Four nodes: one that builds the whole prompt and conditioning, one that lets you watch the shot while it samples, one that writes the file, and one that joins a continuation onto the clip it continued from.

I make short films with these. Everything here exists because I needed it, and the defaults are what I actually use.

<a href="https://youtu.be/aEUM6sihSdI"> <img width="500" height="750" alt="Capture d'écran 2026-08-17 221233" src="https://github.com/user-attachments/assets/d9984423-f965-470b-80c8-3faebffa079b" /> </a> THE RECITATION — a short film made with these nodes ---

Install

Search for CGlide in ComfyUI Manager, or:

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

Restart ComfyUI. The nodes appear under CGlide in the node menu.

Update ComfyUI while you're at it. Honestly, half the H3 problems people post about are an old ComfyUI, not the model. The low-VRAM path below needs 0.31.0 or newer.

Requirements

PyAV — reading reference videos and their soundtracks. Usually already there with ComfyUI.
torchaudio — resampling reference audio.
ffmpeg on your PATH — Glide Video shells out to it. Nothing else uses it.

Models

Everything comes from the Comfy-Org MiniMax H3 repo:

the reference-to-video checkpoint (and/or first-last-frame) → models/diffusion_models/
the text encoder → models/text_encoders/
the video VAE → models/vae/
the audio VAE → models/vae/

Less VRAM: grab the w4a8 model from Kijai's experimental repo — about 11.8 GB instead of 21 — and the int8_convrot video VAE from the same place. Same workflow, same prompts, you only swap the loaders.

Careful, this one is nasty: if your ComfyUI is too old for w4a8 you get a black video with sound. No error, nothing. If you see that, it's the version, not you.

H3 Studio

One node instead of the pile of loaders and text boxes. Nine image slots, three video, three audio, the prompt, and the checks — all in one panel.

It does not touch sampling. Same sampler, same steps as a normal H3 workflow. It builds the prompt and the conditioning, that's it.

Inputs: clip, vae, audio_vae, and first_frame / last_frame for the first-last-frame family.

Outputs: positive, latent, width, height, length, seconds, overlap_frames, source_video, guide_frames.

The last ones are for chaining — see below. Everything else you wire the way you would with the native H3 nodes.

What's in the panel

Both model families in one node. First / last and Omni references is a switch at the top; the accent colour follows so you always know which one you're in.
Drag files onto slots. Or paste — Ctrl+V drops a screenshot straight into the slot under your pointer. Type @ in the prompt and pick a reference to drop its tag in.
Several images at once. The nine image slots take a whole batch, from the picker or from a drop. Drop onto a particular card and the batch fills from that card forward; drop onto the rack and it fills whatever slots are empty and leaves the filled ones alone. A cast usually arrives as a folder, so one file per gesture was the wrong unit.
Drag a picture onto another slot to swap them. Order comes out of your file manager and it's rarely the order you want. The @tags travel with the picture — move @image1 to slot 3 and the prompt is rewritten to @image3, so every sentence you wrote about it still describes it. Hold Alt while releasing to move the picture and leave the prompt alone, for when the prompt is already right and it's the pictures that are out of order.
"Sent to the encoder as". The model sees Picture 1, Picture 2 and so on, and the number comes from which slots are filled, not from the slot number. Leave a gap and your Picture 3 is not the one you think it is. This strip shows you what actually goes out.
Trim bars with playback. Drag anywhere on the track to slide the window and the picture follows, so you can hunt for the right moment. Pick a length, slide it, done.
Prompt check. Amber when the soundscape doesn't cover the whole clip, or when nothing says what the mouths are doing. Those two are most of the mumbling problems. The sound chip also counts words against the clip length, so a line that can't fit in the beat shows up before you render it.
Shot timeline. Type a shot marker and the bar splits. One colour per shot, the first words of each shown in the band, and dragging a boundary rewrites the timestamp in the prompt. Grey means nothing is scripted there — and unscripted time is exactly where H3 invents sound.
Projects. Several clips in one file, click to switch, autosaved. A new clip inherits your references so you're not reloading the same five images every time. Save packed writes a .h3proj.zip with the media inside — my whole twelve-clip film is about 42 MB.

