Cine con IA · Escena

Cine con IA · Escena (CineEscenaH3) is where your text and your pictures become something the sampler can use. In the pack's own graph it's the middle node: sizes and duration come in one side, prompt and model come in the other, and it hands positive, positive_escalar and latent downstream. Underneath, it replaces what would otherwise be five core nodes - the reference-to-video conditioning, up to three reference image slots, the guide-image resize, and the guide anchor - in one box.

What it builds

Give it a clip, the two VAEs, a prompt, and a width/height/length, and it produces the H3 conditioning plus an empty AV latent sized to your frame count. If you connect reference images, they're attached positionally as <Picture 1>, <Picture 2>, <Picture 3> - and this is the detail people miss: the order is the wire, not the text. If your prompt calls the second image <Picture 2>, the image you plugged into referencia_2 had better be the right one.

tamano_referencia is a real quality lever. max reads the reference at a 2048-pixel short side: noticeably better face and texture retention, slower. match downsizes it to the video's own area: faster, and you will lose identity. Default is max and for anything with a face in it, keep it there.

The two guide modes

modo_guia has two settings and they are not a preference, they're different mechanisms.

exacta · fija fotograma 0 center-crops your imagen_guia to the video's exact dimensions and anchors it as an exact latent keyframe at fotograma_guia (0 = the first frame). The opening frame matches your image, pixel for pixel. Use it when the first frame must be right.

flexible · prioriza camara doesn't anchor anything. The image is added as just another reference, taking the next <Picture N> slot, which means it guides identity and style without freezing the composition - and gives large camera moves room to happen.

There's a consequence worth knowing before you use flexible for a big orbit: if the image you feed it shows the whole set, it acts as a background plate, and the author's measurements put background plates at roughly halving the camera move. One identity plate per character, describe the location in text.

positive vs positive_escalar - read this bit twice

The exact-mode anchor carries a latent of the first pass's size. In the second pass the latent has been scaled up, so that anchor no longer fits - and you get a shape mismatch or a broadcast error, which is one of the more annoying ways to lose twenty minutes.

So the node outputs two conditionings. positive goes to Render. positive_escalar - the same conditioning without the guide anchor - goes to Escalar y Refinar. Wire the wrong one and the pack's refine node will actually tell you, in its error message, to go back and connect positive_escalar instead. latent goes to Render, and Render's latent goes to Escalar y Refinar.

Install

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

Restart. No Python dependencies beyond what ComfyUI already has, but this node needs ComfyUI's native MiniMax H3 core nodes to exist - if they don't, it fails with a clear "missing MiniMaxH3ReferenceToVideo" rather than something cryptic.

Traps

length hides a warning you should watch: below 124 frames the node flags that you're under H3's trained 124–362 range. Camera moves are the first thing to disappear.

And remember fotograma_guia counts frames, not seconds. At 24 fps, "anchor my guide at the interesting bit" means frame 96, not 4.