Krea2 Reference Encoder (latent only)

The pack's flagship node, Krea2ReferenceTextEncode, does two jobs at once: it runs your prompt through the Qwen3-VL text encoder and attaches reference latents. Sometimes you only want the second part. Maybe your text conditioning is already encoded and caching it is the difference between an 8-second run and a slow one. Maybe you're chaining references onto an existing positive. That's what Krea2ReferenceEncoder is for - the "latent only" variant, as the display name says.

What it does

Feed it a CONDITIONING (the positive from a normal text encode) and one or more already-VAE-encoded LATENTs, and it appends those latents to the conditioning's metadata under the reference_latents key - the same key the Krea2ModelPatcher looks for. No text is re-encoded, no images go through the VLM, nothing gets tokenized. It's a pure conditioning surgery node.

The slot convention is the same as the rest of the pack, and it matches what the adapter was trained on:

• base_latent - the image being edited, VAE-encoded. Reference slot 1. Per the tooltip, this is your base.
• reference_latent_2 / _3 / _4 - optional style or content references.
• strength - a float from 0 to 2 (default 1.0) that scales every reference latent before it's stored.

The one output is a CONDITIONING that carries your original prompt plus the references, ready for a KSampler running a Krea2ModelPatcher-patched model.

When you'd actually reach for it

Honestly, most people start with the text-encode node and never touch this one. The cases where the latent-only version earns its keep:

• You already have a cached conditioning. If you're iterating on a prompt but swapping references, re-encoding text through Qwen3-VL every time is pure waste. Encode once, swap latents here.
• You're reusing a conditioning built outside this pack - anything that produces a CONDITIONING can feed positive.
• You want strength control. The text-encode node has no per-run intensity knob; this one does, and it's a genuinely useful dial for "how much should the references steer the output."

The two inputs that actually matter are base_latent and strength. base_latent must match the output resolution - the author's docs call it a hard constraint from the training data, and mismatches produce degraded, drifting edits. strength above 1.0 amplifies the reference signal, and past ~1.2 the model tends to overfit to the references and start ignoring your prompt; dial it down before you start fighting the results. Chaining is supported - you can stack a second encoder's output onto the first and the references accumulate.

Installing it

Identical to the rest of the pack; there's nothing model-heavy in the repo itself.

cd ComfyUI/custom_nodes
git clone https://github.com/molbal/ComfyUI-Krea2-MultiRef

Restart ComfyUI, or ComfyUI Manager → Custom Nodes Manager → ComfyUI-Krea2-MultiRef → install. The model files (Krea2 checkpoint, Qwen3-VL text encoder, Qwen-Image VAE) are yours to fetch; the only code dependency beyond stock ComfyUI is einops, which ships with it.

Troubleshooting

• No effect at all = the model isn't patched. Krea2ModelPatcher must sit between the model and the sampler, or reference_latents is silently discarded.
• Reference resolution mismatch. Latents get snapped to multiples of 16, but the base image resolution is not resized - that's on you. Match it to the output.
• A heads-up about the repo docs: the bundled docs mention a negative conditioning input and a matching negative output. The shipped node doesn't have them - positive in, one CONDITIONING out. Don't hunt for a negative socket; with Krea2's refusal quirks, negative prompts are best handled the normal way upstream.

Keep expectations realistic: this is the runtime for an adapter that was still mid-training when its author ran out of GPU budget. The plumbing works; the polish is a work in progress.