Cine con IA · Modelos

Here's the part of local video nobody writes tutorials about: the downloads. Four files, three different destination folders, and a YouTube description with a dead link. Cine con IA · Modelos (CineModelos) shows you what each model family needs, marks what you already have, and downloads the rest into the folder ComfyUI actually reads - with progress and resume.

What it does

Three inputs. familia filters the list: Todas, MiniMax H3, LTX-2.5, Wan 2.2, Hunyuan 1.5. incluir_opcionales adds the variants (turbo LoRAs, ControlNet, upscalers) on top of the essentials. incluir_efectos adds H3's ten ~1 MB effect embeddings - bullet time, kiss cam, and so on. One output, info, and the node is an output node, so the list prints on every run.

For MiniMax H3 the essentials are honest about the size of this hobby: the int8 reference-to-video checkpoint (~21 GB) into models/diffusion_models, the nvfp4 Qwen3-VL text encoder (~15.7 GB) into models/text_encoders, a video VAE (~2.8 GB) and an audio VAE (~0.6 GB) into models/vae. That's roughly 40 GB before you've rendered a single frame. The optional turbo LoRA is another 2 GB, a first/last-frame variant is another 21, and there's an optional Singularity ref2va checkpoint - a complete ~21 GB model, not a LoRA - for people chasing dynamic motion.

How it downloads, and why that's designed this way

The catalog lives inside the node's source, not in the workflow. The author's comment says why, and it's a genuinely good reason: a workflow is a file anyone can publish and edit, so if download URLs travelled inside it, someone else's workflow could carry a button that fetches something else and writes it wherever. Here, the workflow only says which family you want.

The downloader is plain urllib: it checks each destination path, skips anything already present, writes to a .part file and resumes with an HTTP Range request, and can be cancelled. A Hugging Face token is picked up from HF_TOKEN, HUGGING_FACE_HUB_TOKEN, HUGGINGFACE_TOKEN, ~/.cache/huggingface/token or ~/.huggingface/token - never stored in the workflow. It's only needed for gated repos.

That's a live issue on LTX-2.5: that repo requires accepting terms on the web first. Hit a 401/403 and the node tells you exactly that, by name, instead of a bare HTTP code.

Install

The node is part of the pack:

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

Restart ComfyUI and look for Cine con IA → Modelos. It uses only the Python that ComfyUI already ships - no requirements.txt, no pip step, no build. The download routes are registered on ComfyUI's own web server, so if the info panel shows up but the download button does nothing, restart and check the console for "rutas de descarga listas".

Worth knowing while you're installing anything by name: Manager installs run a repository with your user account, and the Comfy Registry had a stretch in 2025 where three malware-carrying packs sat live and were installed 790 times before they were pulled. A pack that downloads multi-gigabyte weights for you is convenient and also exactly the category of code worth reading the source of once. This one is a single 2,376-line nodes.py; you can skim the catalog and the downloader in ten minutes.

The gotchas

Where the file lands is decided by the node, not by the repo. Each catalog entry carries the destination folder - diffusion_models, text_encoders, vae, loras, vae_approx, latent_upscale_models, frame_interpolation, embeddings - and the on-disk filename is the last segment of the repo path. If you renamed things to your own taste, the node will cheerfully download a second copy.

Refresh after downloading. Model selectors in ComfyUI cache their lists. Download the optional Singularity checkpoint and it won't appear in Cargar modelo until you refresh the node definitions or restart, and you have to keep the MiniMax H3 profile selected when you pick it.

One licence worth reading before the download, not after. H3 ships under the MiniMax H3 Community License, whose applicable territory excludes the EU, UK, Republic of Korea and the US - users there aren't licensed to run the local weights or use their outputs. The hosted Hailuo API is the licensed path in those regions. Your call what you do with that, but the pack won't tell you, and a 21 GB download is a bad time to find out.

Wan 2.2 needs two models. The 14B high-noise and low-noise halves are separate downloads, and Cargar modelo has one model slot - the node's own notes say you need two loader nodes. Budget the disk accordingly.