KG Krea 2 Reference Stack Encoder V10

The class name is a mouthful (KGTextEncodeKreaImageReferencesV10), but in the node menu it's KG Krea 2 Reference Stack Encoder V10, and it's the node doing the real work in the pack's V10 line: it takes your prompt plus up to twelve Image Guide Card V10 nodes and merges them into the CONDITIONING a KSampler actually uses. Every guide card is just a config packet for one reference image - job, strength, timing. This is where all of that gets combined with your written prompt into a single piece of conditioning, and V10's version of it adds two things beginners will actually appreciate: it tells you what it did, and it stops your cards from fighting each other.

What it's doing under the hood

The flow is Load Image → Image Guide Card V10 → Reference Stack Encoder V10 → KSampler. Krea 2's text encoder is Qwen3-VL - a vision-language model - and Krea CLIP here is that same encoder object. This node doesn't inject a separate adapter's embeddings the way IP-Adapter does; it runs your reference images and prompt through Krea's own native multimodal encoder, using the per-card directives to decide how each image should be prepared and weighted before that pass.

What you set

Most of this mirrors V9 directly: Krea CLIP, Final image prompt, Written prompt strength (0–10, default 1, raise it when the text should win over the images), Image slider feel (artist friendly / literal / extra gentle), Image detail level (256 to 768), Image framing (keep shape / center crop square / stretch to square), When images guide (smart per-card timing / whole image / layout early, details later), Early-to-final handoff (0–1, default 0.4), and Text/logo guard prompt handling (full rewrite vs keep my prompt words). If you've read the V9 article those are the same knobs doing the same job.

Two fields are new to V10, and they matter more than the rest combined once you're stacking several hot cards:

• Balance strong cards - off - use my values, gentle balance, or strict balance. When several cards are all pulling hard at once, this budgets their total departure from the prompt so they don't simply overpower each other and the result doesn't turn into mush. Leave it on gentle balance until you have a specific reason to trust your own numbers over it.
• Reuse image studies - reuse between runs - faster tuning (the default) vs always re-study. With reuse on, the encoder caches what it studied about each reference by content, so re-running the graph after nudging one card's strength or timing doesn't force every card to re-run its encoder pass. This is the difference between fast iteration and waiting through a full re-study every time you tweak a slider.

The Reference 1–12 guide card slots work exactly like V9, accepting KG_KREA_REFERENCE outputs from Image Guide Card V10 nodes.

The two outputs V9 doesn't have

Alongside conditioning, this node gives you stack_report - a plain-language string of what every card requested, what it actually got, and why - and prepared_references, an IMAGE output that's a contact sheet of exactly what the vision encoder studied after each card's preparation treatment. Wire stack_report into a Show Text node and prepared_references into a Preview Image node the first time you build a multi-card stack; it turns "why isn't this doing what I expected" from a guessing game into something you can actually read. This is the single biggest reason to reach for V10 over V9 once you're running more than one or two cards at a time.

Installing it

Same pack, same steps: ComfyUI Manager (search "ComfyUI Krea Reference") or cd ComfyUI/custom_nodes && git clone https://github.com/kgilper/krea-reference.git, then restart ComfyUI. MIT license, no bundled model weights - you need your own Krea 2 checkpoint, the Qwen3-VL text encoder (~8GB), and the Qwen-Image VAE (~250MB) already loaded. For a working first pass, try the pack's own krea-v10-full-showcase-workflow.json or krea-v10-reference-stack-workflow.json, both wired with the report and preview outputs already connected.

Where people get stuck

If stack_report says a card was skipped or heavily rebalanced when you expected it to hit hard, that's Balance strong cards doing its job - turn it to off - use my values if you specifically want your raw numbers respected over the automatic budgeting. If tuning feels slow, check Reuse image studies is actually on reuse between runs; a stray always re-study will re-run every reference's encoder pass on every single queue, which adds up fast with several cards connected. And as with every node in this pack, it depends entirely on the Krea CLIP input actually being Krea 2's Qwen3-VL encoder - point it at a CLIP from any other model family and there's no equivalent multimodal pathway for it to use.