KG Krea 2 Reference Stack Encoder V9

Search this node by class name (KGTextEncodeKreaImageReferencesV9) and you might not recognize it - in the node menu it's called KG Krea 2 Reference Stack Encoder V9, and that name is the more useful one to remember, because this is the node doing the actual work. Every Image Guide Card V9 you build is just a config packet; this is the encoder that reads them, reads your prompt, and produces the CONDITIONING that goes to KSampler. It's a straight replacement for a normal CLIP Text Encode node, just one that also accepts up to twelve reference images with individual jobs attached.

How it thinks

The whole pack rests on one idea: instead of one image vaguely influencing a result, each reference image gets a stated job - preserve the subject, borrow a style, copy lighting, suggest a material - set on its own guide card. This node is where those jobs get combined with your written prompt into one piece of conditioning. The flow is always Load Image → Guide Card V9 → Reference Stack Encoder V9 → KSampler. Mechanically it isn't IP-Adapter-style embedding injection through a bolted-on adapter; Krea 2's text encoder is Qwen3-VL, a vision-language model, and Krea CLIP here is that same encoder - this node runs your images and prompt through Krea's own native pathway rather than a second model sitting alongside it.

What you actually set

• Krea CLIP and Final image prompt - the same two things any text encode node needs: your loaded Krea 2 CLIP and your prompt text.
• Reference 1–12 guide card - optional slots for KG_KREA_REFERENCE inputs coming from Image Guide Card V9 nodes. Leave them all empty and this behaves like a plain text encoder.
• Written prompt strength - 0 to 10, default 1. Raise this when the text prompt should win over the reference images rather than get diluted by them.
• Image slider feel - how the How strongly this image guides value on each card actually translates: artist friendly - soft at low values (the forgiving default), literal slider values (what you set is what you get, no smoothing), or extra gentle for stubborn references (for an image that keeps overpowering everything else no matter how low you set its strength).
• Image detail level - the default study resolution for cards that don't override it themselves (low - loose idea (256) up to very high - most exact (768)). Lower this globally if references keep dragging in unwanted specifics; individual cards can still override it per-image.
• Image framing - how references get cropped before study: keep the original aspect, center-crop to square, or stretch to square. Matters most when your references and your target output have mismatched aspect ratios.
• When images guide - timing for the whole stack: smart per-card timing, guide the whole image, or layout early, details later.
• Early-to-final handoff - 0 to 1, default 0.4, the crossover point when timing is split between early layout and final details.
• Text/logo guard prompt handling - full guard - rewrite my prompt vs gentle guard - keep my prompt words, for when you've got a text/logo-guard card in the stack and need to decide how aggressively it's allowed to touch your written prompt to enforce that guard.

Output is a single CONDITIONING, ready for a KSampler's positive input - no separate stack report or preview on this version; that's a V10 addition.

Installing it

Same pack as every node here: ComfyUI Manager, search "ComfyUI Krea Reference," or cd ComfyUI/custom_nodes && git clone https://github.com/kgilper/krea-reference.git and restart. MIT licensed, no model weights included - bring your own Krea 2 checkpoint, the Qwen3-VL text encoder (~8GB), and the Qwen-Image VAE (~250MB), loaded through the standard ComfyUI loaders. The fastest way to see this node working correctly on a first run is the pack's own showcase workflow - copy example_assets/krea-reference-examples into your input/ folder and load krea-v9-full-showcase-workflow.json.

Common trip-ups

Nothing shows up in the output image at all if the guide cards aren't actually reaching a Reference N slot - a card left dangling contributes nothing and gives no error. If your text/logo guard card doesn't seem to be doing anything, check Text/logo guard prompt handling: gentle guard keeps your exact prompt words, which can undercut the guard if your prompt itself names the thing you're trying to suppress. And because this node is Krea-2-specific, feeding it a CLIP from any other model family won't error cleanly - it just won't produce anything coherent, since the whole mechanism depends on Qwen3-VL's specific ability to take images as input alongside text.