Instead of four graphs you get one, with a switch column on the left that decides which mode you are in. That is the whole appeal: image-to-video, text-to-video, video-to-video, first/last-frame and audio-driven digital human all live inside the same file, and the model runs at eight steps.

Evidence note: this episode ships no workflow JSON, so no node counts or sampler values are claimed. Every setting below is the author's spoken description of the mirror's graph.

When this is useful

  • One machine, one graph, several output types — useful when you switch modes often.
  • Fast sampling: eight steps on a 5090 for a card-draw workflow.
  • Dubbing or re-voicing an existing clip through the V2V path.
  • Audio-driven digital human by flipping the audio input mode.

How this differs from the previous build

It is a multi-in-one graph rather than a new model, and the differences are practical.

  • Mode is a button, not a file. The left column switches T2V, I2V, V2V and first/last-frame; the rest of the graph is shared.
  • Audio is a mode switch too. Set the audio input to the audio-drive option and an I2V run becomes an audio-driven digital human.
  • Three image slots in I2V map to first, middle and last frame — the same slots used by the first/last-frame mode.
  • Eight-step generation is the selling point; the author pairs it with a 5090 32GB for the fastest card draws.

Running it

  1. Deploy the mirror, open ComfyUI, and select the multi-in-one graph from the left workflow list.
  2. Pick the mode in the left column. The author advises using the buttons rather than tracing the graph, which is dense.
  3. For text-to-video, enter the prompt. Optionally enable the built-in prompt enhancement — the author prefers an external LLM because enhancement slows generation.
  4. Set resolution by preset, or switch the constraint to a pixel budget (the demo uses pixel budget) and set duration in seconds.
  5. For image-to-video, load the image into the first slot; the second and third slots are middle and last frame.
  6. To get an audio-driven digital human, change the audio load mode to audio input and supply the audio.
  7. For video-to-video, load the source clip on the right and set the same size inputs.
  8. Add LoRAs through "add LoRA model", uploading them to ComfyUI/models/LoRA first.

Honest limits

  • The author's own view is that V2V is rarely worth it — it is for dubbing or deliberately degraded lip sync.
  • The graph is large and organised into groups; editing it by tracing nodes is error-prone.
  • Two upscale options trade speed against VRAM headroom, and the author does not give measured benchmarks for either.
  • Audio and sound watermarks only take effect if the corresponding module is loaded on the right-hand side.
  • No workflow JSON ships, so nothing here is file-verified.

Episode resources

Mirror (Compshare, referral code preserved)