IPAdapter Tiled Batch

This is IPAdapter Tiled built for volume. Regular Tiled slices your reference into pieces, encodes each, and applies them with matching masks so a non-square or large reference keeps its detail instead of getting crushed in CLIP vision's square resize. Tiled Batch does the same thing, but adds an encode_batch_size control so you can manage memory when you're processing a lot of frames - an animation, a long image batch - where encoding every tile of every frame at once would blow up your VRAM.

So the "when" is specific: you want tiled reference conditioning and you're running a batch big enough that the plain Tiled node chokes on memory. If you're doing a single still, use IPAdapter Tiled; if you're doing a modest batch, the plain node is probably fine too. Tiled Batch earns its keep on the heavy jobs.

How it works

Mechanically it's identical to Tiled: cut the image into tiles, encode them, build attention masks so each tile's influence lands in the right region, patch the model, and return it. The one addition is batched, memory-aware encoding. That's also why it shares Tiled's three outputs instead of the usual one.

Inputs and outputs worth knowing

The three outputs:

• MODEL - the patched model, into your sampler.
• tiles (IMAGE) and masks (MASK) - the tiles it cut and the masks it built. Preview these to see exactly how the reference got sliced; it's the fastest way to understand a weird result.

The inputs a beginner touches:

• image - the reference to be tiled.
• weight - strength, default 1.0. Ease toward 0.8 if the prompt loses out.
• weight_type - the profile enum (style transfer is a solid default for carrying a look).
• sharpening - a small sharpen on the encoded tiles (0 to 1); keep it low.
• encode_batch_size - the memory dial. Default 0 lets it encode everything in one go; set a positive number (say 4 or 8) to encode tiles in chunks and cap VRAM. If you OOM, this is the knob.

combine_embeds, start_at / end_at, and embeds_scaling behave as across the pack. Optional attn_mask, image_negative, and clip_vision are the usual extras.

Installing the pack

ComfyUI Manager: search ComfyUI IPAdapter plus, install, restart. Or:

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

then restart, keeping ComfyUI updated. CLIP vision encoders in ComfyUI/models/clip_vision, adapter weights in ComfyUI/models/ipadapter (create it if absent), exact filenames for the Unified Loader.

Where people get burned

Same top gotcha as the whole family: the ClipVision/IPAdapter mismatch, a tensor-size error meaning your encoder and adapter aren't the same family (SD 1.5 → ViT-H, bigG/vit-G → their own). First thing to check.

The Tiled-Batch-specific one is memory. If you're here, you're probably running a big batch, and tiling multiplies the encode work - a reference cut into N tiles across M frames is a lot of CLIP passes. If you hit an out-of-memory error, drop encode_batch_size to a small positive value so it processes tiles in chunks rather than all at once. And, like plain Tiled, don't use it to lock a single small subject's identity - tiling can fracture a coherent subject; it's for texture and style across a large area. Pack scope as ever: SD 1.5 / SDXL only, maintenance mode.