Skip to content
,

ComfyUI IPAdapter Guide (2026): Install, Models, Style Transfer and Fixes

7–11 minutes
Smiling woman with curly hair in a cosy living room holding an IPADAPTER card, with the text Copy any style with one image

IPAdapter lets you use an image as a prompt. Give it a reference picture and ComfyUI carries over its style, subject, colours or composition into a new image. Its creator describes it as a “1-image LoRA”: no training, just a reference image and a few nodes.

In ComfyUI, IPAdapter comes from the ComfyUI_IPAdapter_plus node pack by cubiq. It is still the most popular way to do style transfer and reference-based images with SD 1.5 and SDXL models. The setup trips up a lot of people, though, because the model files have to be named and placed exactly right.

This guide covers installing the pack, the exact files and folders, the Unified Loader presets, the weight types for style and composition transfer, how to combine several reference images and how to fix the common errors. It is checked against the pack’s README and source code in November 2026.

TL;DR

  • Install: ComfyUI Manager > search ComfyUI_IPAdapter_plus (cubiq) > Install > restart.
  • Files: CLIP Vision encoders go in models/clip_vision, IPAdapter models in models/ipadapter (create it). Rename files exactly as listed below.
  • Easiest workflow: Load Checkpoint > IPAdapter Unified Loader (preset PLUS) > IPAdapter > KSampler.
  • Start with weight 0.8 and more steps. Use IPAdapter Advanced with weight type “style transfer” to copy only the style.
  • Works with SD 1.5, SDXL and Kolors. Not Flux, Qwen-Image or Z-Image.
  • Note: the pack has been in maintenance-only mode since April 2025, but it still works with current ComfyUI.

What IPAdapter Does

A text prompt can only describe so much. IPAdapter adds a second input: a reference image. It encodes that image with a CLIP Vision model and injects it into the image model, so the result picks up what the reference looks like.

Common uses:

  • Style transfer: make any subject look like a reference painting or photo style.
  • Character consistency: keep the same made-up character’s look across images.
  • Composition transfer: keep the layout of a reference while changing everything else.
  • Image variations: generate new versions of an image without a long prompt.

IPAdapter works alongside ControlNet. ControlNet fixes structure such as pose and edges, while IPAdapter carries over look and style. See our ComfyUI ControlNet guide to combine them.

Install IPAdapter Plus

  1. Update ComfyUI first. The pack’s README says IPAdapter always needs the latest ComfyUI.
  2. Open Manager > Custom Nodes Manager.
  3. Search for ComfyUI_IPAdapter_plus by cubiq and click Install.
  4. Restart ComfyUI.

Manual install from your ComfyUI folder:

cd custom_nodes
git clone https://github.com/cubiq/ComfyUI_IPAdapter_plus

Do not have the Manager? See how to install ComfyUI Manager.

A note on maintenance: on 14 April 2025 the author put the repository into “maintenance only” mode. It still loads and works, and crucial fixes may still be merged, but do not expect new features.

Download the Models (Exact Names and Folders)

You need two kinds of files: a CLIP Vision encoder and an IPAdapter model. The Unified Loader only finds files with these exact names, so rename them after download. All are on the h94/IP-Adapter page on Hugging Face.

CLIP Vision encoders: ComfyUI/models/clip_vision

Rename to Source file Used by
CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors h94/IP-Adapter models/image_encoder/model.safetensors Most SD 1.5 and SDXL “vit-h” models
CLIP-ViT-bigG-14-laion2B-39B-b160k.safetensors h94/IP-Adapter sdxl_models/image_encoder/model.safetensors vit-G models only

IPAdapter models: ComfyUI/models/ipadapter (create the folder)

File Base What it is
ip-adapter_sd15.safetensors SD 1.5 Basic, average strength
ip-adapter_sd15_light_v11.bin SD 1.5 Light impact
ip-adapter-plus_sd15.safetensors SD 1.5 Plus, very strong
ip-adapter-plus-face_sd15.safetensors SD 1.5 Portraits
ip-adapter-full-face_sd15.safetensors SD 1.5 Stronger face model
ip-adapter_sdxl_vit-h.safetensors SDXL Standard SDXL model
ip-adapter-plus_sdxl_vit-h.safetensors SDXL SDXL Plus
ip-adapter-plus-face_sdxl_vit-h.safetensors SDXL SDXL portraits
ip-adapter_sdxl.safetensors SDXL vit-G model, needs the bigG encoder

Most people only need the ViT-H encoder plus ip-adapter-plus_sdxl_vit-h (for SDXL) or ip-adapter-plus_sd15 (for SD 1.5). You can keep models in sub-folders, and you can point to another location by adding an ipadapter entry in extra_model_paths.yaml. Our ComfyUI model folders cheat sheet lists every folder.

Build Your First IPAdapter Workflow

Start from the default text-to-image workflow with an SDXL or SD 1.5 checkpoint.

  1. Add IPAdapter Unified Loader. Connect the MODEL output of Load Checkpoint to its model input. Set preset to PLUS (high strength).
  2. Add IPAdapter. Connect the loader’s model and ipadapter outputs to it.
  3. Add Load Image with your reference picture and connect it to the IPAdapter image input.
  4. Connect the IPAdapter MODEL output to the KSampler model input (replacing the direct checkpoint link).
  5. Set weight to about 0.8, write a short prompt, and click Queue.

