Cine con IA · Salida

The unglamorous end of the pipeline: turn a latent back into pictures and sound, optionally smear in some extra frames, and hand out something a Save node understands. Cine con IA · Salida (CineSalida) does the job of five core nodes - video decode, audio decode, interpolation model load, frame interpolation, and video assembly - and its video output drops straight into Save Video.

The mechanism

VAEDecode on the vae_video, VAEDecodeAudio on the vae_audio (H3's latent carries both, which is why the pack makes you load two VAEs), then optional frame interpolation, then CreateVideo with images, fps and audio assembled into a single VIDEO object.

interpolar is a four-item list: no · 24 fps, x2 · 48 fps, x3 · 72 fps, x4 · 96 fps. It's a multiplier, not a target, and the tooltip is precise about what it does - it invents intermediate frames so motion looks smoother and does not change the duration. Nothing gets longer. The same four seconds just has twice as many pictures in it.

modelo_interpolacion reads models/frame_interpolation. fps_base defaults to 24, which is H3's native rate, and the tooltip warns that changing it speeds the clip up or slows it down rather than re-timing the generation. If you want 48 fps output, the correct move is x2 interpolation at fps_base 24 - that's why the interpolate options are labelled with their resulting frame rates.

Five outputs: video, fotogramas, audio, fps, and info. Wire video to Save Video and you're done. fotogramas is the decoded IMAGE batch - handy for frame grabs, for feeding a still into an image model, or for an image-based Save node. audio is the AUDIO track on its own, worth having if you want to master it separately. fps is the final rate including interpolation.

The bits that actually trip people

Interpolation needs a model, and its absence is quiet. Turn interpolation on with no modelo_interpolacion selected and you get your frames back uninterpolated, with "falta el modelo de interpolacion" buried in info. No error, no crash, no interpolated frames. Check info (it prints the frame count, final fps, real duration, and what it actually did) before you conclude interpolation "doesn't work on my machine".

Audio decode is meaningless off H3. On non-H3 profiles, Cargar modelo loads the audio VAE slot with the video VAE, so decoding "audio" from a silent latent gives you a track of nothing useful. H3 and LTX-2.5 generate audio; Wan 2.2 and Hunyuan 1.5 don't.

Interpolation is a smoothing trick, not a quality upgrade. It won't fix a mushy shot or a broken camera move - it'll make a mushy shot smoothly mushy. Get the first pass right, then interpolate.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/chaLords/ComfyUI-Cine-con-IA.git

Restart ComfyUI, look for the Cine con IA category. The pack ships no Python dependencies of its own; the interpolation model is a separate download that goes in models/frame_interpolation, and the pack's Modelos node will tell you which files each family expects and where. If you don't intend to interpolate, you can skip that download entirely - this node runs without it.

Traps

The real one is expectations. Everyone wants to solve short clips by interpolating, and it doesn't work that way: a 5-second clip is still 5 seconds and still has the same amount of motion, just more samples of it. The field's actual constraint is clip length, which is why H3's trained window is 124–362 frames and why the pack nudges you about durations under 5.2 seconds in Duración.

Also, double-check fps on the way out. Two workflows, same clip, one at 24 and one at 48 after interpolation: if you then feed the 48 fps version into a tool expecting 24, you've just made a slow-motion video. It's an easy mistake and it looks like a model failure.