Glide Preview

ComfyUI's preview shows you one frozen latent frame while it samples. So you sit there for ten minutes looking at a still and you have no idea if the motion is working.

Drop this between your model and the sampler and you watch the shot move as it denoises.

mode — animated, contact sheet, or stock.
decoder — latent2rgb is free but mushy. Put taeh3.safetensors (from madebyollin) in ComfyUI/models/vae_approx/ and pick it here for real picture. Restart before it shows in the list.
every_n_steps — raise it if previewing is costing you sampling time.
max_resolution — costs decode time as well as bandwidth.

Honest limit: it caps around 7.2 fps and I can't fix that. H3 stores about one latent frame per 3.35 output frames, so that's every frame there is. It's a preview, not the render — but it's enough to kill a bad take at step four instead of at minute ten. If the decoder fails it falls back to the stock preview and your run keeps going.

Glide Video

Frames and audio into one file, in one ffmpeg pass, no second file to mux.

Presets from very compatible up to 4:4:4 10-bit and lossless FFV1. The 4:4:4 10-bit ones are visibly better on H3 output — less flat colour, more nuance. Read the tooltip, each preset says what it's for.
fps is playback rate only. Setting it wrong changes speed, it does not add or drop frames. H3 is 24.
save_metadata embeds the prompt and workflow in the file comment. Drag-drop restore into ComfyUI is reliable for MKV, hit or miss for MP4.
fallback_on_failure retries with a more compatible preset instead of throwing away a finished render. It says so loudly when it does.

The 4:4:4 and ProRes presets don't play in a browser, so the node writes a small playable copy next to the real file just for the preview. The real file is the one in your output folder — the copy is only so you can watch it in the node.

Continuing from your last render

There are two ways to carry one clip into the next, and they do different jobs.

As a guide — the CONTINUE FROM slot. Frame-exact: the model picks up the motion and the sound where the last clip left off. For continuing a shot.

As a look reference — an ordinary video slot. Same light, same colour, same room, but the camera is free to be anywhere. For the next shot in the same scene.

Both have a last render button that grabs the newest video in your output folder, so you're not going to the output folder and dragging files around. The copy happens server side, so nothing big goes through your browser. Hover before clicking and it tells you which file it's about to take and what it is.

On a reference video slot the button does the look-reference version: sound stays off — same room doesn't mean same soundtrack, and an @videoaudio tag would pull the previous clip's mix into this one — and it scans the clip for the window that reads clearest, then writes that window into the trim where you can see it and drag it if you disagree. Hold Alt to take the whole clip untrimmed and place the window yourself.

The window picker is a score, not a promise. If nothing in the clip reads well it says so in amber and you pick one yourself.

Glide Join — extending a clip

H3 gives you a few seconds at a time. This is how you get more.

H3 Studio's CONTINUE FROM slot anchors the tail of your previous clip, so the model continues the motion and the sound instead of starting fresh. Glide Join then assembles the result onto the source.

Wiring:

H3 Studio ──positive──> sampler ──> VAE Decode ──images──> Glide Join ──> Glide Video
          ──overlap_frames──────────────────────────────>
          ──source_video────────────────────────────────>
                                    VAE Decode Audio ──audio──>

Wiring overlap_frames and source_video across means the window length is typed in one place. Change the window in the node and the join follows.

Widgets:

seam_mode — early_cut keeps the continuation's version of the overlap and colours it onto the source. One continuous model trajectory, so motion carries best. early_scurve is the same join ramped over a few frames, for when the speed change at the join is visible. hard_cut keeps the source's own frames instead.
seam_blend_frames — ramp length, only read by early_scurve.
match_levels — the model re-renders the overlap in its own colour, so this fits gain and offset per frame against the source and holds the last fit afterwards. Leave it on.
match_tail — H3 rounds its audio grid up, so every clip carries about 8 ms more sound than picture. That grows at every join. Leave it on.

