Easy Apply IPAdapter (Embeds)(ipadapterApplyEmbeds)

The Easy Apply IPAdapter (Embeds) node (class ipadapterApplyEmbeds) applies an IP-Adapter to a diffusion model using precomputed embeddings instead of a reference image. It is intended for workflows that already have EMBEDS from an encoder or embedding-generation node, and it exposes the core IPAdapter controls in a compact “Easy Apply” interface. The node returns a patched MODEL ready for a sampler, plus an IPADAPTER output for chaining further adapter steps. It appears in the EasyUse/Adapter category.

Inputs

Required

• model (MODEL)
  The base diffusion model to which the IP-Adapter will be applied.

• clip_vision (CLIP_VISION)
  A loaded CLIP Vision model. Even though this node uses precomputed embeddings, the IPAdapter pipeline still needs the CLIP Vision model internally. Connect the same CLIP Vision model used when generating the embeddings.

• ipadapter (IPADAPTER)
  The IPAdapter model/configuration to apply.

• pos_embed (EMBEDS)
  The positive embeddings that describe the style, content, or subject to transfer. This is the main conditioning source for the adapter.

• weight (FLOAT, default = 1, range = -1 to 3)
  Overall strength of the IPAdapter. Lower values weaken the effect; higher values increase it. A value of 0 effectively disables the adapter.

• weight_type (COMBO)
  Selects the IPAdapter weighting preset. Different presets change how the adapter’s attention is blended with the prompt, allowing trade-offs between style transfer, composition, and prompt adherence. The exact preset names are provided by the IPAdapter implementation.

• start_at (FLOAT, default = 0, range = 0 to 1)
  The fraction of the denoising process at which the IPAdapter starts being applied. Keep at 0 to apply it from the beginning of generation.

• end_at (FLOAT, default = 1, range = 0 to 1)
  The fraction of the denoising process at which the IPAdapter stops being applied. Keep at 1 to apply it through the end of generation.

• embeds_scaling (COMBO)
  Controls how the precomputed embeddings are used inside cross-attention. Available options:

  • V only: injects only the value component of the embedding.
  • K+V: injects both key and value components from the embedding.
  • K+V w/ C penalty: same as K+V, with an additional cross-attention penalty.
  • K+mean(V) w/ C penalty: uses keys and mean-pooled value vectors, with the same cross-attention penalty.

Optional

• neg_embed (EMBEDS, optional)
  Negative embeddings used to steer the result away from unwanted styles or content. Leave disconnected if no negative embedding is available.

• attn_mask (MASK, optional)
  An optional attention mask that limits the IPAdapter influence to specific regions of the generated image. Leave disconnected to apply the adapter globally.

Outputs

• MODEL
  The IPAdapter-patched diffusion model. Pass this to a sampler or to other model-processing nodes.

• IPADAPTER
  The IPAdapter object that was used. This can be passed to another Easy Apply IPAdapter node when stacking multiple adapters in a single workflow.

Usage Notes

• This node does not accept an image input. If you have a reference image, use the regular Easy Apply IPAdapter node instead; use this node when you already have EMBEDS from a previous IPAdapter encoding step.
• pos_embed is required; neg_embed is optional. If only one embedding is available, leave neg_embed disconnected.
• Use start_at and end_at to control where in the sampling schedule the adapter is active. For example, applying IPAdapter only in the middle of generation can help preserve overall composition while still transferring style.
• The IPADAPTER output is useful for chaining: take the MODEL output from one Easy Apply IPAdapter node and feed it into the model input of another, while reusing the same adapter configuration through the IPADAPTER output.