Cine con IA · Cargar modelo

MiniMax H3 is a 33B omni-modal model with a ~21 GB int8 checkpoint and a text encoder that eats another 15 GB. The launch thread's mood was "surely this shouldn't need a supercomputer," followed immediately by a 3060 owner asking whether "slow" meant minutes or geological timescales. So the interesting part of Cine con IA · Cargar modelo (CineCargarH3) isn't that it loads a model - it's that it already knows which memory tricks this particular model responds to.

What it loads

Four file pickers, straight from your ComfyUI folders: modelo (from models/diffusion_models), codificador_texto (models/text_encoders), vae_video and vae_audio (models/vae). Yes, two VAEs - H3 generates sound natively, so the audio VAE is a real component, not a leftover. On families without audio, the node loads the audio VAE slot with the video VAE and says so in info rather than opening a second file nothing will use.

perfil is what makes this a multi-family loader. Set it to MiniMax H3, LTX-2.5, Wan 2.2 or Hunyuan 1.5 and it changes the CLIP mode, picks the right sigma-shift node, and decides whether the VRAM patches are even applicable. Set it to Personalizado and it does nothing up front - it guesses the family from your filenames and tells you which one it guessed. That guess is decent, not magic.

Five outputs: model, clip, vae_video, vae_audio, and an info string. The two VAEs fan out to Escena and Salida; model goes to both Render and Escalar y Refinar.

The knobs worth touching

trocear_atencion and trocear_ffn both default to 16 and both mean the same thing: split the heavy math into N passes instead of one. The tooltips are honest about the trade - identical result, lower peak VRAM, somewhat slower. There is no quality cost. If a render OOMs, this is the first lever, not the last.

Here's the catch you'll meet if you're not on H3: the chunking only applies to MiniMax. On other families the node logs "troceo omitido (solo sirve en MiniMax)" and moves on. It's not broken, it's not applicable.

Both patches come from ComfyUI-KJNodes (MiniMaxLowVRAMAttention and MiniMaxChunkFeedForward, the latter called with a 2048 sequence threshold). If KJNodes isn't installed, the loader keeps going and writes "FFN sin trocear (falta KJNodes)" into info. Nothing fails. You just don't get the relief.

shift_video (6.0) and shift_audio (3.0) are H3's sigma shift. This is the flow-matching generation's rough replacement for picking a scheduler, and it's model-specific - the pack ships per-family defaults and applies them with whichever shift node your ComfyUI has.

vista_previa decodes the in-progress render with a tiny VAE from models/vae_approx so you can watch the shot evolve. Cheap, and worth leaving on while you're dialling in a prompt.

Up to four LoRAs chain in slots lora through lora_4, applied in order and before the memory patches, so the patches wrap the already-modified model. Pick one and the front end proposes 0.75 as a starting strength. The tooltip names one specifically for H3 skin and face realism: h3-realism-people-t2v-i2v-r2v.safetensors.

Install

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

Restart ComfyUI. The pack declares zero Python dependencies - it uses only what ComfyUI already ships. You do want a recent ComfyUI (0.34.2+) with native H3 nodes, the model files themselves, and KJNodes if you want the chunking and live preview.

The gotcha that will bite you

Wan 2.2 14B is two models - a high-noise and a low-noise half that hand off mid-sampling. This loader has one model slot. On Wan you need two Cargar modelo nodes and a sampler that switches between them; the pack is upfront that the full pipeline today is H3, and the other profiles are for loading and prompting. Don't buy a 30 GB download expecting one node to wire it.

Second: an H3 int8 checkpoint and its nvfp4 text encoder are two separate files with two separate budgets. If the pair doesn't fit, quantize the encoder harder before you touch the diffusion model.