Text Stage

An LLM inside a ComfyUI graph is a mainstream pattern now - prompt enhancement, captioning, structured prompt generation - and Text Stage is ComfyTV's entry point for it. It's a Generate stage that runs a local text workflow and hands the result out as text, so you can chain "generate a description" straight into a video or image stage without copy-pasting between apps.

It's a ComfyTV Generate stage: type what you want, pick a workflow, hit Run, and get COMFYTV_TEXT downstream.

The inputs

• workflow - which backend text workflow to invoke, defaulting to Local Qwen3 4B. This is the honest part: the tooltip says "Placeholder for now." Like the pack's other workflow-driven stages (Upscale, SubtitleGen), this runs a local ComfyUI workflow, and the shipped one needs a real model: qwen_3_4b.safetensors in ComfyUI/models/text_encoders/ (the pack's models doc links the HuggingFace source). No model, no output.
• main_prompt (multiline) - your intent for this stage. The tooltip notes that upstream text inputs are treated as additional context, and the node combines them (newline-separated) with your main prompt before calling the workflow.
• texts, images, videos (autogrow inputs) - additional context slots. Wire an upstream text stage here and its output joins the prompt; images/videos get passed along to the workflow as inputs it can use.
• custom_params - internal JSON of user-defined parameter attachments; leave it alone.

The project_id / parent_output_id / force_run_token inputs are internal frontend bookkeeping - ignore them.

How it works

Text Stage packages your combined prompt (main + upstream texts) and any attached media, then invokes the selected text workflow through the pack's local-workflow runner. Because it's workflow-driven, the LLM actually running is whatever your chosen workflow loads - the stage is the plumbing around it. The "additional context" behavior is what makes it feel like a node rather than a chat box: upstream stages feed it, and its output feeds the next generation stage.

Install

ComfyTV ships ~190 stages in one pack; this node comes with it.

cd ComfyUI/custom_nodes
git clone https://github.com/jtydhr88/ComfyTV

Restart ComfyUI fully and look under ComfyTV → Generate. ComfyUI Manager finds it by searching "ComfyTV". The pack itself declares zero Python dependencies, but the Qwen3 4B workflow needs its ~2GB+ model downloaded into models/text_encoders/ (see docs/models.md for the exact filename and link). On ComfyUI Desktop or macOS, clone into the running instance's absolute path from the startup log rather than trusting the relative cd.

Troubleshooting

• No output / workflow errors. Almost always the model: the default workflow needs qwen_3_4b.safetensors in text_encoders/. Get that file and the stage has something to run.
• "Placeholder for now" vibes. The tooltip is candid - this stage is early in the pack's roadmap. It works when you have a text workflow to point it at, but don't expect a polished product.
• Output repeats the input. A bare chat model echoes your prompt; that's the workflow's prompting, not the stage. Swap or configure the workflow's system prompt.

If you already run a local LLM for prompt work, this is the node that lets ComfyTV use it in the graph. Just remember the model download is on you.