Video Stage

This is the node that turns a prompt into an actual moving clip inside ComfyTV, and it's the reason most people install the pack. Drop one on the canvas, pick a workflow, type a prompt, hit ▶ Run, and a few seconds later you're scrubbing an MP4 preview right in the node body. It's a per-node run, not a trip through ComfyUI's global queue - everything happens inside the stage, and whatever it produces gets stored as a project snapshot.

What it actually does

Video Stage is ComfyTV's text/image-to-video generator. It wraps a curated ComfyUI video workflow - the pack ships four Local LTX 2.3 variants: T2V (text only), I2V (one image as the first frame / appearance ref), FLF2V (two keyframes, start and end), and IA2V (an image plus an audio track, for lipsync or music-driven motion). Same 22B LTX 2.3 base with a Gemma 3 text encoder, a Lightning LoRA, and a 2× spatial upscaler under all four, just wired differently.

Why LTX 2.3? It's the speed tier of local video: seconds per clip on consumer VRAM rather than minutes, which makes it the natural pick for a canvas app where you iterate. Draft fast here, and refine elsewhere if you want more polish.

The inputs that matter

• workflow - pick the mode. T2V wants only a prompt; I2V wants one image wired into images; FLF2V wants exactly two in start→end order; IA2V wants an image and an audio track and usually has generate_audio left off.
• main_prompt - your scene, camera move, mood. Upstream text stages get appended as context.
• resolution / aspect_ratio - short-side tier (720P default) plus ratio; set 9:16 for vertical shorts.
• duration_s - clip length in seconds (1–120, default 5). The subgraph derives fps and frame count from this.
• generate_audio - whether the workflow also synthesizes an audio track.

The output is a single video of type COMFYTV_VIDEO - a snapshot URL stored in the project, not a native ComfyUI VIDEO tensor. That's the whole architecture in a nutshell: ComfyTV stages talk to each other in snapshots, and only Bridge nodes (→ / ← ComfyTV Video) let you cross into a normal tensor-based graph. Want to hand the clip to a plain SaveVideo? Bridge it.

Installing and the model tax

Install is the same for every ComfyTV node: cd ComfyUI/custom_nodes && git clone https://github.com/jtydhr88/ComfyTV, restart ComfyUI, and it appears under the ComfyTV category. ComfyUI Manager can install it too if you prefer.

The real cost here isn't the install - it's the models. Unlike every other node in this pack, Video Stage needs actual weights. The README's models.md lists the LTX 2.3 files (22B fp8 checkpoint, Gemma 3 text encoder, the Lightning and LoRA files, plus the upscaler) spread across checkpoints/, text_encoders/, loras/, and latent_upscale_models/. Altogether that's a big download, so budget for it before you hit Run and stare at an error. MiniMax H3 workflows are also available if you want a heavier, higher-fidelity alternative.

Common gotchas

• I2V ignores your reference? You're on T2V - pick the I2V workflow and make sure the image is actually wired into images.
• FLF2V with one image? It needs two keyframes, in order.
• IA2V fails or audio drifts? audio is required, and its length should roughly match duration_s.
• Can't connect to normal ComfyUI video nodes? That's expected - you're holding a COMFYTV_VIDEO snapshot. Insert a ← ComfyTV Video Bridge to convert it.

One more thing worth knowing: the "Local LTX 2.3 T2V" default label is just the preset - you can import any ComfyUI workflow as JSON and bind it to this stage in the sidebar editor, so it'll happily drive your own model stack instead.