Easy Apply IPAdapter (From Params)(ipadapterApplyFromParams)

The Easy Apply IPAdapter (From Params) node applies an IPAdapter model to a diffusion model using a prebuilt IPADAPTER_PARAMS object. It is the params-first counterpart to the regular Easy Apply IPAdapter nodes: rather than setting every adapter option directly on this node, you construct the settings upstream and pass them in through ipadapter_params. This keeps workflows modular and makes it easy to reuse the same adapter settings across different models or experiments.

Inputs

model

The base diffusion model that the IPAdapter will be applied to. The output MODEL is a modified version of this model with the IPAdapter conditioning applied.

preset

Selects the IPAdapter preset to load. The available options are populated by the package and ComfyUI based on the installed presets. The preset determines which IPAdapter weights and default behavior are used during application.

ipadapter_params

An IPADAPTER_PARAMS object containing the IPAdapter settings to apply. Build this upstream using the appropriate Params node, then pass it here. This object typically carries the input image, weights, start/end percentages, and other adapter controls.

combine_embeds

Controls how IPAdapter embeddings are combined when multiple embeddings are present.

• concat – concatenates the embeddings.
• add – element-wise addition.
• subtract – subtracts one embedding from another.
• average – averages the embeddings.
• norm average – averages then normalizes the result.
• max – element-wise maximum.
• min – element-wise minimum.

Choose the method that matches the desired result: average and concat are common for mixing style or appearance, while add/subtract are useful for more explicit embedding arithmetic.

embeds_scaling

Controls how the IPAdapter embedding tensors are scaled.

• V only – uses only the visual features.
• K+V – uses both key and value features.
• K+V w/ C penalty – uses keys and values with a cross-attention count penalty.
• K+mean(V) w/ C penalty – uses keys and mean-pooled values with a cross-attention count penalty.

The default V only is the standard choice for most IPAdapter models; the other modes are available for matching specific IPAdapter variants or advanced tuning.

cache_mode

Controls what is cached between successive applications.

• insightface only (default) – caches only insightface models, which is useful for face-related IPAdapter usage.
• clip_vision only – caches CLIP vision outputs.
• ipadapter only – caches the loaded IPAdapter model.
• all – caches everything for maximum speed on repeated runs.
• none – disables caching to save memory.

Choose a higher caching level when reusing the same adapter repeatedly. Use none when memory is limited or when the adapter/preset changes unpredictably.

optional_ipadapter

Optional IPADAPTER input. If connected, this preloaded IPAdapter is used instead of loading from preset. This is useful when you already have an IPAdapter loaded from another node and want to avoid reloading it.

image_negative

Optional IMAGE input for a negative reference image. When supplied, it is used to build negative IPAdapter embeddings, which can help suppress unwanted features from being transferred into the result.

Outputs

MODEL output

The modified diffusion model with the IPAdapter applied. Connect this to a sampler or another model-modifying node.

IPADAPTER

The IPAdapter model/pipeline that was used. This can be passed to other nodes, for example to avoid reloading the adapter or to chain further adapter operations.

Usage Notes

• This node expects an already-constructed IPADAPTER_PARAMS object. If you need to set all parameters directly on the node, use a different Easy Apply IPAdapter node instead.
• The preset and optional_ipadapter inputs provide two ways to obtain the adapter: choose a preset for automatic loading, or connect an existing IPAdapter to bypass loading entirely.
• When running repeated generations with the same preset, keep cache_mode at all to avoid reloading models every time. If you switch adapters frequently, prefer insightface only or none.
• embeds_scaling and combine_embeds should match the values expected by your IPAdapter model/preset; mismatched settings may produce degraded results.