LingBot Latent Benchmark Sink

LingBotLatentSink is an output node that does almost nothing: it accepts a LINGBOT_LATENTS tensor, prints its shape to the UI ("LingBot latent shape: (…)" ), and goes home. No VAE decode, no video encode, no pixels anywhere. It's the video-equivalent of hanging a test light on a circuit to see if power flows.

So why would you want it? Because it's the pack's benchmark end-point. The README's own FP8-vs-BF16 timing claims come from exactly this kind of run: a matched sampler test that stops at latents and never pays for decode. When you're tuning settings or comparing CFG modes, you don't want VAE decode and MP4 encoding noise in your timing - you want to isolate the denoise loop. Wire the sampler's latents into this sink instead of LingBotVAEDecode and you get a pure sampler-timing measurement, plus the latent shape as a free sanity check that your frame count and geometry are what you think they are.

It's an output node (is_output_node: true), so it sits at the end of a graph. It takes the place of decode + save for benchmarking runs, or you can keep it as a side branch off a live graph - ComfyUI will execute it for its terminal status without disturbing the main decode path. The shape it reports is a nice confirmation of the 4n+1 math: a 3-second, 20 FPS clip should show up as 61 frames in the tensor's temporal axis.

The one thing to watch

Because the sink replaces the decode path in a benchmark, it's easy to forget you're running a sink and queue a "render" expecting a video - you'll get nothing but a shape string. That's correct behavior, just not a video. When you're done timing, rewire the sampler output into LingBotVAEDecode.

Also worth knowing: the latents arrive as float32 on CPU (the sampler detaches and moves them off the GPU before returning), so this node is cheap - it won't distort the very measurement it's meant to take.

Install

It's in the pack, so:

cd ComfyUI/custom_nodes
git clone https://github.com/ALX-CODE/lingbot-video-1.3b-fp8
pip install -r requirements.txt

or ComfyUI Manager → search "LingBot" → restart. No model files, no extra dependencies - it consumes whatever latent type the pack defines, so it only appears once the pack is loaded.

When you'd actually use it

• Benchmarking FP8 vs BF16 on your own GPU, since the pack's published numbers are single-machine (RTX 5080, one smoke test) and explicitly not universal.
• Tuning cfg_execution modes without decode noise - compare sequential_sage against the batched options and read the fallback reason from the sampler's status line.
• Debugging frame counts fast: the reported shape tells you the temporal axis before you spend a minute decoding.

Beyond that, it's a polite little nothing-node. Every pack needs one, and this one is honest about what it does.