Krea 2 Encode 🆎

Krea 2 is the open-base-model story of 2026 - a 12B diffusion transformer with a Qwen3-VL text encoder that runs at CFG 1.0 on its turbo variant, and the reference-latent mechanism is how you make it outpaint or edit from an existing image. Krea 2 Encode 🆎 is the node that wires all of that into your graph: it encodes a prompt pair and attaches reference latents in one step.

The practical win is the negative. Most Krea 2 work runs at CFG 1.0, where the negative conditioning does nothing at all. A normal two-encoder setup burns a second text encode on a negative that never gets used; this node produces both positive and negative outputs from one encode, so the turbo graph stays lean. The negative is still there - plug it in anyway; it's harmless and keeps the graph conventional - it just didn't cost you an extra encode.

What it takes in

• clip (required) - the Krea 2 text encoder.
• prompt (required) - describe the whole finished canvas, not just the new area. Krea 2 is generating all of it and matching the reference where it must; a prompt that only describes the outpainted strip is how you get nonsense elsewhere.
• negative_prompt (optional) - ignored at CFG 1.0, which is where turbo lives. An empty negative is the normal case.
• vae + reference - wire both and the reference image gets encoded to latents and appended to the positive conditioning. The reference output of Load Image + Pad 🆎 is the intended source (the unpadded original). Leave them unwired and this is a plain two-prompt encoder.
• extra_image - a second reference, e.g. a style or character plate.
• vlm_reference (default off) - also feed the references to the vision tower so the text encoder describes them. Keep it off for outpainting: the latents already carry the picture, and a description tends to pull the result toward a paraphrase of the source.

The mechanism worth knowing

The reference is fitted to a 384px long edge and a multiple of 16 before encoding. That's not arbitrary: Krea 2's VAE downsamples by 8 and the DiT patchifies by 2, so an odd edge would land on a partial patch and produce a mangled reference. The node handles it so you don't have to think about it - but it's also why you should feed it a clean reference output rather than a raw crop at a random size.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ausboss/ComfyUI-AusBoss.git

Restart ComfyUI and search for AusBoss (Manager: ComfyUI-AusBoss). No extra Python dependencies; minimum ComfyUI 0.27.1. The model itself - the Krea 2 checkpoint, text encoder, and VAE - comes from wherever you load your Krea 2 stack (Comfy-Org publishes the loader pieces).

Troubleshooting

If the reference is being ignored, check that both vae and reference are connected - without the VAE, references are skipped silently, not errored. If the outpaint result looks like the reference but doesn't actually continue it - colors and style carry over, geometry doesn't - that's the "reference latents arrive with no position" problem, and it's what the Krea 2 Outpaint Model Patch 🆎 is for. And if output text has drifted into a paraphrase of the source, flip vlm_reference off.