KG Krea 2 Image Guide Card V11

Reference conditioning on Krea 2 is not IP-Adapter. There's no CLIP image embedding, no decoupled cross-attention adapter to load, and the trained identity kits - PuLID, InstantID, FaceID - never got a Krea port because they're SDXL-era technology built on ArcFace embeddings. What you have instead is a multimodal text encoder: Krea 2 reads images and text in the same context window, so a reference image is something you can describe a job for.

That's what this card is: one image, one set of instructions, one packet out.

What it does

The card holds a single reference image plus every setting that tells the stack how to use it. It does no encoding, but it's not a dumb container either - it's where the image gets prepared, and preparation is where most of the actual control lives. Washes, blurs and detail reduction are how you stop a reference from donating things you didn't ask for.

The output is guide_card, type KG_KREA_REFERENCE, and it goes into Reference 1 guide card … Reference 12 guide card on the Reference Stack Encoder V11:

Load Image → Image Guide Card V11 → Reference Stack Encoder V11 → sampler

A card on its own conditions nothing. No stack, no effect.

The four fields to touch first

Use image for is the job. The dropdown runs 33 entries out of the box: the built-in jobs - keep the same subject, copy pose and layout, copy lighting and mood, suggest the visual style, suggest material or texture, copy big shapes only, suggest the color palette, use the background/setting, copy the camera framing, avoid copying text/logos, mood board only, balanced, and manual tuning - followed by any custom recipes you've installed, since the pack bundles recipe packs as YAML and picks them up without a restart. The recipes aren't just strength presets: they pre-tune the image preparation and the per-layer emphasis for that job. Which is why the card's own docs warn that the manual controls "are not universally independent overrides of a quick recipe".

How strongly this image guides goes 0 to 3 and defaults to a barely-there 0.2 - deliberately, because the default "artist friendly" strength curve is soft at low values (below 1.0 it's roughly strength^1.6). What you actually get is also capped by the job: the style and material recipes top out at 0.65, palette and mood board at 0.9, background and framing at 1.2, lighting at 1.25. Cranking the dial past the recipe's ceiling does nothing, so if a card feels dead, check the cap before you blame the image.

Guide direction is the feature that makes this whole thing more interesting than a strength knob: toward this image or away from this image. Away turns the reference into a counter-example - the encoder subtracts its signed contribution. It's the honest way to think about it: steer away from this look, not ban this content. For words and logos specifically, avoid copying text/logos plus the encoder's text guard is the intended route.

When this card guides is recipe decides, whole image, early layout only or final details only. Early is composition, late is finish.

The dials you'll meet later

Manual mode borrows picks one of 12 roles and hands the tuning over - ignore it until a built-in job is nearly right. Prepare image by offers eight treatments, from use image as-is through remove color, soften tiny details, blur words and texture and the shape washes; this is the guaranteed information filter, and the encoder will show you the result. Study this image at sets the study resolution (256 / 384 / 512 / 768, or the stack's setting) - higher is more exact and copies more unwanted content. Frame this reference by prepares the reference, not your output. Color kept and Small details kept are 0–1. Subject copying is a policy, Early layout guidance and Final detail copying are relative phase weights, Maximum image pull caps influence, Shape copied and Overall style reach are manual-mode contributions to structure and style, and Structure layers pull / Finish layers pull bias encoder layers - layers, not spatial masks, despite how they read.

Also: keep the same subject is a policy, not a face lock - nothing here embeds an identity vector, so the same person across scenes is still a LoRA or an edit model.

Install

ComfyUI Manager, search Krea Reference; or into custom_nodes:

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

Restart and refresh. No pip dependencies and no weights ship with the pack - you need Krea 2's model, the Krea Qwen3-VL encoder and the Qwen image VAE.

Where people get burned

A card with How strongly this image guides at 0 is skipped and costs nothing - but its Load Image node still needs a valid filename, so a missing file errors before the stack ever decides to ignore it.

Raising Shape copied and Overall style reach together while you're debugging one problem is the classic own-goal. And use the stack's prepared_references output - it's the contact sheet of what the encoder actually studied, which settles most "why did it do that" questions in one glance.