VOID Loader

VOID Loader is the middle of the pack's four-node chain - it takes the VAE and text encoder from the two loader nodes, adds the two VOID transformer checkpoints, and assembles the actual inference pipelines. Everything before it is plumbing; everything after it is the part that runs. Get this node right and the rest of the workflow is mostly a formality.

The thing you need to understand up front: VOID is a two-pass system, and this node loads both passes. Pass 1 does the base object-removal inpainting. Pass 2 is a warped-noise refinement pass that cleans up temporal consistency - the flicker and smear you'd otherwise get when the model has to invent a region across frames that don't quite agree. Netflix shipped them as two separate ~10 GB checkpoints, and the loader loads both and builds a pipeline for each.

What it takes in

Five inputs, and only the two model dropdowns are things you'll think about:

• pass1_model / pass2_model - dropdowns that list files in ComfyUI/models/void-model/. Right now there's exactly one option each: void_pass1.safetensors and void_pass2.safetensors.
• void_vae / void_text_encoder - the custom-type outputs from VOID Load VAE and VOID Load Text Encoder.
• memory_mode - this is the one that matters, because it's your VRAM budget.

Memory modes

| Mode | VRAM | Notes | |---|---|---| | fp8 + cpu offload (~12 GB) | ~12 GB | Default and recommended. fp8-quantised transformer, offloads to CPU between steps. | | cpu offload (~16 GB) | ~16 GB | bf16 transformer, CPU offload. | | full GPU (~40 GB) | ~40 GB | Everything resident. Fastest, and demands a serious card. | | sequential (~min VRAM) | minimal | Layer-by-layer offload. Slowest, but it'll run on almost anything. |

Start with the default and only change it if you're hitting OOM or you're on a 24GB+ card and want the speed. Full GPU at 40 GB is a luxury pick; on a 4090 or better you'll notice, on a 12GB card it's a hard crash.

The two gotchas that actually bite

The model files are ~20 GB and are not auto-downloaded. There's no "download" button. Grab void_pass1.safetensors and void_pass2.safetensors from netflix/void-model and put them in ComfyUI/models/void-model/ - the pack creates that folder for you on install. If either file is missing, the loader throws FileNotFoundError and prints exactly which path it expected. This is the most common "VOID doesn't work" report, and it's always a download problem.

The diffusers version check is a hard stop. VOID is built on CogVideoX, which needs diffusers>=0.33.1 - newer than stock ComfyUI's 0.29.x. The pack checks your version at import and refuses to start with an error telling you to pip install diffusers>=0.33.1. Do that, then be ready for the follow-up: bumping diffusers can break other custom nodes that pin older versions. The README's advice is sound - pin the offenders individually. The one-time cost here is real but the fix is mechanical.

Also worth knowing: transformers and assembled pipelines are cached by file path and memory mode, so re-runs after the first load are fast, and switching memory modes builds a fresh pipeline instead of corrupting the old one's CPU-offload hooks.

The output

One VOID_PIPELINE output, wired into the void_pipeline socket on VOID Inference. That's it. If you're building the VOID workflow from scratch the wiring is: Load VAE + Load Text Encoder → Loader → Inference, with the Quadmask Builder feeding Inference's mask input.

The setup cost is the ~20 GB of weights and the diffusers dance. Once that's done, this node is a set-and-forget - which is exactly what you want from the heaviest node in the graph.