VOID Text Encode (AP)

Let's be blunt: this node is a convenience wrapper, and the pack's own source says as much - it's "identical to CLIPTextEncode but grouped under AP/VOID for clarity." You could swap in a stock CLIPTextEncode and get the same conditioning. What this node actually buys you is a tidy graph and the pack's prompt conventions sitting where you'd expect them, right next to the rest of the VOID nodes.

But the underlying requirement is not a formality, and it's the thing worth understanding: VOID was trained on T5-XXL embeddings, and nothing else works. Feed it a clip from a CLIPLoader pointed at t5xxl_fp16.safetensors (in models/clip/) and you're fine. Feed it a CLIP L or G model and you get what the README calls "silent conditioning" - the sampler runs, nothing blows up, and the output looks like your prompt was ignored entirely. That's the most confusing failure mode in this whole pack because it fails quietly, so wire the T5-XXL loader first and double-check it.

The inputs are exactly two: text (multiline, so write a full sentence, not tags) and clip. Output is a single CONDITIONING that feeds the positive input of VoidSampler - run a second one for negative (quality artifacts, watermarks, "low quality, blurry" and so on). The prompt tip from the README is the single most important sentence in the pack, so treat it as law: describe the scene after the object is removed, what the clean background should look like - not the removal itself. So "wooden table, warm lighting" is right; "a person being removed from the scene" is wrong. VOID fills the gap by generating the background you describe, and if you describe the removal, it has nothing useful to generate. This mirrors the inpainting wisdom from the KB - you're prompting what should be there, the way you'd prompt a fill region.

A couple of implementation details worth knowing because they save you from hunting for nodes that don't exist. Prompts are internally padded or trimmed to 226 tokens - the model's training sequence length - so you do not need a padding node in your graph. And the encode is genuinely the stock CLIPTextEncode path: tokenize, encode_from_tokens, return conditioning. Nothing exotic.

Install is the shared pack story: ComfyUI Manager → search "AP Netflix VOID", or git clone https://github.com/adampolczynski/AP_Netflix_VOID into ComfyUI/custom_nodes, restart, no extra pip dependencies. The heavy lifting is the model downloads - T5-XXL plus the VOID checkpoint plus the CogVideoX VAE. And one last honest note: this is a beta wrapper around a research model, so even a perfect prompt won't guarantee a perfect removal - the quadmask is doing most of the work, and you'll be iterating there. But get the T5-XXL and the after-state prompt right, and you've removed the two silent killers from the equation.