FL KSampler Plus V2

This is the one to use. FL_KsamplerPlusV2 is the second cut of Fill-Nodes' tiled sampler, and it exists because the original had a rough edge: when it split your prompt conditioning across tiles, it didn't carry everything cleanly, and results could smear. V2 does a complete conditioning split across tiles - including the pooled outputs and tensor-based conditioning that the first version glossed over - plus two new knobs. If you're choosing between this and FL_KsamplerPlus with a blank slate, pick this.

The core job is the same as V1: slice a too-big image into an overlapping grid, sample each tile so only one is in VRAM at a time, and blend the overlaps with cosine-weighted masks so the seams disappear. That's how you sample large output on a card that would otherwise OOM - the same idea as Ultimate SD Upscale and Tiled Diffusion, run inside a single node.

What's new over V1

Two inputs the original doesn't have:

• conditioning_strength (0.1–5.0, default 1.2) - a multiplier on the per-tile conditioning. The 1.2 default nudges the prompt influence up slightly per tile, which helps each tile stay on-theme without you rewriting the prompt. Push it down toward 1.0 if tiles feel over-cooked, up if they drift off-prompt.
• debug_mode (on/off) - when on, the node emits a debug_info STRING output describing what it did with the tiling and conditioning. Genuinely useful when a tiled result looks wrong and you can't tell whether it's the grid, the overlap, or the conditioning to blame.

Everything else mirrors V1. The tiling controls that matter:

• x_slices / y_slices (1–8) - the grid. Smallest that fits VRAM.
• overlap (default 0.25) - raise toward 0.3–0.4 if grid lines show.
• input_type - latent or image; feed an image and you must also wire a vae to encode it.

Outputs are the passthrough set - model, positive, negative, latent, vae, image - plus debug_info. Wire image to your save node, or latent to keep going.

Installing it

It ships in ComfyUI_Fill-Nodes. Via ComfyUI Manager, search Fill-Nodes, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_Fill-Nodes

then restart ComfyUI. Nothing extra to download for the node itself - it runs on the model, VAE, and conditioning you connect.

The caveat that carries over

Same honest warning as the V1 writeup: tiled sampling is not tiled upscaling with a tile condition. Nothing here anchors each tile to a source image the way ControlNet Tile does, so at high denoise the tiles can invent divergent detail and you get a subtle patchwork. Keep denoise conservative for refinement passes, turn debug_mode on when something looks off, and lean on conditioning_strength and overlap before you blame the model. For flat-ish content it's clean; for faces or text straddling a seam, watch it.