CPA EmptyPrompt

This is the strangest node in the pack and possibly the most sensible one. It has no inputs. It has one output, conditioning. And it exists so that you never download the thirteen-gigabyte text encoder.

The reasoning: LTX-2.5's alpha-gen IC-LoRA runs on an empty prompt. Not a short prompt, not a default prompt - empty. The matte is decided by the plate that guides it, not by language. That means the conditioning tensor feeding the sampler is a constant. A constant can be computed once, saved to disk, and shipped.

So it is. data/empty_prompt_ltx-2.5.safetensors, 24 KB, captured inside ComfyUI so it's bit-identical to what running gemma4 through its text encoder would produce. The node loads it and returns it in the shape ComfyUI wants, with pooled_output set to None and unprocessed_ltxav_embeds set true - the flags that tell the LTX stack these embeddings are already done, thanks.

Why this matters more than it sounds

The LTX-2.5 text encoder is gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors, 15.4 GB on disk, and the pack's workflow doesn't need it at all. On a machine squeezing the pipeline into 16 GB of RAM with models streaming off an NVMe, not loading a 13 GB encoder isn't a minor optimisation - it's the difference between the pipeline fitting and not.

It also removes a whole class of failure. No encoder means no encoder version mismatch, no "why is my prompt being ignored" confusion, no downloading a gated text encoder repo just to cut out a person.

What you do with it

Nothing. That's the point. On the main node, the text_encoder widget reads built-in empty prompt (no text encoder needed) and stays there. The pack's settings guide is unusually blunt about this one - the entry under Advanced says, essentially, "never".

If you're the curious type and you do install the gemma4 encoder, the node still exists and the option still works, and the pack's notes say the result is the same. You'd be spending 15 GB and a model load to get an identical matte, which is the kind of thing worth knowing you can do and never doing.

In the graph

It's the positive conditioning for every alpha-gen pass: the guide node, the CFG guider, the sampler. One node, fanned out to every pass, including the native-resolution tiles. Same in the v13 overlap passes and the v15 Clean Plate pass. A single constant conditioning is also why the sampler can run cfg=1.0 and get away with it - there's no prompt to guide toward, so there's no reason to pay for classifier-free guidance.

That is worth pausing on, because it's the deepest structural fact about this whole pipeline. Alpha-gen isn't generating "a person on a white background" from words. It's a video model taking a plate plus an IC-LoRA guide and producing a matte. The prompt slot exists because the architecture has one, and the pack fills it with the cheapest valid value.

Install

Nothing - it ships in the repo's data/ folder and requires no download.

cd ComfyUI/custom_nodes
git clone https://github.com/rik-python/ComfyUI-Comfy-Pro-Alpha-Generator.git
pip install -r ComfyUI-Comfy-Pro-Alpha-Generator/requirements.txt

If the file goes missing - a partial clone, someone tidy deleting "that small file" - you'll get an error loading it when the graph runs, and the fix is a fresh git pull in the pack folder. It's 24 KB; there's no clever recovery needed.

The only real hazard

Do not try to be clever with it. Because it's a conditioning output with the same type as any CLIP encode, you can wire it into other things, and it will behave exactly as badly as you'd expect: it's an LTX-2.x AV embedding with unprocessed_ltxav_embeds set, not a general CLIP conditioning. It's here to feed one model family's samplers, and the pack keeps it that way.

The other thing worth saying plainly: if you ever find yourself wondering why "alpha-gen ignores my prompt", it isn't ignoring it - there is no prompt. The subject is chosen by SAM 3 clicks, the edges by the plate. Words are for the Find box, not the sampler.