KG Krea 2 Image Guide Card V9

This node exists so a reference image can have exactly one job instead of a vague, unspecified influence over the whole result. Instead of dropping a picture in and hoping the model figures out what you meant by it, a guide card makes you say it out loud: keep the subject, borrow the lighting, suggest the material, avoid the logo. V9 is the original version of that idea in this pack, and it's deliberately frozen - the README states saved V9 workflows keep working unchanged even as V10 was built alongside it. If you're not sure which version to reach for: V9 if you want the simpler control surface and don't need per-card timing or a counter-example direction; V10 if you do (see the V10 article for what that adds).

How it fits the workflow

A guide card doesn't encode anything by itself - it's a config packet describing one image. You wire its guide_card output into a KG Krea 2 Reference Stack Encoder V9, which is the node that actually combines your written prompt and up to twelve guide cards into the conditioning that goes to KSampler. Stack several cards on one encoder and give each a distinct role: a product shot responsible for identity, an abstract image responsible for style, a third for lighting, a fourth purely to guard against copying text or logos out of a reference. That's the mental model this whole pack is built around, and it's what makes multi-reference workflows tractable instead of a pile of images all pulling in slightly different, unstated directions.

Mechanically, this doesn't work like IP-Adapter - there's no separate adapter model bolted on and no CLIP image-embedding injection through extra cross-attention. Krea 2's text encoder is Qwen3-VL, a vision-language model, and the stack encoder's Krea CLIP input is that same object. The guide card preps your image and attaches directives; the stack encoder runs it through Krea's own native encoder alongside your prompt. It's using the model's own eyes, not a second model.

What you set

• Reference image - the picture this card is about.
• Use image for - the recipe, and the field that does most of the work: balanced, keep the same subject, copy pose and layout, copy lighting and mood, suggest the visual style, suggest material or texture, copy big shapes only, avoid copying text/logos, or manual tuning if none of the eight presets fit and you want to hand-set everything below.
• How strongly this image guides - 0 to 3, default 0.2. The pack's own starting-values table: 0.03–0.08 for a tiny nudge (text/logo guard, stubborn prompts), 0.55–0.9 for the working band where style, palette, lighting and material actually land (several recipes cap themselves around 0.65), 0.9–1.2 for subject or pose anchors where you want real structural pull.

Everything else - Manual mode borrows, 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 - is what a recipe sets on your behalf. Leave them alone until one specific card is misbehaving; then drop into manual tuning and adjust the one field that's actually wrong rather than guessing across all twelve.

Output is a single guide_card of type KG_KREA_REFERENCE. It's a custom type - it only plugs into a Krea Reference Stack Encoder, not a generic conditioning slot.

Installing it

ComfyUI Manager: search "ComfyUI Krea Reference." Manual: cd ComfyUI/custom_nodes && git clone https://github.com/kgilper/krea-reference.git, restart ComfyUI. MIT license, no bundled model weights - you supply your own Krea 2 checkpoint, the Qwen3-VL text encoder (~8GB), and the Qwen-Image VAE (~250MB) through the standard loaders. Nothing in this pack replaces those; it's purely conditioning logic on top of a Krea 2 pipeline you already have running.

Where people get stuck

The recipe strength default (0.2) is intentionally subtle - appearance recipes whisper below about 0.5 by design, so if a card seems to do nothing, that's usually the first thing to raise before assuming it's broken. The output type mismatch is the other common trip: guide_card won't go anywhere except a Reference N guide card slot on the V9 stack encoder, so a card sitting unconnected or wired somewhere it can't attach just silently contributes nothing. And since this depends entirely on Krea 2's own encoder, it does nothing meaningful if the Krea CLIP feeding the stack encoder isn't actually Krea 2's - point this at an SDXL or Flux setup and there's no equivalent pathway for it to hook into.