Cine con IA · Optimizador H3

Local video generation is a VRAM argument you have with your own graphics card, and H3's full weights are reportedly around 42.5 GB with no published consumer hardware floor. So the settings question - how many steps, how much chunking, whether a second pass fits - is usually answered by finding out the hard way.

This node tries to answer it before the render instead of after. It detects your GPU, applies a profile for your VRAM tier, checks the requested resolution and frame count against a memory model, and hands the decisions to the rest of the graph as one config object.

What it decides, and how

The profiles are the core of it. Each one sets a step count (20 across the board), the sampler and scheduler (res_multistep / simple), the attention and FFN chunking, whether the refine pass runs, and its scale and step profile. 8 GB is the survival setting - 32/32 chunking, refining off. 16 GB is the balanced one: 20/20 chunking with a 1.27× second pass. 24 GB steps down to 16/16 and a 1.44× refine. AUTO reads your card's total VRAM and picks the nearest profile - total, not free, and deliberately so, because ComfyUI's cache state shouldn't change your preset between runs. Pick a preset above what's in the machine and it tells you.

The four 0–100 sliders all do one concrete thing each, which is more than you can say for most "quality" knobs:

• calidad moves the first-pass step count, roughly one step per ten points (so 70 → 100 buys you three more steps on a 20-step profile);
• detalle picks the refine step profile: 3, 4 or 5 steps;
• resolucion nudges only the refine scale, never your requested width and height;
• ahorro_vram adds up to twelve extra chunks on top of the profile's base, so at its default of 50 you get six more than the profile number - worth knowing, because the chunking integers the node emits will be higher than the profile table suggests.

Then the Memory Planner classifies the request. It scores your width × height × frames against a reference load of 416×736×192 (the author's measured config), multiplies by the square of the refine scale when the second pass is on - refinement cost grows with area, so a 2× refine is roughly 4× the work - discounts for chunking, and compares that to a per-card capacity table. The verdict is SAFE, TIGHT or RISKY, with a recommendation attached: more chunking, less refine scale, a shorter clip.

None of this is a benchmark, and the README says so. The concrete numbers come from real renders on one 16 GB RTX 4060 Ti, including an OOM at a ~20.9 GB peak when a 416×736×192 job was pushed through a 2× refine pass with no chunking - which is why that combination is classified as RISKY. 8, 12, 24 and 32 GB cards have no measured data behind their rows. Treat it as a sensible starting position, not gospel.

Inputs and outputs that matter

You set width, height and frames to match what the rest of the workflow generates, plus modo and perfil. Advanced ignores the sliders and exposes the real thing - steps, sampler, scheduler, denoise, both chunking values, refine scale and refine step count - which is where you go the moment a profile is wrong for your card. The optional bottom three are about progressive sampling: muestreo (Normal / Progresivo) plus a transition step count and an initial scale, and remember the tooltip's promise, which is correct: progressive saves time, not VRAM.

Outputs are where the wiring gets interesting, because CINECONIA_H3_CONFIG is consumed by exactly one node in the pack, the Render optimizado H3 sampler. Everything else the rest of your graph needs comes out as ordinary typed values so you can wire them by hand:

• trocear_atencion and trocear_ffn (INT) into the Load model node's chunking fields - that node applies the chunking for real, and notes in its own output whether KJNodes was available to do it;
• refinar (BOOLEAN), escala_refinado (FLOAT) and pasos_refinado into Escalar y Refinar, which respects a value arriving by cable and greys out its own buttons;
• perfil_activo and info (STRING) wherever you want to read them.

Install

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

Restart and it shows up as "Optimizador H3" under Cine con IA/H3. No Python dependencies are installed by the pack. ComfyUI Manager may not find it - registered as cine-con-ia, versions pending review - so clone it or grab the release ZIP. The interface is Spanish throughout.

Two optional companions matter. ComfyUI-KJNodes is what actually performs the attention and FFN chunking at load time; without it the loader continues happily and just notes "attention without chunking", so if your VRAM numbers don't move, check that first. And the progressive path needs SelfLift plus an H3 latent upscaler in models/latent_upscale_models.

Known rough edges

The 8 GB profile turns the refine pass off, and if you asked for refining you'll get a note in the info line about it - Advanced is how you force it anyway, at your own risk. Below roughly 7.75 GB of VRAM the planner stops guessing and reports RISKY, because there's no profile underneath it. On a machine with no detectable CUDA, AUTO can't resolve and the planner goes to UNKNOWN asking you to pick a VRAM size manually.

And the licence caveat over this entire pipeline: H3's open weights are under a community licence that excludes the US, EU, UK and South Korea. Outside those regions this is a genuinely good local setup; inside them, the hosted API is the licensed path, and no amount of chunking changes that.