Easy Apply IPAdapter (Advanced)(ipadapterApplyAdvanced)

The Easy Apply IPAdapter (Advanced) node applies an IPAdapter to a base model with a wide set of controls for embedding composition, attention masking, LoRA strength, provider selection, caching, and tiled processing. It is designed as a flexible, all-in-one IPAdapter application node that can handle FaceID-style presets and advanced workflows.

Inputs

Required

• model (MODEL) — The base diffusion model that will have the IPAdapter applied to it.
• image (IMAGE) — The reference image used as the image prompt. This image drives the content and style injected into the model.
• preset (COMBO) — Selects the IPAdapter preset configuration. Available options depend on the installed IPAdapter presets/models (for example, standard, Plus, FaceID variants, etc.).
• lora_strength (FLOAT, default = 0.6, range 0–1) — Strength of the LoRA associated with the selected preset. This is especially relevant for FaceID-style presets that use a LoRA to help preserve identity. 0 disables the LoRA effect, and 1 applies full LoRA strength.
• provider (COMBO, default = "CUDA", options: CPU, CUDA, ROCM, DirectML, OpenVINO, CoreML) — Computation provider used for InsightFace/FaceID operations. Choose the provider that matches your hardware and runtime.
• weight (FLOAT, default = 1, range -1–3) — Global strength of the IPAdapter influence on the model. Higher values increase the effect; negative values can invert the influence.
• weight_faceidv2 (FLOAT, default = 1, range -1–5) — Extra weight specifically for FaceID v2 presets. Use this to fine-tune face similarity when using a v2 FaceID preset.
• weight_type (COMBO) — Determines how the IPAdapter weight is applied over the denoising process. Different methods create different ramping curves for the influence.
• combine_embeds (COMBO, options: concat, add, subtract, average, norm average) — Method used to combine image embeddings, such as when both a positive and negative image are provided, or when multiple embeddings are present.
• start_at (FLOAT, default = 0, range 0–1) — Starting point of the denoising loop at which IPAdapter begins to apply its effect. 0 means from the very first step.
• end_at (FLOAT, default = 1, range 0–1) — Ending point of the denoising loop at which IPAdapter stops applying its effect. 1 means it stays active until the final step.
• embeds_scaling (COMBO, options: V only, K+V, K+V w/ C penalty, K+mean(V) w/ C penalty) — Controls how the image embeddings are projected into the model's attention mechanism. Different scaling modes affect the balance between keys and values.
• cache_mode (COMBO, default = "all", options: insightface only, clip_vision only, ipadapter only, all, none) — Determines which intermediate results are cached for reuse. Caching improves performance when applying IPAdapter multiple times.
• use_tiled (BOOLEAN, default = false) — Enables tiled processing for large images, reducing memory usage at the cost of some processing overhead or quality trade-offs.
• use_batch (BOOLEAN, default = false) — Processes the input images as a batch instead of individually, which can be useful when applying multiple image prompts at once.
• sharpening (FLOAT, default = 0, range 0–1) — Applies sharpening to the input image before feature extraction. Higher values can increase detail and definition.

Optional

• image_negative (IMAGE, optional) — A negative reference image. Its embeddings are combined with the main image using the combine_embeds mode to help steer the generation away from unwanted features.
• attn_mask (MASK, optional) — An attention mask that limits the IPAdapter influence to specific regions of the image. Areas outside/inside the mask can be controlled depending on mask semantics.
• clip_vision (CLIP_VISION, optional) — A CLIP Vision model used for encoding the input image. If omitted, the model associated with the selected preset/IPAdapter is used.
• optional_ipadapter (IPADAPTER, optional) — An already configured IPAdapter object. Use this to reuse a previously loaded or customized adapter instead of building one from the preset.
• layer_weights (STRING, default = "") — Optional string for fine-grained control over the contribution of individual IPAdapter layers. Leave empty for default layer behavior.

Outputs

• MODEL — The modified model with the IPAdapter applied. Connect this to a sampler or other model-processing nodes.
• IMAGE — The processed image from the IPAdapter input/preprocessing stage. This can be used for inspection or passed to other nodes.
• MASK — The attention mask used by the node, after any internal processing. Useful for debugging or reusing the effective mask downstream.
• IPADAPTER — The configured IPAdapter object. You can reuse this adapter in other compatible nodes to avoid rebuilding it.

Usage Notes

• The provider and lora_strength settings are particularly important for FaceID-style presets. Make sure the selected provider matches your InsightFace installation and that LoRA strength is balanced for identity preservation without distorting the generated result.
• Use start_at / end_at to control when IPAdapter influence occurs in the denoising process. Applying IPAdapter only in later steps can preserve more composition freedom, while earlier application yields a stronger overall effect.
• combine_embeds becomes relevant when image_negative is provided. Modes like concat and average can produce very different steering behavior.
• Toggle use_tiled if you run into memory issues with high-resolution reference images or attention masks.
• cache_mode = "all" is recommended when reusing the same adapter or reference image across multiple k-samplers. Set it to "none" if memory usage is a concern or your workflow uses many distinct inputs.
• The IMAGE and MASK outputs are provided for debugging or for chaining into workflows that need to visualize or post-process the reference image and attention mask.