Cine con IA · Render

If you have built a video graph by hand, you know the shape: a KSamplerSelect dangling off to the left, a BasicScheduler, a BasicGuider, a RandomNoise, all feeding a SamplerCustomAdvanced, all wired to the same model and latent. It works. It's also five boxes whose only job is to spell out settings you'd rather see in one place. Cine con IA · Render (CineRenderH3) is that consolidation - the node does not reimplement sampling, it calls the core nodes for you.

What's actually happening

Under the hood the node builds the exact same five-node rig: BasicScheduler with your scheduler and step count, KSamplerSelect with your sampler, BasicGuider on your conditioning, RandomNoise on your seed, then SamplerCustomAdvanced. The advance-mode sampler stack is not modified, just hidden.

That matters for two reasons. First, your result matches the stock graph, so anything you know about samplers and schedulers still applies. Second, if the ComfyUI install is too old to have those nodes, you get a plain "missing the advanced sampling nodes" error rather than a mystery.

The inputs

model, positivo and latente are the wire-ins - model from Cargar modelo, positive from Escena, latent from Escena. No negative conditioning input, and that's not an oversight: H3 doesn't take one. The pack's Prompt node emits a negative output because LTX, Wan and friends use negatives, but the H3 pipeline doesn't have anywhere to plug it.

pasos defaults to 8. That's the H3 baseline. Load the turbo LoRA in Cargar modelo and you're looking at 4 steps instead - noticeably faster, slightly softer. Which is a trade, not a free win.

sampler defaults to er_sde, pulled from whatever ComfyUI currently lists (44 entries on a current install). scheduler defaults to beta. Together those are the modern flow-matching pairing: the old "Karras everything" habit is actively counterproductive on straight-trajectory models, where an aggressive sigma redistribution is a distortion rather than a correction. If you change one thing here, change pasos; if you change two, experiment with the sampler and leave the scheduler alone.

semilla is a big int with control_after_generate - the dropdown that decides what happens to the number after a run. This is the single most reliable beginner injury in ComfyUI: by default it fires after the generation, so the seed in the box is the one that will be used next, not the one that just made the clip you liked. Flip it to fixed before you go hunting, or set widget control mode to Before in ComfyUI's settings and stop thinking about it forever.

denoise defaults to 1.0 and for a first pass that's where it belongs - the latent coming in is empty.

Outputs

latent and info. The latent goes into Escalar y Refinar if you're doing a second pass, or straight into Salida if you're not. info echoes steps, sampler, scheduler and seed, which is more useful than it sounds when you're comparing six renders and can't remember which was which.

Install

Part of the pack:

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

Restart. The pack itself installs no Python packages; the real requirements are a current ComfyUI with native MiniMax H3 support and the model files, which the pack's Modelos node will fetch into the right folders. Optional but recommended: KJNodes, which gives the loader its attention/FFN chunking. The render node works without it - it just isn't where that pressure relief comes from.

Traps

Eight steps is the baseline, but the pack's camera recipes were verified at 20 steps with res_multistep / simple and the turbo LoRA off. A 4-step turbo render suppresses camera movement almost entirely, so if a dolly or orbit came out as a still-ish shot, check whether you're running turbo before you blame the prompt.

And keep the seed fixed while you're tuning. Change one thing at a time - randomising the seed and changing the sampler together tells you nothing about either.