VOID PQ5 Encode Prompt

The PQ5 sampler doesn't accept text - it accepts text embeddings. This node is the bridge: it runs your prompt and negative prompt through the VOID model's T5 text encoder and hands the resulting tensors to the sampler. It's the "what should be there instead" half of the equation, and it's deliberately shaped like a CLIP text-encode node, so it'll feel familiar the moment you look at it.

Worth knowing up front: in VOID's typical wiring your prompt is basically "empty background" - the default - because the quadmask already tells the model where the object was and what to repair. The prompt isn't describing the whole output; it's nudging the fill. You'll often run this node without touching a single field.

How it works

It calls the pipeline's encode_prompt with the standard CogVideoX/T5 machinery. The one behavior worth knowing: do_classifier_free_guidance is set to true only when guidance_scale > 1.0. That flag determines whether the negative prompt gets its own separate embedding path - at the default guidance of 1.0 (the model family's tuning), there's no CFG, so the negative prompt is effectively inert. If you raise guidance above 1.0, the negative prompt starts mattering, so put something real in it.

The inputs that matter

• model - from VOID PQ5 Load Model.
• prompt (STRING, default "empty background") - the edit intent. For object removal, leave the default; for pass-2 refinement you might describe the scene you want to preserve.
• negative_prompt (STRING, default empty) - only consulted when CFG is on (guidance > 1.0). The upstream VOID config ships a default negative about low quality/watermark/distortion; you can paste something like that in.
• guidance_scale (FLOAT, default 1.0) - this controls the CFG flag as much as the strength. Keep it matched to what the sampler uses.
• max_sequence_length (INT, default 226) - T5 sequence budget. The VOID defaults are set for this model's training; you rarely need more, and more costs VRAM.

Outputs:

• prompt_embeds and negative_prompt_embeds (both PQ5_PROMPT_EMBEDS) - wire both into the sampler's corresponding inputs. Keep the pairing intact: they were encoded together under the same CFG decision.

Install & notes

Standard pack install (Manager → "ComfyUI-NetflixVoid", or clone into custom_nodes). No downloads of its own - the T5 encoder loads as part of the pipeline bundle from VOID PQ5 Load Model, which needs the text encoder folder at ComfyUI/models/text_encoders/void (that's the usual missing-piece error). Runs on first use; afterwards the pipeline cache makes re-encoding cheap.

Gotcha: because guidance default is 1.0, if you set guidance to 7 on the sampler but leave this node at 1.0, the embeds were encoded without CFG while the sampler tries to run with it - the flags are encoded into the embed path, so keep the two guidance values in sync.