EasyKSampler (Tiled Decode)

This is the Easy-Use sampler that decodes its result in tiles instead of all at once. Sampling produces a latent; turning that latent back into a viewable image is the VAE decode step, and on large images that decode is where a lot of people run out of VRAM - the sampling fit fine, then the decode tried to hold the whole full-res image at once and crashed. easy kSamplerTiled chops the decode into tile-sized pieces so it stays within budget. If you're rendering big, this is the sampler that lets you do it on a smaller card.

How it works

It's the back half of the Easy-Use split-sampler setup. Your loader and pre-sampler have already stuffed the pipe with everything - model, conditioning, latent, sampler settings - so this node's whole job is: run the sampling, then decode the resulting latent to pixels one tile at a time, stitching the tiles back together. Only the decode is tiled; the diffusion itself runs normally. That means it fixes decode-time out-of-memory, not sampling-time out-of-memory.

The inputs that matter

• pipe - the pipe carrying all your settings from easy preSampling (or a full loader). This node reads its steps, cfg, sampler and seed from there, which is why there are no sampler knobs on the node itself.
• tile_size (default 512, range 320-4096) - the size of each decode tile. Smaller tiles use less VRAM but add a little overhead; 512 is a fine default and you rarely need to touch it. Drop it toward 320 if you're still hitting OOM on a very large image.
• image_output - what to do with the result: Preview, Save, Hide, Hide&Save, Sender, Sender&Save, or None. Preview shows it without writing a file; Save writes to your output folder.
• save_prefix - the filename prefix when you're saving.
• Optional model - lets you override the model in the pipe for this pass without rebuilding the loader.

Outputs: the pipe (so you can keep chaining) and the decoded image. This is a terminal output node, so it'll display or save on its own.

How to install it

Via ComfyUI Manager: search "ComfyUI Easy Use", install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

Then install.bat / pip install -r requirements.txt, restart ComfyUI. Preinstalled on comfy.icu.

Common issues

Know what problem this actually solves. If your workflow dies during sampling - the diffusion pass itself - tiled decode won't save you, because the OOM is happening before decode ever runs. This node only helps when the crash is at the VAE-decode step, which is the classic symptom on large or upscaled outputs. Diagnose where the crash is before assuming this is the fix.

Second, tiled decode can occasionally leave faint seams where tiles meet, especially on smooth gradients like skies. It's usually invisible, but if you spot a grid pattern, bump tile_size up so there are fewer, larger tiles. On a card with plenty of VRAM for the image you're making, you don't need this node at all - the plain easy kSampler decodes in one shot and skips the tiling overhead entirely.