IPAdapter Advanced (Clip Size Aware)

IP-Adapter is the tool you reach for when you want a generation to look like a reference image - style, palette, general vibe - without training a LoRA. This node is a modified IP-Adapter Advanced node (credited to cubiq's ComfyUI_IPAdapter_plus, the pack most workflows use for this), with one fix: instead of always resizing your reference image down to a hardcoded size before encoding it, it checks the actual input resolution the CLIP vision model it's given supports, and resizes to that.

Why that matters: CLIP vision encoders are normally locked to 224x224. This pack's other node, Extend Clip Vision Input Size, patches an encoder to accept something bigger - 448x448, to match the author's own experimental IP-Adapter finetune for NoobAI-XL style transfer. Feed that larger encoder into a normal IP-Adapter node and it'll still crush your reference image down to whatever's hardcoded, throwing away the point of the bigger encoder. This node is the other half of the fix - "size aware" so the two pieces work together.

Mechanically it's the same trick as every IP-Adapter node: extra cross-attention layers read embeddings from your reference image and run alongside the model's existing text cross-attention, rather than replacing it. Your prompt still does its job; the reference image steers style on top of it.

The inputs that matter

Required:

• model - your checkpoint's MODEL output. For this finetune, that means an SDXL-family model - NoobAI-XL is what it was built for.
• ipadapter - an IPADAPTER object from an IPAdapter model loader, pointed at the author's finetuned weights (linked from the README/Civitai page).
• image - your reference image, the style or look you want carried into the output.
• weight - how hard the reference pushes, -1 to 5, default 1, step 0.05. The wider IP-Adapter community's rule of thumb for style transfer is 0.6-0.8; push much past 0.8 and the reference starts overriding your prompt instead of blending with it.
• weight_type - one of 15 presets (linear, ease in/out, weak input/output/middle, strong middle, style transfer, composition, strong style transfer, and a few more) controlling how influence is weighted across layers and steps. Given the finetune, "style transfer" or "strong style transfer" is worth trying before linear.
• combine_embeds - concat / add / subtract / average / norm average. Only matters with more than one reference image; ignore it for a single image.
• start_at / end_at - where in sampling the adapter is active, 0 to 1. Running it across every step softens detail; ending early (0.6-0.8 instead of 1.0) is the standard fix for mushy outputs.
• embeds_scaling - V only / K+V / K+V w/ C penalty / K+mean(V) w/ C penalty. Internals-level attention tuning; leave it default unless chasing a specific artifact.

Optional:

• image_negative - a reference image for what you don't want, the inverse of the main image input.
• attn_mask - a MASK restricting where on the canvas the adapter applies, for regional conditioning.
• clip_vision - the one that makes this whole pack worth using. Feed it the output of Extend Clip Vision Input Size (the sibling node) and your reference image gets resized to the actual encoder resolution - 448 for the author's finetune - instead of a smaller hardcoded default.

Output: a single MODEL, patched with the cross-attention layers baked in. Wire it straight into your KSampler like any other model output.

Installing it

• ComfyUI Manager - search "ComfyUI-ExtendIPAdapterClipVision" and install.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/vahlok-alunmid/ComfyUI-ExtendIPAdapterClipVision, restart ComfyUI.

This pack replaces the apply step only - you still need cubiq's ComfyUI_IPAdapter_plus installed for the loader nodes (IPAdapter Model Loader, CLIP Vision Loader) that feed this one. You'll also need the model files: the author's finetuned IP-Adapter weights from the Civitai link in the README, a NoobAI-XL (or compatible SDXL) checkpoint, and an SDXL-class CLIP vision encoder (ViT-bigG-14).

Common issues

Doesn't do anything useful on Flux, SD3, or other non-SDXL checkpoints. This is an SDXL-family finetune, built for NoobAI-XL specifically. IP-Adapter as a family has never shipped a working Flux port from any mainstream implementation - plug in the wrong base model and you'll get nothing coherent, not a graceful fallback.

Reference image overpowering the prompt. Drop weight toward 0.5-0.6, or shorten end_at so the adapter isn't influencing the last chunk of sampling steps.

Output looks soft or lower-detail than expected. Same fix - running IP-Adapter across 100% of steps costs sharpness. Pull end_at down.

Behaves like a normal, size-locked IP-Adapter node despite the name. Check clip_vision specifically - it's optional, so it's easy to leave unconnected. Skip it and you lose the entire reason this node exists over the stock one. Feed in the output of Extend Clip Vision Input Size, not a raw CLIP Vision Loader output.

No wide community thread to search when something's unclear. Single-author, low-traffic pack built around one person's own finetune - the README is close to the entire documentation surface. If you hit something odd, cubiq's ComfyUI_IPAdapter_plus docs and issues are the closer analog, since this node is a direct modification of that one.