Sampler Timer Start

Here's a question you'll ask the first time you try tiled video sampling: is this actually faster than just doing it full-frame, or am I just paying the tiling tax in a new way? You can't answer it by feel - tiled sampling changes per-step cost, cache behavior, and total time all at once. Sampler Timer Start is one half of the pack's answer: a way to measure real sampler wall-clock time without changing a single sampler setting.

The mechanism is tidy. The node takes your MODEL, clones it (model.clone() plus a sampler wrapper - the input model is left untouched), and returns the instrumented model plus a sampler_timer handle. You connect the instrumented model to your sampler's model input, run the sampler, and the wrapper records each call's wall-clock time, syncing CUDA at the measurement boundaries so you're not measuring in fantasy time.

The inputs and outputs

• model - connect the output of your checkpoint/model loader. That's the only input.
• model (output) - the instrumented model, wire it into the sampler's model input (or wherever the MODEL enters your guider/custom sampler pipeline).
• sampler_timer - the timing handle, wire it to Sampler Timing Result.

Because the input model is cloned, your sampler behaves identically - same weights, same sampling - you just get timing for free. That's the "no settings touched" promise, and it's why this is safe to drop into an existing workflow for a measurement run.

Where it fits

The README's intended shape: Timer Start → KSampler (or KSampler Advanced / SamplerCustom) → Sampler Timing Result, with the timer handle strung between the two timing nodes. The measurement covers only the sampler - model loading, VAE decode, everything else is excluded - which is exactly what you want when comparing per-tile vs full-frame sampling cost. Timing resets for every queued workflow execution, so results are always fresh.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/maDcaDDie2000/comfyui-video-tiler

Restart; it appears under Video Tiler alongside its counterpart. No extra dependencies, no model downloads - the whole pack rides on PyTorch, which ComfyUI already has. Apache 2.0, vibe-coded for personal use, not actively maintained. The clone-based design means your main workflow is safe if the node ever misbehaves: worst case it doesn't run, it doesn't corrupt your model. That's a small comfort, but with a pack this unmaintained, it's a real one.