Kimodo Preview (2D)

Text-to-motion generation is a black box until you look at it. Kimodo Preview (2D) is the fastest way to do that: it renders the generated motion as a skeleton stick figure and hands you a normal ComfyUI image - so it plugs into the same preview/save tools you already use for pictures. It won't win any beauty contests, but it's the difference between queueing an export blind and catching "the robot moonwalked again" in two seconds.

How it works

Feed it the motion object from Kimodo Sampler (or the cleaned one from Kimodo Post Process - either works). Three knobs decide what you see:

• sample_index - which of your num_samples batch variations to show (0-indexed). If you generated 4 samples, this is your paging control.
• frame_index - which frame of the clip to draw (0-indexed). Preview is single-frame, so scrub by hand if you want a specific pose.
• image_size - output square side, 256–1024, default 512. Bigger is easier to read, costs nothing but pixels.

Output: one image (IMAGE type). Wire it to ComfyUI's built-in preview/save-image nodes and it behaves exactly like a rendered frame - you can even stack several Preview nodes set to different frames for a quick pose contact sheet, which is a legitimately handy trick for judging a walk cycle.

Why single-frame is fine

Motion quality is mostly about a few key poses: foot plant, weight transfer, the moment a turn starts. Scrub frame_index through those and you'll see 90% of what's wrong. For actual motion timing you'll want Kimodo Preview 3D, which animates, or the GIF/MP4 outputs on the export nodes - but this node is the cheap sanity check that belongs between sampling and every export. I'd make it a permanent resident in your workflow even after you get the fancy preview working.

Install

Part of Kimodo Motion Bridge - ComfyUI Manager (search the pack title) or clone + python -m pip install -r requirements.txt + python install.py, restart. Node is under Kimodo. It needs the same model pipeline upstream, so it inherits the pack's real requirement: ~17GB VRAM generation on the sampler, and the gated Llama text encoder for first-time encoding.

The honest caveats

• It draws a skeleton, not a character. If you're testing whether your final Mixamo character will look good, this won't tell you - that's what the FBX/Unity/Rive outputs are for.
• Single frame means you can miss slide or float artifacts between frames. Don't use a passing Preview as proof the motion is clean; that's Post Process + the 3D preview's job.

Quick loop: sample a batch, Preview sample 0 at a couple of key frames, and only then decide which sample deserves the export budget.