothing else. Those deltas get re-added at weights derived from your card values - at the default reach, a card at +3 contributes a coefficient of 1 and a card at +6 contributes 2.

Two things fall out of that design. Duplicate and reversed cards are combined before encoding - two "brightness +3" cards are one axis at +3, not a double dose, and a "brightness +3" alongside "brightness -3" cancels to nothing and costs no encoder pass. Timing is per phase: the early and final sampling phases are budgeted separately, so early layout only and final details only cards don't compete for the same capacity.

The fields that matter

Krea CLIP takes the Krea Qwen3-VL encoder. Final image prompt is your actual scene description - the sliders are adjustments relative to whatever you write here, not standalone instructions.

Overall slider reach and Combined slider budget are the volume knobs, and both are manual mode only. Reach multiplies every card's request; the budget caps the sum of absolute coefficients separately for each phase, default 2, which is exactly one full-range axis at reach 1. Zero on either one bypasses pole encoding completely and gives you the plain prompt, pixel-identical.

Slider scaling mode is the field worth reading the tooltip on: "Automatic computes reach and budget from active cards; the two manual values are ignored. Manual preserves saved workflows." In automatic, the stack measures the real load from your active cards, keeps reach at 1 while that load is under 6, and only then starts scaling down and reports the reduction. Two facts to internalize: automatic ignores the manual widgets including their zeros, and it never amplifies small values to fill the budget. If you want a plain-prompt comparison branch, set the cards to zero or switch to manual and set reach to zero.

Early-to-final handoff (default 0.4) is the phase boundary. Reuse slider studies caches the encoder work so you can drag values around without re-paying for it - changing values reuses the study, changing the prompt or the pole text does not. The cache holds at most two entries and is bound to the CLIP's patch revision, so if you modify the model in place outside Comfy's revision protocol, you have to select always re-study.

Outputs

conditioning goes to the sampler's positive input. slider_report is a plain string - wire it to a Preview Any node and actually read it. It tells you the scaling mode, the effective reach and budget, how many encoder passes ran, how many distinct axes survived cancellation, the per-card early and final weights, and every card that was skipped and why. When a slider "does nothing", the report is usually already telling you why.

Install

ComfyUI Manager, search Krea Reference; or:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/kgilper/krea-reference.git

Restart, refresh the browser. No pip dependencies, no extra weights, no training. You still need Krea 2's model, encoder and Qwen image VAE.

Where people get burned

The budget is shared, and that's the whole design. Add a fourth hot slider and the first three get quieter. Budget 2 is a coefficient ceiling, not a safe perceptual maximum, and a small budget is stronger limiting, not automatically better output. Start at ±2 to 4.

Adding a near-zero slider costs extra work. Below 0.25 total axis weight the node blends the plain and steered conditioning branches so the first nonzero step isn't a jump, which means an extra denoiser branch. Zero stays exactly plain.

Version mismatches bite here specifically. Preset and timing come from V11 cards, so a V11 card feeding a V1 stack loses both silently. Older V1 cards work fine in the V11 stack, using whole-image timing. And the automatic scaling mode landed in the GitHub source after the 0.5.0 registry build - if Slider scaling mode isn't in your node, update the pack, don't assume you're in automatic when you aren't.

Wrong encoder type errors loudly. The stack refuses to run on a host without Krea Qwen token-span muting rather than silently degrading.

One scope note so you don't burn an evening: the reference stack and the slider stack are separate encoders. You can't merge their conditioning and call it a combined image-anchor-plus-slider node - the pack says explicitly that it doesn't ship one.