VOID PQ5 Load Model

This is the entry point to the PQ5 inference stage - the part of the VOID workflow that actually regenerates your video. It takes a transformer checkpoint, a VAE, a text encoder, and a pile of repo-local configs, and assembles them into the CogVideoX-based pipeline that everything downstream (encode, sampler, decode) shares. "PQ5" is the naming from the upstream VOID runtime this pack vendors; the checkpoints come in pass-1 and pass-2 variants, and the README's two-pass workflow loads this node twice with a different checkpoint each time.

The important thing to know before you use it: this node has no fallbacks. The README says it outright. If your checkpoint, VAE, or text encoder paths are missing or wrong, it raises a hard error instead of silently loading something else. That's a feature in disguise - silent fallback in a model loader is how you get a "why does my video look like that" mystery at 2am - but it means you have to get the model files in the right places first.

How it works

Two dropdowns: checkpoint (populated from ComfyUI/models/checkpoints) and vae (from ComfyUI/models/vae). On load it resolves both paths, then verifies the text encoder directory exists at ComfyUI/models/text_encoders/void - that one isn't a dropdown, it's a fixed path, so it's the most common thing to be missing. Then it builds the pipeline: CogVideoX transformer loaded from your checkpoint with the repo-local pq5_assets configs (scheduler, tokenizer, model_index), the VAE weights, and the T5 text encoder. The whole bundle gets cached keyed by the file paths, so re-running with the same model is cheap. The pipeline uses model CPU offload with float8 quantized transformer weights by default, which is how a 5B-class video model runs at all on consumer GPUs.

The inputs that matter

• checkpoint - the VOID transformer checkpoint (the pass-1 void_pass1.safetensors-style file) sitting in ComfyUI/models/checkpoints. If the folder is empty the dropdown shows "<no checkpoints found>" and running throws.
• vae - a matching VAE checkpoint in ComfyUI/models/vae.

Output: a single model (PQ5_MODEL) wire that fans out to VOID PQ5 Encode Prompt, VOID PQ5 Encode Video, VOID PQ5 Sampler, and VOID PQ5 Decode Video.

Install & model setup

Pack install is standard (Manager → search "ComfyUI-NetflixVoid", or clone into custom_nodes). The model files are not. You need to place, by hand:

• Transformer checkpoint(s) → ComfyUI/models/checkpoints/
• VAE checkpoint → ComfyUI/models/vae/
• Text encoder folder → ComfyUI/models/text_encoders/void/

plus the pack's own pq5_assets folder stays where it shipped (configs, tokenizer, scheduler live there). Where you get the checkpoints: the VOID framework's Hugging Face space / upstream VOID-PQ5 releases - the pack itself downloads none of it. If you're running the two-pass workflow, you'll do this twice: load void_pass1.safetensors, run, then load a pass-2 checkpoint in a second VOID PQ5 Load Model and run the sampler+decode again.

Common issues

• "No checkpoint files found" / "No VAE files found" - you skipped the manual model setup. Get the files in the right folders.
• "Required text encoder directory not found: .../text_encoders/void" - the classic. The text encoder is a whole folder, not a single file; drop it exactly at that path.
• First load is slow - building a 5B video transformer with CPU offload takes a real while. The pipeline cache makes subsequent runs fast. If you switch checkpoints constantly, keep VOID PQ5 Unload Cache handy to clear the bundle cache.