CS SeC-4B Model Loader

Before you can run CS Video Segment (SeC-4B), you need this node to exist in your graph. CS SeC-4B Model Loader is the dedicated loader for SeC-4B - "Segment Concept," the OpenIXCLab video segmentation model that a chunk of the community is quietly swapping in for SAM2-based tracking. The short version: SAM2 tracks what an object looks like, and SeC-4B understands what it is. The model itself is 4 billion parameters of vision-language + video memory, so no, it doesn't fit in a LoRA slot - it's a real download.

The pack pulls its single-file weights from HuggingFace (VeryAladeen/Sec-4B), scans them out of ComfyUI/models/sams/SeC-4B, and hands you a ready-to-run model instance on a SEC_MODEL output.

How it works

The loader is a thin registry around the model. It picks the weight file, chooses a device, and - the nice part - caches the loaded model, so if your workflow has multiple segment nodes or you re-run a queue, you're not re-loading 7 GB of weights every time. On first run, if the chosen file isn't on disk, it auto-downloads it from HuggingFace to models/sams/SeC-4B. The BF16 file is about 7.35 GiB; the FP16 is smaller. Config files and tokenizer ship inside the pack, so no separate setup there.

The inputs that matter

• model_file - SeC-4B-bf16.safetensors (default) or SeC-4B-fp16.safetensors. The loader honors each file's native precision; for most people, just use the BF16 default.
• device - auto (picks CUDA if present), cpu, or a specific gpu0/gpu1. On CPU the weights get cast to float32 and FlashAttention is disabled - technically works, realistically slow. This model wants a GPU.
• use_flash_attn - default on; falls back to standard attention if your environment lacks FlashAttention.
• allow_mask_overlap - default on; lets multiple objects' masks overlap. Leave it.

Output

One output: SEC_MODEL, which plugs into the model input of CS Video Segment (SeC-4B). It's not a standard MODEL type - it's the pack's own type, so don't expect it to feed a KSampler.

Installing it

Pack install is the usual CineStyle story:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_CineStyle.git

or ComfyUI Manager → ComfyUI_CineStyle. The heavy part isn't the install - it's the weights. Either let the loader auto-download when you first execute (needs working HuggingFace access), or grab the file yourself and drop it here:

ComfyUI/models/sams/SeC-4B/SeC-4B-bf16.safetensors

Direct links are in the README. The pack's requirements.txt pulls transformers, peft, accelerate, timm and friends - this is a real dependency install, not a five-second one.

Where people get burned

• First execution looks hung. Downloading 7.35 GB with progress printed to console, then loading a 4B model, takes a while. Give it time; don't kill the process.
• Auto-download fails in restricted environments. If you can't reach HuggingFace, the download errors and you must fetch the file manually. Know your network before you queue.
• VRAM. This is a 4B model, not a toy. If the segment step OOMs, that's expected behavior, not a bug - see the segment node's offload_video_to_cpu and auto_unload_model options.