Nodes/Vantage-Nodes/Vantage Text Encode (Krea2)
ComfyUI Node

Vantage Text Encode (Krea2)

Krea2 (K2) text conditioning with optional vision prompting. Reference images are fed through the Qwen3-VL vision path; an optional per-image mask crops the image to the masked region. No VAE is used because Krea2 has no reference-latent pathway.

By vantagewithai·Created 7 months ago·Updated 2 months ago· 22
Vantage Text Encode (Krea2)
  • clip
  • model
  • images
  • masks
  • MODEL
  • CONDITIONING
prompt
vision_megapixels1.0
mask_padding0.00
conditioning_rescalingtrue
multiplier4.00
per_layer_weights1.0,1.0,1.0,1.0,1.0,1.0,1.0,2.5,5.0,1.1,4.0,1.0
prompt_adherence_enhancertrue
strength1.00
system_prompt
CategoryVantage/Krea2

Inputs (13)

NameTypeDefaultDescription
clipCLIPCLIP/text-encoder object used to tokenize and encode the prompt for Krea2.
modelMODELKrea2 model to patch with the optional prompt-adherence enhancer and return downstream.
promptSTRINGMain text prompt. If reference images are connected, this text is fused with vision tokens before Krea2 conditioning is encoded.
vision_megapixelsFLOAT1.00.1–8Maximum allowed size in megapixels for each reference image before the Qwen3-VL vision encoder. Larger images are downscaled; smaller ones are kept at native size and never upscaled.
mask_paddingFLOAT0.000–1Extra context kept around each mask before cropping, expressed as a fraction of image size added on each side. 0 makes a tight crop; 0.1 keeps about 10% surrounding context.
conditioning_rescalingBOOLEANtrueEnables conditioning rescaling with multiplier and optional per-layer weights. Disable to pass the raw encoded conditioning through unchanged.
multiplierFLOAT4.00-1000000000–1000000000Global multiplier applied to the final conditioning tensor when conditioning_rescaling is enabled.
per_layer_weightsSTRING1.0,1.0,1.0,1.0,1.0,1.0,1.0,2.5,5.0,1.1,4.0,1.0Comma-separated per-layer gains for the 12 Krea2/Qwen3-VL tap layers. Leave as default for the tuned profile, or edit to rebalance specific layers before the global multiplier is applied.
prompt_adherence_enhancerBOOLEANtrueEnables the Krea2 prompt-adherence enhancer wrapper on the model. This adjusts txtfusion behavior during sampling to strengthen prompt following.
strengthFLOAT1.000–1Strength of the prompt-adherence enhancer. 0 disables the enhancer effect; 1 applies the full tuned profile.
system_promptoptSTRINGOptional system-instruction override for how the VLM interprets the reference image together with your prompt. Leave unconnected to use Krea2's trained descriptor. Provide only the instruction text; the node adds the chat-template scaffolding automatically.
imagesoptIMAGEOptional reference image batch. Each image is converted into Qwen3-VL vision tokens and fused with the text prompt for Krea2 conditioning.
masksoptMASKOptional mask batch aligned to the reference images. Each mask crops its corresponding image to the masked region before vision encoding; empty masks leave the full image unchanged.

Outputs (2)

NameTypeDescription
MODELMODEL
CONDITIONINGCONDITIONING