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.
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)
| Name | Type | Default | Description |
|---|---|---|---|
| clip | CLIP | CLIP/text-encoder object used to tokenize and encode the prompt for Krea2. | |
| model | MODEL | Krea2 model to patch with the optional prompt-adherence enhancer and return downstream. | |
| prompt | STRING | Main text prompt. If reference images are connected, this text is fused with vision tokens before Krea2 conditioning is encoded. | |
| vision_megapixels | FLOAT | 1.00.1–8 | Maximum 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_padding | FLOAT | 0.000–1 | Extra 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_rescaling | BOOLEAN | true | Enables conditioning rescaling with multiplier and optional per-layer weights. Disable to pass the raw encoded conditioning through unchanged. |
| multiplier | FLOAT | 4.00-1000000000–1000000000 | Global multiplier applied to the final conditioning tensor when conditioning_rescaling is enabled. |
| per_layer_weights | 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 | Comma-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_enhancer | BOOLEAN | true | Enables the Krea2 prompt-adherence enhancer wrapper on the model. This adjusts txtfusion behavior during sampling to strengthen prompt following. |
| strength | FLOAT | 1.000–1 | Strength of the prompt-adherence enhancer. 0 disables the enhancer effect; 1 applies the full tuned profile. |
| system_promptopt | STRING | Optional 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. | |
| imagesopt | IMAGE | Optional reference image batch. Each image is converted into Qwen3-VL vision tokens and fused with the text prompt for Krea2 conditioning. | |
| masksopt | MASK | Optional 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)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| CONDITIONING | CONDITIONING | — |