Nodes/Krea2_ComfyUI_Advanced/Eric Krea2 Vision Prompt
ComfyUI Node

Eric Krea2 Vision Prompt

A ComfyUI node in Eric/Krea2 with 11 inputs and 1 output.

By EricRollei·Created about a month ago·Updated 10 days ago· 5
Eric Krea2 Vision Prompt
  • krea2_pipeline
  • image1
  • image2
  • image3
  • conditioning
◄prompt►
◄vision_positionbefore prompt►
◄vision_megapixels0.40►
◄vision_processor_sourceH:\Testing\Qwen3-VL-4B-Instruct-heretic-7refusal►
◄max_sequence_length512►
◄print_prompttrue►
◄vision_templatepicture_n►
CategoryEric/Krea2

Inputs (11)

NameTypeDefaultDescription
krea2_pipelineKREA2_PIPELINE—
promptSTRINGYour instruction/description. Combined with the reference image(s) inside Krea2's own trained descriptor template - not a generic image-caption template.
image1optIMAGEReference image 1. Fed through the vision path only - no VAE, nothing for Krea2's DiT to silently discard.
image2optIMAGEReference image 2 (optional).
image3optIMAGEReference image 3 (optional).
vision_positionoptCOMBObefore promptWhere the 'Picture N: <image>' placeholders sit relative to your prompt text inside the user turn. Matches the community node's convention.
vision_megapixelsoptFLOAT0.400.05–4PER-IMAGE ceiling in megapixels, before auto-sharing (see below). Krea2's text budget is fixed at 512 tokens total (prefix + all images + your prompt) - it doesn't grow with more images, so this node automatically divides ~0.40 MP of shared budget across however many images are connected (1 image ~0.40 MP, 2 images ~0.20 MP each, 3 images ~0.13 MP each) and takes the smaller of that share and this ceiling. Lower this to reserve more room for a long prompt; it never raises the auto-share, only caps it further.
vision_processor_sourceoptSTRINGH:\Testing\Qwen3-VL-4B-Instruct-heretic-7refusalFolder with a preprocessor_config.json for the Qwen3-VL image processor. Krea2's own diffusers folder ships text-only (no image config), so this points at any full Qwen3-VL-4B checkpoint folder that has one - the vision tower is identical, only the image-preprocessing CONFIG (resize/normalize rules) is read from here.
max_sequence_lengthoptINT51264–1024Must match the value your generation node uses (Krea2 default 512).
print_promptoptBOOLEANtruePrint the assembled prompt text + token/image-grid shapes to the console. Turn this on for your first test of this node.
vision_templateoptCOMBOpicture_nHow the vision blocks are labeled inside the user turn - match the LoRA's training. picture_n: 'Picture 1: <vision>' markers - matches ai-toolkit REFERENCE training (ostris lineage, e.g. Style Reference LoRA; pair with the Reference Latents node's ostris_t0 recipe at ~0.15 MP here). bare_edit: bare vision blocks, no labels - matches the identity/instruction EDIT lineage's grounded encode (predict_velocity_edit; pair with the edit_frame recipe). Those LoRAs train the VLM view at ~384-768px, so vision_megapixels ~0.30-0.45 is in-distribution here. For CFG > 1 edits (Raw removals), ground the negative the same way: this node with an EMPTY prompt and the same image.

Outputs (1)

NameTypeDescription
conditioningKREA2_CONDITIONING—