Text Encode Krea 2 Ostris Edit

The input half of Krea 2's edit trick

Krea 2 isn't an edit model, and this node is a big part of why you don't need one. Text Encode Krea 2 Ostris Edit is the encode side of Ostris's edit workflow: it takes your prompt, up to three reference images, and optionally a VAE, and produces the conditioning that the companion Krea 2 Ostris Edit Model Patch needs to make the model actually edit.

Use the two nodes as a pair. This one builds the conditioning; the patch makes the model read it. On its own it's just a fancy text encoder - which is fine, because the fanciness is the point.

How it works

Krea 2 uses the Qwen3-VL text encoder, a vision language model, and that's exactly why this is possible. The node pushes your reference images through that encoder using Krea's own conditioning template, with Picture N: vision placeholders in the same layout used during training. That layout is the load-bearing detail: edit LoRAs trained with ai-toolkit saw images labeled exactly this way, so when your prompt refers to "the image" or "the first picture", the model knows which one you mean.

Two sizing rules, both matching training. Images sent to the Qwen3-VL encoder are downscaled (never upscaled) to fit 384×384 total pixels. And when a VAE is connected, each image is also VAE-encoded and attached to the conditioning as reference latents, sized to fit 1MP. Those latents are what the model patch consumes.

The inputs that matter

• clip - from a CLIPLoader set to the Krea 2 type. Required, and it has to be the Qwen3-VL-based Krea 2 encoder, not a plain CLIP.
• prompt - your edit instruction, plain text. For edit LoRAs this is usually a trigger phrase plus whatever you want changed about the image.
• vae - connect the Qwen-Image VAE if you want reference latents. Which you do. Without it the images still become vision tokens, but there's nothing for the model patch to latch onto.
• image1 to image3 - your references. Three is a hard cap from the node itself.

Wiring

CLIPLoader (krea2) -> Text Encode Krea 2 Ostris Edit (prompt + images + VAE) -> positive
CLIPLoader (krea2) -> Text Encode Krea 2 Ostris Edit (negative, no images)     -> negative

The negative pass is just a prompt - no images, no VAE. Plenty of edit LoRA workflows run with a plain empty negative and it's fine.

Install

Same pack for both nodes. Search "Krea 2 Ostris Edit" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit.git

restart, done. No extra dependencies. It's a first-party pack from Ostris (the ai-toolkit developer), so it tracks the trainer's edit support.

The gotchas

The big one: your text encoder checkpoint must include the Qwen3-VL vision weights, or the images can't be encoded at all. This is the classic Krea 2 setup failure - the model needs three files (the checkpoint, the Qwen3-VL text encoder at roughly 8GB, and the Qwen-Image VAE at roughly 250MB), and people grab only the first, then report the node as broken. Install all three before you touch this pack.

Second: forget the VAE and the patch node has nothing to work with. The reference silently stops mattering and you'll think the LoRA is weak.

Third, manage expectations. The method is experimental and the community says so plainly: Krea 2 alters the input image, drifts on lighting and color, and occasionally misbehaves on wide aspect ratios. It's not Qwen-Image-Edit. But "train an edit LoRA for Krea 2 and run it in ComfyUI" is something nothing else gives you, and that's the trade you're making here.