generated

Easy Apply IPAdapter (FaceID Kolors)(ipadapterApplyFaceIDKolors)

Easy Apply IPAdapter (FaceID Kolors) node from ComfyUI-Easy-Use.

ipadapter Apply Face ID Kolors

model
image
image_negative
attn_mask
clip_vision
optional_ipadapter
preset
FACEID PLUS KOLORS
lora_strength
0.60
provider
CUDA
weight
0.80
weight_faceidv2
1.00
weight_kolors
0.80
weight_type
COMBO
combine_embeds
COMBO
start_at
0.000
end_at
1.000
embeds_scaling
COMBO
cache_mode
all
use_tiled
use_batch
sharpening
0.00
Easy Use

This node is a one-stop wrapper for applying IPAdapter FaceID with the Kolors preset. It bundles the usual FaceID setup — face embedding extraction, FaceID LoRA scaling, embedding composition, and IPAdapter attention injection — into a single Easy Apply node. The preset is fixed to FACEID PLUS KOLORS, so the node is ready for face-conditioned generation with a Kolors-compatible model.

Inputs

InputTypeDefaultDescription
modelMODEL—The diffusion model to apply IPAdapter conditioning to.
imageIMAGE—Input face image(s). Used both for InsightFace face embedding extraction and CLIP vision features.
presetCOMBOFACEID PLUS KOLORSFixed preset. Only FACEID PLUS KOLORS is available.
lora_strengthFLOAT0.6Strength of the FaceID LoRA applied before IPAdapter conditioning. Lower values weaken identity transfer; higher values can make the face more prominent but may cause artifacts.
providerCOMBOCUDAExecution provider for InsightFace face analysis. Options: CPU, CUDA, ROCM, DirectML, OpenVINO, CoreML. Use CPU when no compatible acceleration is available.
weightFLOAT0.8Overall weight of the IPAdapter conditioning. Range: -1 to 3.
weight_faceidv2FLOAT1Weight for the FaceID v2 embedding component. Range: -1 to 5.
weight_kolorsFLOAT0.8Weight for the Kolors embedding component. Range: -1 to 5.
weight_typeCOMBO—Weighting mode used when applying IPAdapter attention across the UNet. The exact options come from the underlying IPAdapter implementation.
combine_embedsCOMBO—How the FaceID and Kolors/CLIP embeddings are combined. Options: concat, add, subtract, average, norm average.
start_atFLOAT0Start of the IPAdapter influence interval. Range: 0 to 1.
end_atFLOAT1End of the IPAdapter influence interval. Range: 0 to 1.
embeds_scalingCOMBO—Controls how embeddings are injected into attention. Options: V only, K+V, K+V w/ C penalty, K+mean(V) w/ C penalty.
cache_modeCOMBOallWhat to cache between calls. Options: insightface only, clip_vision only, ipadapter only, all, none.
use_tiledBOOLEANfalseEnables tiled processing, which can reduce VRAM usage when working with larger images.
use_batchBOOLEANfalseEnables batch handling for the input images and embeddings.
sharpeningFLOAT0Sharpening amount applied during IPAdapter processing. Range: 0 to 1.
image_negativeIMAGE—Optional negative reference image. Used as negative face/embedding conditioning.
attn_maskMASK—Optional attention mask. Restricts the IPAdapter effect to specific regions of the latent/image.
clip_visionCLIP_VISION—Optional CLIP Vision model. If not provided, the node relies on the preset/cached model handling.
optional_ipadapterIPADAPTER—Optional IPAdapter model. Use this to supply an externally loaded IPAdapter model instead of relying on the internal preset setup.

Outputs

The node definition provided by the source exposes no output slots.

Usage Notes

  • The preset input is fixed to FACEID PLUS KOLORS, so no preset switching is needed for normal use.
  • Tune lora_strength before adjusting weight. If the face identity is too weak, try increasing lora_strength, weight, or weight_faceidv2. If the output looks overfitted or distorted, lower lora_strength first.
  • start_at and end_at control the denoising interval where the IPAdapter is active. The default full range (0 to 1) applies conditioning throughout. A narrower interval can help preserve composition while retaining identity.
  • cache_mode controls caching of InsightFace, CLIP vision, and IPAdapter resources. all is convenient for repeated generations; none is safer when inputs change often and stale cached features could be a problem.
  • image_negative can be supplied for negative face conditioning, which may help suppress unwanted facial features.
  • attn_mask is useful when the face conditioning should only affect a specific part of the image, such as the face region.
  • If the available hardware does not support CUDA, switch provider to CPU or another appropriate provider; otherwise InsightFace face analysis may fail at runtime.

Comments

Sign in with GitHub to join the discussion.

Loading comments…