Skip to main content
The ComfyTV / Generate group has the stages that create new content from a prompt: Text, Image, Video, Music, Speech, and 3D Model. Every generator works the same way: type a prompt, pick a workflow (model) from the dropdown, click ▶ Run. The result shows in the node’s preview and flows to anything wired downstream. Under every workflow dropdown sit two buttons:
  • 🔗 Link workflow — the recommended way to use your own workflow. Pick one straight from ComfyUI’s own workflow library; edit it in ComfyUI and save, and ComfyTV picks up changes automatically. See custom-workflows.md.
  • ⬆ Upload workflow — import a .json file from disk, no restart.
Hover a workflow entry in the stage manager to ★ star it as the default for that stage kind, and use the preset bar on the stage card to save/recall parameter presets.

​
Image Stage

Image stage
  • Prompt: the main text box. Upstream Text nodes get appended; upstream images feed the current workflow’s LoadImage (only the i2i workflow consumes them). @image_N tokens reference the wired image slots inline.
  • workflow — shipped: Local SD1.5 (t2i), Local SD1.5 I2I (image-to-image), Image Ideogram4 T2I — plus anything you link or upload.
  • resolution / aspect_ratio / batch_size: target size tier, shape, and how many images per Run.
Two outputs: images is the whole batch from this Run; image is the single thumbnail you picked on the node (defaults to the first one until you click).

​
Image-to-image

Pick Local SD1.5 I2I, wire a reference image into the images slot, write the prompt, Run.

​
Video Stage

Video stage
  • workflow — two families ship today:
    • LTX 2.3 (video only):
      • Local LTX 2.3 T2V — text → video.
      • Local LTX 2.3 I2V — image → video.
      • Local LTX 2.3 FLF2V — first-last-frame → video. Wire two images on images (start + end keyframes).
      • Local LTX 2.3 IA2V — image + audio → video.
    • MiniMax H3 (generates video with sound in one pass — ambience, foley, even dialogue):
      • Local MiniMax H3 T2V — text → video+audio.
      • Local MiniMax H3 FLF2V — first-last-frame → video+audio.
      • Local MiniMax H3 R2V — multi-reference: wire several images (character sheets, scene, props, style) and optional audio references, then address them as @image_N / <Audio N> in the prompt to lock identity and continuity across shots. The H3 workflows ship with Lightning LoRAs wired in, roughly halving generation time (see models.md).
  • resolution / aspect_ratio / duration: output size, shape, length.
  • audio input — required for IA2V; used as reference audio by H3 R2V; the other workflows don’t use audio.
Once you have a clip, the whole video suite — editing, color, keying, compositing, FX — takes over from there.

​
Text Stage

Text stage Local LLM text generation (built-in Qwen3 4B). Use it to expand a prompt, write a description, or feed other stages’ context slots.

​
Music Stage

Music stage Text-to-music. Two workflows ship:
  • ACE-Step v1 Song — the lightweight default.
  • MiniMax Music 3 — higher-quality full songs.
Shared parameters:
  • Prompt: free-form tags — genre, mood, BPM, instrumentation.
  • Lyrics (optional): empty = instrumental; non-empty = vocal track.
  • Duration: slider (1–240 s, default 30).
Output is a single audio file. From there the audio suite (EQ, stem split, reverb, …) and the Music nodes can pick it up.

​
Speech Stage

Speech stage Text-to-speech. Ships with a Kokoro TTS workflow: type the line to speak, pick a voice in the workflow’s parameters, Run. Output is an audio clip — wire it into a Video Stage’s audio input (IA2V), a Mux Audio node, or the audio suite.

​
3D Model Stage

3D model stage Generates a 3D model (GLB) via a model-kind workflow. Three ship today:
  • Hunyuan3D 2.1 — image → textured mesh.
  • MoGe-2 Depth Mesh — image → geometry via monocular depth estimation.
  • TripoSplat Gaussian — image → Gaussian splat (loads into Scene3D too).
Model files for all three are listed in models.md; you can also link or upload your own model-kind workflow. The node embeds an orbit preview; after you orbit, it auto-captures a screenshot into its image output, so downstream image stages can consume the model’s look. The GLB itself flows to the 3D nodes — Scene3D, mesh ops, materials; see the Node Reference.