IPAdapter ClipVision Enhancer Batch V2

IPAdapter ClipVision Enhancer Batch V2 lives in the "dev" menu, and it's the node you pull out when your reference image is busy and the plain adapter keeps missing the point. CLIP vision squashes your reference into a small embedding, and for a simple image that's plenty - but for a crowded scene, a detailed illustration, or a small subject lost in a background, that single embedding drops most of what's in the picture. This node fixes that by tiling the image, encoding each tile, and merging the results back into one richer conditioning.

How the enhancement works

Under the hood it runs the same IP-Adapter apply path as the Advanced node, with two extra dials that change how the image is encoded rather than how it's applied. enhance_tiles (default 2, range 1–16) splits the reference into an N×N grid; each tile gets its own CLIP vision pass, then the tile embeddings are recombined so the model sees both the whole scene and its parts. enhance_ratio (default 0.5) is the blend between the full-image embedding and the tiled one. At 1.0 you lean on the full image, at 0.0 you lean entirely on the tiles - 0.5 splits the difference, which is why it's the default. The Batch variant adds unfold_batch and an encode_batch_size input (0 to 4096) so you can enhance a stack of frames without running out of VRAM - same batch semantics as the other Batch nodes in the pack.

When it's worth the cost

The honest trade-off: tiling multiplies the encoding work, so generation gets slower, and more tokens flowing into the adapter can push detail around in ways you didn't ask for. Reach for it when the reference is the whole point - a reference painting with a distinctive texture, a group photo where you want the composition - and stick with enhance_tiles: 1 (which effectively disables the enhancement) for simple portraits where the standard path is already good. This is a "fix the specific problem" node, not a "always leave it on" node.

The inputs you'll touch

All the usual apply-node inputs are there: model, ipadapter, image, weight (default 1.0, up to 5), weight_type, start_at/end_at, embeds_scaling, plus optional image_negative, attn_mask, and clip_vision. The two stars are enhance_tiles and enhance_ratio above. Single output: MODEL, straight into your KSampler.

Installing it

Same as the rest of the pack. ComfyUI Manager: search "ComfyUI_IPAdapter_plus_V2". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_IPAdapter_plus_V2

Restart. Adapter weights in ComfyUI/models/ipadapter, CLIP vision in ComfyUI/models/clip_vision.

Gotchas

This pack is a renamed copy of cubiq's V2 codebase - it's a snapshot, so it won't track upstream fixes on its own. And the dev-category label is fair warning: tune enhance_tiles on a single frame before you run it on a whole batch, or you'll burn a lot of VRAM discovering that tile 4 was always the problem.