The README suggests lowering the weight to at least 0.8 and increasing the number of steps for better results. The pack’s examples folder on GitHub has ready-made workflows for every feature, which you can drag into ComfyUI.

The simple IPAdapter node has a weight_type with three options: standard, prompt is more important and style transfer. Try “prompt is more important” if the reference overpowers your text.

Unified Loader Presets Explained

Preset Strength Notes
LIGHT Low SD 1.5 only
STANDARD Medium Good general start
VIT-G Medium Needs the bigG encoder
PLUS High Most popular for style and subject
PLUS FACE Portraits Focuses on the face of the reference
FULL FACE Stronger portraits SD 1.5 only

If a preset fails with a “model not found” error, the matching file is missing or misnamed. The legacy IPAdapter Model Loader plus a Load CLIP Vision node work with any file name, but you must pick the files by hand.

Weight, Start/End and Weight Types

Setting What it does Start with
weight How strongly the reference is applied 0.6-0.8
start_at / end_at Which part of the sampling steps IPAdapter is active 0.0 / 1.0
weight_type (Advanced) How the weight is spread across the model’s layers linear, or style transfer
combine_embeds How several reference images are merged concat
embeds_scaling How the image embedding is injected V only (default)

The IPAdapter Advanced node offers these weight types: linear, ease in, ease out, ease in-out, reverse in-out, weak input, weak output, weak middle, strong middle, style transfer, composition, strong style transfer, style and composition, style transfer precise and composition precise.

The README notes that changing the weight type in IPAdapter Advanced is a good way to make the result follow your prompt more closely.

Style Transfer and Composition

This is where IPAdapter shines with SDXL.

  • Style only: use IPAdapter Advanced with weight type style transfer (or style transfer precise). The colours, lighting and brushwork carry over, but the subject comes from your prompt.
  • Layout only: use weight type composition. The arrangement of the reference is kept, but not its look.
  • Both from different images: use the IPAdapter Style & Composition SDXL node, which takes one image for style and another for composition, each with its own weight.
  • Dedicated nodes: IPAdapter Precise Style Transfer and IPAdapter Precise Composition give the same effects with fewer settings.

Style transfer works best on SDXL. If the subject from the reference leaks into your image, lower the weight or try the “precise” variant.

Use Several Reference Images

You can feed more than one reference image:

  1. Batch two or more images together (for example with a Batch Images node) and connect the batch to IPAdapter Advanced.
  2. Choose how they are merged with combine_embeds: concat, add, subtract, average or norm average.
  3. For finer control, use IPAdapter Encoder on each image, merge them with IPAdapter Combine Embeds, and apply with IPAdapter Embeds.

You can also mask a reference to one area of the image with the attn_mask input, which is handy for putting two different styles in two halves of a picture.

FaceID Models

The pack also supports FaceID models, which focus on facial features. They need the insightface Python package, and most FaceID models also need a matching LoRA in models/loras. The IPAdapter Unified Loader FaceID loads the right LoRA automatically if the files follow the README’s names.

Only use face references of made-up characters or of people who have given you permission. Do not use IPAdapter or FaceID to recreate real people without consent.

What About Flux and Newer Models?

IPAdapter Plus supports SD 1.5, SDXL and Kolors. It does not work with Flux, Qwen-Image, Z-Image or video models. Those models usually have their own ways to use a reference image:

Troubleshooting

Problem Likely cause Fix
“IPAdapter model not found” on Unified Loader File missing or wrong name Rename files exactly as listed; check models/ipadapter
“Missing CLIPVision model” Encoder missing or misnamed Put the renamed ViT-H file in models/clip_vision
Size mismatch error SD 1.5 adapter with SDXL checkpoint, or wrong encoder Match the adapter to your base model; vit-G needs bigG
IPAdapter nodes red Pack not installed or ComfyUI too old Install via Manager and update ComfyUI
Output copies the reference too much Weight too high Lower to 0.5-0.7, or use “prompt is more important”
Reference has little effect Weight too low or LIGHT preset Use PLUS and weight 0.8-1.0
FaceID error about insightface insightface not installed Install it in ComfyUI’s Python environment
Does nothing with Flux Not supported Use Flux Kontext instead

For more ComfyUI errors, see our guides to the “prompt outputs failed validation” error and LoRA not working fixes.

FAQ

Is IPAdapter the same as a LoRA?

No. A LoRA is trained on many images. IPAdapter uses a single reference image at generation time, with no training.

Which IPAdapter model should I use?

For SDXL, start with ip-adapter-plus_sdxl_vit-h. For SD 1.5, ip-adapter-plus_sd15. Use the face models only for portraits.

Is IPAdapter Plus still maintained?

It has been in maintenance-only mode since April 2025. It still works, but new features are unlikely.

Can I use IPAdapter with ControlNet?

Yes. Apply IPAdapter to the model and ControlNet to the conditioning in the same workflow.

How much VRAM does IPAdapter need?

It adds a little on top of your base model, mainly for the CLIP Vision encoder. If SDXL already runs, IPAdapter will usually run too.

Can IPAdapter work with img2img?

Yes. Feed a latent from an encoded image into the KSampler and lower the denoise, while IPAdapter adds the reference style.

Sources

Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Smiling woman with curly hair in a cosy living room holding an IPADAPTER card, with the text Copy any style with one image