KG Krea 2 Reference Stack Encoder V11

Most reference systems bolt something onto the model: an adapter with trained weights, a separate encoder, a second network in the loop. This node does the opposite. It never touches the diffusion model - it builds conditioning, by making the text encoder answer a few extra questions.

How it works

Your prompt and up to twelve prepared reference images go into the Krea Qwen3-VL encoder together, tagged as image tokens inside the chat template - the same way an LLM "sees" an image in a conversation, since Krea 2's encoder is a vision-language model rather than a CLIP. That's one encode with everything active.

Then, for each card, the encoder runs again with that one reference muted, and the node subtracts: the difference between "everything" and "everything minus this image" is that image's isolated contribution. Those per-reference deltas get re-added in proportion to each card's requested strength, with a per-layer gain from the card's role and manual dials, and every layer's scale is clamped to ±6 so one hot card can't push a single layer off the map.

Two consequences. First, cost grows roughly linearly with active cards - each card with a non-neutral target is an extra encoder pass, which is why twelve references is a limit rather than a target. Second, because the deltas depend on the content and never on the strengths, the node can cache the expensive part and let you re-tune strengths for free.

Then there's the timing. When images guide decides whether cards act across the whole denoise, per their own recipe timing, or split into two phases at Early-to-final handoff (default 0.4) - early layout versus final detail, stitched with different conditioning strengths across the timestep range.

The settings that matter

Krea CLIP is the Krea Qwen3-VL text encoder. Not a generic CLIP, and if you swap the wrong file in you'll find out fast.

Final image prompt should describe the picture you want, not list the references - the images are input, this is the output. Written prompt strength (0–10, default 1) scales the text side only; any value other than 1 costs one more encoder pass.

Image detail level (256 / 384 / 512 / 768) is the study size for references. Medium 384 is the balanced default; higher reads more exact detail and copies more unwanted content with it. Image framing prepares the reference - full shape, center square or stretch - not your output canvas.

Image slider feel maps requested strengths through a curve: artist friendly is soft at low values, literal slider values passes them straight through, and extra gentle is for stubborn references. Set it and leave it alone while you A/B; changing it mid-comparison makes every result incomparable.

Balance strong cards is the V11 headline. With several hot cards, the node sums the peak layer targets and, if the total exceeds the budget, scales every signed target toward zero by a single factor - small budget means stronger limiting. It ignores the pooled channel - Krea's Qwen encoder has none. Off by default, and the docs are refreshingly blunt that this bounds coefficients, not perceived change.

Reuse image studies is the cache switch: studies are keyed on a SHA-256 of the prepared image bytes plus the CLIP's patch revision, and unknown revisions bypass reuse entirely.

Outputs

conditioning goes to the sampler's positive input. stack_report is a readable string - wire it to Preview Any and read it when something's off. prepared_references is an IMAGE batch: a contact sheet, one frame per card, padded onto a shared dark canvas, showing exactly what the encoder studied. That third output is the debugging tool nobody knows exists. When a reference seems to be doing the wrong thing, look at it first.

Install

ComfyUI Manager → search Krea Reference, or:

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

Restart, refresh. requirements.txt in the pack says there are no pip dependencies beyond a normal ComfyUI installation, and no model weights ship with it. You need three files from elsewhere: the Krea 2 diffusion model, the Qwen3-VL encoder, and the Qwen image VAE. Missing the last two is the single most common Krea complaint - people install the checkpoint and conclude the model is broken.

Where people get burned

Balance strong cards changed its labels in V11 - gentle sign-preserving balance and strict sign-preserving balance - and there's no automatic migration, so a workflow saved against V10 needs that widget reselected by hand. Nothing silently converts; you just get the old numbers.

If you patch the model in place outside Comfy's revision protocol, the study cache can't tell it changed. Set always re-study when you're doing that.

And keep expectations calibrated: the pack documents reference borders and weak fine-texture transfer as real limitations, and says plainly that strong settings can still move identity, framing and style. The honest workflow is to start around ±2 to 4 and look at both directions.