Nodes/ComfyUI-CGlide/Glide Preview
ComfyUI Node

Glide Preview

Watch the shot move instead of staring at one frozen frame

By CGlide·Created 2 months ago·Updated about 24 hours ago· 39
Glide Preview
  • model
  • model
modeanimated
fps24.0
columns4
every_n_steps2
max_resolution512
sampler_previewfalse
decoderlatent2rgb

ComfyUI's built-in preview shows you a single frozen latent frame while the sampler works. For a video model that's almost useless - you sit there for ten minutes watching a still and you have no idea whether the motion is working. Glide Preview drops between your model and the sampler and plays the whole shot as it denoises, so a bad take dies at step four instead of at minute ten.

The wiring is the easy part. model goes in, model comes out unchanged - it's a pass-through that sits on the model path and intercepts the preview instead of the generation. You put it between your model loader and the sampler, and that's the whole graph change.

The settings that matter

  • mode - animated plays the whole shot on the node while it samples; contact sheet tiles the frames into one still (with columns controlling tiles per row); stock is ComfyUI's frozen first frame.
  • fps - playback rate for the preview. Here's the honest limit: it caps around 7.2 fps, and the author can't fix that. H3 stores about one latent frame per 3.35 output frames, so 7.2 is literally every frame there is. Set it higher and it just uses them all.
  • every_n_steps - redraw cadence (default 2). Raise it if previewing is costing you sampling time.
  • max_resolution - the longest edge of the preview sent to the browser (default 512). It also decides how far the TAE decoder upsamples, so raising it costs decode time as well as bandwidth.
  • sampler_preview - also draw ComfyUI's single-frame preview on the sampler node. Off by default in animated mode, since this node already shows the whole thing.

The decoder input is where the visual quality is decided. latent2rgb is H3's colour approximation - free, but mushy. For real picture, put taeh3.safetensors (from madebyollin) in ComfyUI/models/vae_approx/ and pick it here. You need to restart before it shows up in the dropdown. If the decoder fails mid-run, the node falls back to the stock preview and your generation keeps going - a failure here never costs you a render.

Install

Part of the ComfyUI-CGlide pack:

cd ComfyUI/custom_nodes
git clone https://github.com/CGlide/ComfyUI-CGlide

Restart ComfyUI and it appears under the CGlide category. No extra dependencies for the preview itself - the pack's only real requirement is PyAV, which is normally already installed with ComfyUI, plus ffmpeg on your PATH if you use the pack's video output node. Like every node in this pack, a .py change needs a full ComfyUI restart, not a browser refresh.

Is it a render? No. It's a preview, capped at every latent frame the model actually stores, and the animated mode leans on a decoder that is either mushy or needs a manual download. But if you've ever burned a ten-minute render on a take where the camera slid off the subject in the first two seconds, you'll take that trade without thinking twice.

CategoryCGlide

Inputs (8)

NameTypeDefaultDescription
modelMODEL
modeCOMBOanimatedAnimated plays the whole shot on this node as it denoises. Contact sheet tiles the frames into one still. Stock is ComfyUI's frozen first frame.
fpsFLOAT24.00.5–24Playback rate for the preview. H3 stores one latent frame per ~3.35 output frames, so 7.2 fps is every frame there is - anything higher simply uses them all.
columnsINT41–12Tiles per row in the contact sheet.
every_n_stepsINT21–20Redraw at most this often. Raise it if previewing is costing you sampling time.
max_resolutionINT512256–4096Longest edge of the preview sent to the browser. Also decides how far the TAE decoder upsamples, so raising it costs decode time as well as bandwidth.
sampler_previewBOOLEANfalseAlso draw ComfyUI's single-frame preview on the sampler node. Off by default in animated mode, since this node already shows the whole shot.
decoderoptCOMBOlatent2rgblatent2rgb is H3's colour approximation - free, but mushy. A TAE gives real picture: put taeh3.safetensors in ComfyUI/models/vae_approx/ and pick it here.

Outputs (1)

NameTypeDescription
modelMODEL