KG Krea 2 Image Guide Card V10

If you've ever dropped a reference image into a workflow and watched it half-drag the subject, half-drag the lighting, half-drag nothing you wanted - that's the problem this node exists to fix. A guide card doesn't let an image vaguely influence the result. It makes you answer two questions before the image gets anywhere near the sampler: what should Krea 2 actually borrow from this picture, and how hard should it lean on it. One card, one image, one job.

This is the V10 half of a two-node system. The card itself doesn't do any encoding - it's a config packet. You wire its guide_card output into a KG Krea 2 Reference Stack Encoder V10, which is the node that actually runs your images and prompt through Krea 2's text encoder and produces the conditioning. Stack up to twelve cards on one encoder and each one owns a different job: one card keeps your product recognizable, another borrows the lighting from a mood board, a third steals a material without touching the subject, a fourth exists purely to stop the model from copying a logo. That's the whole design idea - multi-reference workflows stop being a guessing game once every image has a stated purpose.

Worth knowing before you dig into the knobs: this isn't IP-Adapter. It doesn't bolt on a separate adapter model or inject CLIP image embeddings through extra cross-attention layers the way IP-Adapter does for SD 1.5 and SDXL. Krea 2's text encoder is Qwen3-VL - a vision-language model, not a plain text encoder - and the Krea CLIP input on the stack encoder is that same object. The guide card's job is to prep your image (crop it, wash the color out of it, decide how much of it "counts") and hand it, plus a pile of directives, to Krea's own native encoder alongside your written prompt. It's closer to using the model's own eyes than strapping a second model onto it.

What you actually touch

Everything funnels through a handful of fields; the rest is fine-tuning you can ignore on your first pass.

• Reference image - the image this card is about.
• Use image for - the recipe. This is 90% of the node. Pick keep the same subject, suggest the visual style, copy lighting and mood, copy pose and layout, avoid copying text/logos, and so on - 33 choices total in V10, since it adds palette, background, camera-framing and mood-board recipes on top of V9's set, plus your own custom recipes if you drop schema-validated YAML/JSON into the pack's custom_recipes/ folder. manual tuning unlocks every raw dial (Manual mode borrows, layer pulls, etc.) if a preset doesn't fit.
• How strongly this image guides - 0 to 3, default 0.2. The README's own table is the cheat sheet: 0.03–0.08 for a tiny nudge (text/logo guard), 0.55–0.9 for style/material/lighting (many recipes cap themselves around 0.65), 0.9–1.2 for subject/pose anchors.

V10 adds four controls V9 doesn't have, and they're the actual reason to use V10 over V9: Guide direction, which lets a card push away from an image instead of toward it (a counter-example - "not this palette"); When this card guides, per-card timing (whole image vs early layout only vs final details only); and Structure layers pull / Finish layers pull, manual dials over the structure-vs-finish conditioning split. Everything else - Prepare image by, Color kept, Small details kept, Study this image at, Frame this reference by, Subject copying, Early layout guidance, Final detail copying, Maximum image pull, Shape copied, Overall style reach - exists on both versions and is what a recipe sets for you under the hood; leave them on defaults until a specific card is misbehaving.

Output is a single guide_card of type KG_KREA_REFERENCE - a custom type that only a Krea Reference Stack Encoder understands. It won't plug into a plain CONDITIONING slot.

Installing it

ComfyUI Manager: search "ComfyUI Krea Reference." Manual: cd ComfyUI/custom_nodes && git clone https://github.com/kgilper/krea-reference.git, then restart ComfyUI. That's it - MIT licensed, no submodules, and it ships with zero model weights. It's pure conditioning logic; you still need your own Krea 2 checkpoint, the Qwen3-VL text encoder (~8GB), and the Qwen-Image VAE (~250MB) loaded through the normal loaders. This node does nothing useful on any other model family - the whole trick depends on Krea 2's specific encoder.

Where people get stuck

If nothing seems to happen, the two usual suspects: the guide_card output isn't wired into a Reference N guide card slot on the stack encoder (it silently can't go anywhere else), or the strength is genuinely too low - 0.2 default is deliberately subtle, and appearance recipes whisper below about 0.5 by design. If a style reference starts dragging its own subject into your shot, drop Study this image at a notch (lower detail level = looser idea, less literal copying). And remember this needs a real Krea-2-compatible CLIP/VAE pair loaded upstream; point it at an SDXL or Flux pipeline and the Krea CLIP input just won't accept what you're feeding it.