Choose a standalone H3 workflow based on the input you already have and the result you need. Use the master as a reference map: it loads all 16 groups enabled, so selecting a group does not isolate a run.

Watch on YouTube · Video published 2026-09-22 · Companion reviewed 2026-09-23

Start with your input

The September 2026 walkthrough covers sixteen routes, including preparation utilities and experimental tasks. They are not sixteen guaranteed production outcomes. Choose one route, use a short test, and inspect identity, motion, audio and the end of the clip before scaling up.

WorkflowUse it when
01 · Text to videoYou have a written scene and no visual starting frame.
02 · Image to videoYou want movement from an existing still.
03 · First + last frameYou have a plausible beginning and ending image.
04 · Reference to videoYou need separate identity, motion and audio references.
05 · Image + supplied voiceYou have a starting still and a speech recording.
06 · Reference clip preparationYou need a trimmed clip and endpoint images.
07 · Multi-character sceneYou want distinct people in a shared scene.
08 · Character + product + locationYou want references assigned to different scene elements.
09 · Storyboard to videoYou have guide images for points within a shot.
10 · Motion / camera referenceYou want a reference to guide how the shot moves.
11 · Two-character dialogueYou need to test turn-taking and speaker attribution.
12 · Continue a clipYou want a short extension from existing footage.
13 · Performance transferYou need a pose/performance-guided experiment.
14 · Object removalYou want to investigate removal; the shown test failed.
15 · End on this imageYou have an intended ending composition.
16 · Multilingual performanceYou need speech/performance tests in different languages.

Set up one route safely

  1. Install the model components and nodes listed for that standalone. Begin with the official native H3 guide and the package’s dependency manifest. FL2VA and REF2VA are different routes; do not treat their accelerators as interchangeable.
  2. Open one standalone JSON. The current 302-node master intentionally loads every group enabled. Merely clicking or selecting one group does not prevent the other enabled outputs from running.
  3. Load your own images, clips or audio into the indicated input nodes. The downloadable workflow pack excludes source performances and model weights. Historical filenames in a graph identify slots to replace, not files that are bundled.
  4. Keep the graph’s compatible frame count, dimensions and schedule for the first test. Record your seed, model revision, accelerator and input filenames. Then change only one variable.
  5. Review the entire exported clip, including its final frames. Check audio duration, unwanted cuts, facial changes, hand movement and unexpected titles or endcards.

Prompt for a shot, not a list of adjectives

Suggested image-to-video practice prompt

The person in the starting image gives one small wave, lowers their hand, and settles into a relaxed pose. The camera makes a slow, steady move forward. Keep the same room, clothing and soft side lighting throughout the shot. Finish on a calm expression without a scene cut.

For supplied speech, include the exact intended words and allow enough time for them. Preserving the original audio in the export is different from proving that every mouth movement matches it. Do not use someone else’s likeness, voice or performance without the rights needed for the project.

What the demonstrations do and do not establish

The object-removal example did not reliably remove the object and remains experimental. Image-plus-voice can generate unwanted endings. Reference-guided performance can drift, and a guided final frame does not guarantee a plausible path to that frame. These are inspection points, not problems solved by adding more steps. The prepared graphs and episode evidence are distinct from a fresh compatibility test on your machine.

Sources and downloads

Related lessons