Leave the CONTINUE FROM slot empty and the whole thing is a normal generation — H3 Studio sends nothing, Glide Join passes the frames straight through. One workflow for both, no rewiring, no bypassing.

Two things worth knowing:

The window is locked to the tail of the clip, on purpose. A window from the middle produces material that follows on from the middle, and splicing that after the last frame skips everything between — it reads as a hard cut and no seam mode can fix it.
Colour drift down a chain is real but small when the window is a true tail. If you see a step at the join it's more likely to be motion than colour.

The chroma badge — read this one

There's a little badge on the continuation waveform. Green for 4:4:4 and 4:2:2, orange for 4:2:0. It's not decoration.

Here's what happened. I had a continuation that kept coming back slightly out of sync — Glide Join kept printing frame correspondence was off by -2. So I went looking at the anchor window, the frame counts, the way the tail is read. Wasted a good while in there.

It was the codec. I was rendering AV1, which is 4:2:0 — three quarters of the colour is gone before the model ever reads those frames. Weaker guide, less exact anchor. I rendered the same pair again as H.264 4:4:4 and the correction line just disappeared. Then ProRes, which is 4:2:2 — also clean. So the line sits between 4:2:0 and 4:2:2, and anything above that is fine.

So: don't chain from a 4:2:0 file. The badge tells you what's in the slot before you render.

This is about the guide path only. A look reference carries no timing, so there's no badge on a reference video slot — the same file that's a bad guide is a perfectly good reference.

Render all — chaining a whole project

Doing the above by hand for five clips is a lot of dragging files around. So this does the loop for you.

Open the Project panel, pick a mode at the bottom, hit Render all:

chained — each render goes into the next clip's CONTINUE FROM. Glide Join stitches as it goes, so the last file is the whole thing. There's no separate assembly step.
carry look — each render goes into the next clip's video reference slot instead, with the window picked for you. Separate clips, same room. This is the one I use most: it holds a location across a scene without pinning the camera to where the last shot ended.
separate — same loop, no continuation. Just queues all your clips one after another.

While it runs the node is locked and shows which clip it's on. Stop finishes the clip it's rendering and stops there — everything already done stays done. And if a render fails it stops, tells you which clip, and keeps the finished ones, so you can pick it up from there instead of starting over.

If you don't like the result, Revert in the Project panel puts the project back exactly as it was before the run.

Pick chained with a 4:2:0 preset selected in Glide Video and the node asks you first — every link in the chain pays for that, not just one join. Carry look doesn't ask, because a reference doesn't care.

Skipping a clip

Each row in the project list has a dot. Click it and that clip drops out of the run — the row dims, the dot hollows, and Render all passes over it. The clip stays in the project and in the saved file. Click again to bring it back.

This is for when one shot in the middle of a finished sequence needs another pass and you don't want to re-render the ten that already work.

Two limits worth saying out loud:

Each link re-encodes the whole video so far. Clip 5 means clip 1's frames have been through the encoder five times. On H.264 4:4:4 10-bit at crf 12 I can't see it, and FFV1 is lossless so it's free — but don't run a long chain on a share preset. You'd be stacking generations and handing every link a 4:2:0 guide.
One mode for the whole project, for now. Per-link — where each gap between two clips picks its own mode, so a cut inside a scene and a cut between scenes aren't forced to be the same thing — is what I actually want, and it's coming. This is the simple version first.

Careful with this

A .py change needs a full ComfyUI restart, not a browser refresh.
If a node's widget list changes shape between versions, delete and re-add the node. ComfyUI matches saved widget values by position and will smear old values across new widgets.
The node is new. There will be bugs — open an issue and tell me what you did, that's genuinely more useful than a star.

Credits

MiniMax for H3, and the ComfyUI team for the core H3 nodes this builds on.
madebyollin for the TAE decoder that makes the preview readable.
Kijai for the quantised models that put this on smaller cards.

Development assistance from Claude (Anthropic).