EditUtils: Krea2 Model Config lrzjason

Krea 2 is a 12B diffusion transformer from June 2026 that, unusually, shares its text encoder and VAE with the Qwen family - the text encoder is Qwen3-VL and the VAE is Qwen-Image's. Krea2ModelConfig_EditUtils is the node that encodes that fact for the EditUtils pipeline. One instruction input, one model_config DICT output, and the whole config is the answer to a single question: how should the encode chain treat Krea 2?

The answer, which looks odd until you know the model: as a Qwen. The dictionary this node emits sets model_name: "qwen" (the routing key that sends EditTextEncode_EditUtils down the Qwen encode path), vae_unit: 8 (matching the Qwen-Image VAE's spatial unit), a config_for: "krea2" tag used for logging, and a llama_template built from an instruction. Because Krea 2 uses Qwen's encoder and VAE, reusing the Qwen encoding machinery isn't a shortcut - it's the correct pipeline, and the code comments say as much.

The instruction input is where this node actually differs from the pack's other model configs: it comes pre-filled with a sensible default - "Describe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:" - and if you clear it, the node falls back to that same text. That's a deliberate Krea 2-specific system prompt for the Qwen3-VL encoder: it asks the VLM to describe what's in the reference thoroughly before the edit instruction is applied, which is the pattern that gets Krea 2 (and Qwen edit) to hold composition steady while changing content. Leave it alone until you have a reason not to.

So where does it sit in a workflow? Exactly the same slot as every other config node: Krea2ModelConfig_EditUtils → EditTextEncode_EditUtils.model_config, with image configs coming from a Config Preparer chain and the patched model from Krea2EditApply_EditUtils. That's the four-node spine of a Krea 2 edit here - and the repo's own "Simple Krea2 Depth" workflow is exactly that, with a depth LoRA thrown in. It's also a nice demo of the pack's swap trick: change this one node for a Flux2KleinModelConfig_EditUtils and the whole encode chain re-targets without rewiring anything else.

Two things to keep in your back pocket about Krea 2 itself, since they'll color every edit you run: the open weights went through an alignment pass the hosted model didn't, so faces can come out flattened and some prompt clauses silently dropped (community workarounds exist and cost a little quality), and it renders best when prompted with long, specific natural language. Neither is this node's doing - but they're the context for why your Krea 2 edits behave the way they do. Install: ComfyUI Manager (search "EditUtils") or git clone https://github.com/lrzjason/ComfyUI-EditUtils into custom_nodes, restart. The Krea 2 checkpoint, Qwen3-VL encoder, and Qwen-Image VAE are yours to download.