AD_LTX_vae_combine

This is the finish line of Apt_Preset's LTX_video chain. After AD_LTX_sampler does the actual diffusion, this node VAE-decodes the resulting image latent back to pixels, decodes the audio latent back to a waveform if you gave it one, and muxes everything into a proper VIDEO output - plus a couple of extras that make chaining multi-segment generations less painful.

LTX-2 and LTX-2.3 are built to generate synchronized audio and video in the same model, which is the whole reason this node has audio inputs and outputs sitting right next to the image ones instead of being a separate step. If you never fed audio in, those inputs are just optional and this node still works fine as a plain image-to-video decoder.

The inputs and outputs that matter

• img_latent / img_vae (required) - the sampled latent and its VAE, straight from AD_LTX_sampler.
• frame_rate (default 24) - the frame rate to encode the output video at.
• lowGpu_mode (default false) - flip this on if you're VRAM-constrained; it presumably decodes in smaller chunks at the cost of some speed.
• trim_latent_end (default 0, range 0–4096) - lets you drop trailing frames from the latent before decode, useful for cutting off any padding frames that got added to satisfy LTX's frame-count grid rather than leaving them in your final video.
• audio_latent / audio_vae (optional) - if you sampled audio alongside video, plug these in to get it decoded and muxed too.

Outputs: image (the raw decoded frames as an IMAGE batch), audio (AUDIO), video (VIDEO - the muxed, ready-to-save result), and last_image (IMAGE) - the final frame on its own, specifically useful for feeding back as the starting image of the next generation segment in an extension workflow, which is how you build longer clips out of several LTX passes strung together.

How to install it

Via ComfyUI Manager: search "ComfyUI-Apt_Preset". Manually:

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

Then install.bat (Windows, double-click) or pip install -r requirements.txt (Linux/Mac), and restart ComfyUI. No separate model download for this node - it decodes with whatever VAE came through the chain from AD_LTX_load_model.

Common issues & troubleshooting

Video output has extra frames at the end that don't belong. That's very likely padding LTX added to hit its frame-count grid. Use trim_latent_end to cut them before decode instead of trimming the finished video in post.

Out of memory on decode. VAE decode is often the actual VRAM spike in a video pipeline, not sampling. Flip lowGpu_mode on before you start cutting resolution or frame count elsewhere.

No audio in the output despite generating some. Check that both audio_latent and audio_vae are actually wired in - this node needs both to produce sound; if only one is connected, or neither, you'll silently get a video-only output with no error.

Chaining segments and the continuation looks like a jump cut. Grab last_image from this node and feed it as the conditioning image for your next segment's AD_LTX_load_model (or wherever your workflow takes an image-to-video starting frame) - that's the intended continuity hook. If you're regenerating from scratch each segment instead of chaining off last_image, that's the source of the visible cut.