IPAdapter Tiled Batch V2
Tiled detail for a whole sequence of references
- model
- ipadapter
- image
- image_negative
- attn_mask
- clip_vision
- MODEL
- tiles
- masks
The batch version of IPAdapterTiledV2 - same tiling trick, but built for when your image input is a sequence of references rather than a single picture. It slices each reference into 256px tiles, conditions the model on each tile with its own attention mask, and treats every image in the batch as its own tiling job. If you're animating with IP-Adapter and your reference frames carry detail you care about, this is the node; the single-image version would fold the whole batch into one conditioning and blur the very detail you tiled to preserve.
Mechanically it's the tiled pipeline with unfold_batch = True: each image in the batch gets encoded, tiled, and conditioned separately instead of being merged into a single average conditioning. On top of that it adds one input the non-batch tiled node doesn't have: encode_batch_size (0–4096, default 0). That's the memory-batching control - it caps how many tile-encodes run at once, so on a tight-VRAM card you can set it to something small like 2 or 4 and trade a little speed for not OOM-ing. 0 means "let it run all at once."
The inputs that matter:
image- a batch of reference images. Match this to however many frames your latent batch uses.weight,weight_type,start_at,end_at- the standard influence controlssharpening(default 0) - contrast-adaptive sharpening per tile; the detail-recovery dialembeds_scaling- injection mode for the tile embeddingsencode_batch_size- the VRAM limiter mentioned above
Outputs: MODEL, plus tiles and masks (the extracted tile images and their attention masks) so you can preview exactly what each frame is conditioning on.
When you'd reach for it
Frame-accurate detail in animation: each frame of a sequence carries its own high-detail reference, and you want each frame conditioned on its own tiles. Also the honest default for anyone whose still-image tiled runs are fine but whose batch runs stutter - the encode_batch_size knob is exactly the escape hatch for that.
Install
Part of chflame163/ComfyUI_IPAdapter_plus_V2, the V2 fork of cubiq's pack. Manager → search ComfyUI_IPAdapter_plus_V2, or:
cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_IPAdapter_plus_V2
Restart. No pip requirements.
Gotchas
Batch length is the classic failure: the reference batch needs to line up with your frame count, and a mismatch silently means some frames get conditioned by the wrong tile-set or skipped. The tile cap (max 4 tiles, with a hard center-crop beyond 4:1 aspect) applies per frame - a panorama in a batch gets its edges thrown away just like the single version does. And per-frame tiling is the most compute-hungry node in this pack; on a 6GB card set encode_batch_size low before it crashes, not after. The KB's general advice still applies: IP-Adapter across all steps softens detail, so keep end_at below 1.0 when you care about sharpness.
Inputs (13)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| ipadapter | IPADAPTER | — | |
| image | IMAGE | — | |
| weight | FLOAT | 1.00-1–3 | — |
| weight_type | COMBO | 15 options: linear, ease in, ease out, ease in-out, reverse in-out, weak input, +9 | |
| start_at | FLOAT | 0.0000–1 | — |
| end_at | FLOAT | 1.0000–1 | — |
| sharpening | FLOAT | 0.000–1 | — |
| embeds_scaling | COMBO | 4 options: V only, K+V, K+V w/ C penalty, K+mean(V) w/ C penalty | |
| encode_batch_size | INT | 00–4096 | — |
| image_negativeopt | IMAGE | — | |
| attn_maskopt | MASK | — | |
| clip_visionopt | CLIP_VISION | — |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| tiles | IMAGE | — |
| masks | MASK | — |