Nodes/Krea 2 Identity Edit/Krea2 Edit (source patch)
ComfyUI Node

Krea2 Edit (source patch)

Adds the krea2_edit in-context source-preservation path (source latent as frame=1 tokens) to a Krea2 model.

By lbouaraba·Created about a month ago·Updated 21 days ago· 477
Krea2 Edit (source patch)
  • model
  • source_latent
  • source_latent_b
  • ref_boost_mask
  • vae
  • source_image
  • source_image_b
  • target_latent
  • MODEL
ref_boost1.00
ref_boost_a1.00
fit_modefit
Categorykrea2edit

Inputs (11)

NameTypeDefaultDescription
modelMODEL
source_latentLATENT
source_latent_boptLATENT2nd reference (subject photo) for multi-ref LoRAs -> RoPE frame=2, training-matched order: scene first, subject second
ref_boostoptFLOAT1.000–1000reference-fidelity dial: multiplies target->reference attention. Applies to the LAST ref (= the subject in two-ref workflows, the only ref in single-ref). 1.0 = off, >1 pulls harder toward the reference's appearance, <1 loosens. Optimal value is model-specific
ref_boost_aoptFLOAT1.000–1000same dial for the FIRST ref (= the scene in two-ref workflows). No effect in single-ref workflows. 1.0 = off
fit_modeoptCOMBOfithow an image source fits a mismatched output aspect ratio (needs vae + source_image connected): fit = resample the source to the target grid at a centered offset — matches how this model was trained (default, use this); crop (legacy) = center-crop to the target AR then resize (v1/v1.1 geometry, only for older weights)
ref_boost_maskoptMASKoptional region on the (last) reference to boost, e.g. the face; empty = whole reference
vaeoptVAERECOMMENDED with source_image: enables the blur-proof pixel-space path (crop+resize in pixels, encode internally) — immune to input/output resolution mismatches
source_imageoptIMAGEsource as IMAGE (with vae connected): overrides source_latent with exact pixel-space fitting — fixes blurry results from mismatched resolutions
source_image_boptIMAGE2nd reference as IMAGE (with vae)
target_latentoptLATENTRECOMMENDED with vae + source_image: wire the SAME latent you feed KSampler.latent_image. Lets the node VAE-encode the source here, before sampling starts, instead of on the first step — otherwise the VAE is pulled onto the GPU mid-sampling and can evict part of the diffusion model, slowing every remaining step on VRAM-tight setups

Outputs (1)

NameTypeDescription
MODELMODEL