Vantage Text Encode (Krea2)

Krea 2 is a 12B image model that got a lot of "finally, prompt adherence that doesn't fight you" goodwill when its weights opened, and this node is the ComfyUI-side text encoder for it. It's not a plain CLIP encode: it patches the Krea 2 model with an optional prompt-adherence enhancer, can fuse reference images (and masks) into the conditioning, and lets you rebalance the conditioning per layer - all of it influenced by ideas from the ComfyUI-Krea2T-Enhancer, Conditioning-Rebalance, and Krea2TextEncoder projects the pack credits.

The two knobs are the whole personality of this node, so let's be honest about what they do.

The reference-image fusion

Connect images (an IMAGE batch) and optionally masks, and your prompt text gets fused with Qwen3-VL vision tokens before Krea 2 conditioning is encoded. That's your "show the model what you mean" path - style reference, composition reference, character reference, all mixed with the text prompt. Two controls shape it:

• vision_megapixels (default 1) - the cap for each reference image before the vision encoder. Bigger images get downscaled to this budget; smaller ones stay native. This is your VRAM/quality dial for vision fusion.
• mask_padding (default 0) - when a mask is connected, each image crops to its masked region before encoding. 0 is a tight crop; 0.1 keeps ~10% surrounding context on each side. For "just this object in the frame" references, a little padding reads better.
• system_prompt - optional override for how the VLM interprets the reference images together with your prompt. Leave it unconnected to use Krea 2's trained descriptor; if you override, give only the instruction text - the node adds the chat-template scaffolding itself.

The two boosters

• conditioning_rescaling (default on) with multiplier (default 4) and per_layer_weights - scales the encoded conditioning tensor by a global multiplier plus per-layer gains for the 12 Krea 2/Qwen3-VL tap layers. The default weight string (1.0,1.0,1.0,1.0,1.0,1.0,1.0,2.5,5.0,1.1,4.0,1.0) is a tuned profile; the big bumps on the later layers are doing the heavy lifting. Turn rescaling off to pass raw conditioning through untouched. This is where the Conditioning-Rebalance DNA lives.
• prompt_adherence_enhancer (default on) with strength (default 1) - wraps the model so the txtfusion behavior during sampling gets boosted to strengthen prompt following. 0 disables it, 1 applies the full profile. If Krea 2 feels like it's drifting from your prompt, this is the switch to test first.

Outputs are MODEL (the patched model - run it into your sampler) and CONDITIONING (feed it to the sampler's conditioning input). clip is the CLIP/text-encoder object that does the tokenizing and encoding.

Install

Part of Vantage-Nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/vantagewithai/Vantage-Nodes.git
pip install -r requirements.txt

or ComfyUI Manager → search "Vantage-Nodes" → Install → restart. You need a Krea 2 model and its matching text encoder/CLIP from elsewhere - this node doesn't download them; it encodes with what you connect.

Common issues

• "Why does my prompt do nothing?" - check prompt_adherence_enhancer and conditioning_rescaling are on (they default on, but a saved workflow may have flipped them). Then confirm the clip is actually Krea 2's text encoder, not a generic CLIP.
• Vision fusion eats VRAM - lower vision_megapixels. The vision encoder budget is the main lever.
• Reference crops too tight - raise mask_padding. A tight crop loses context; the fractional padding exists exactly for this.