Video

The Video tab turns a prompt — and, if you like, an image or a clip — into a video with sound. Pick a mode, write the prompt, choose Quality and Length, press Generate (⌘⏎ from the prompt box).

Modes

The chips across the top of the form. Each one says what it takes.

Hailuo H3 renders Text and Image; the other modes are LTX’s.

Image: Anchor or Inspire

Drop a PNG, JPG or WEBP into Reference image — it is cover-cropped to the render size. On LTX-2.5 a Reference use row appears:

• Anchor — animate this image. Your picture is frame one and the clip moves it.
• Inspire — new shot from it. The clip takes the character or look into a new scene, and you will not see your picture as frame one. Good for “this character, new scene”.

On Hailuo H3 the image is always the first frame.

FFLF and Keyframes

• FFLF — Start frame (frame 0) opens the clip and its shape sets the output size; End frame (last frame) closes it. A close-up at the end holds a face through the clip.
• Keyframes — choose 3 keyframes to 8 keyframes. Start and End are fixed; the others are the beats in between.
• Both need the LTX High add-on (the Q8 model) and pick High quality on their own. If it is missing, Generate says so and tells you where to install it.

Extend

Pick the Source video, set Extend by (seconds) (0.5 to 10, default 2) and Direction (After or Before). Extend has its own two qualities:

• Q8 Draft — 12 steps, safe on a 64 GB Mac.
• Q8 Pro — 30 steps, wants 96 GB or more.

Stay on Q8 Draft unless the Mac has the memory.

Writing the prompt

Describe the scene and the sound. Sound is generated with the picture; a prompt with no sound in it comes out near silent.

• Enhance (LTX) rewrites your prompt in the style the model was trained on. With LoRAs on, it reads their notes and keeps their trigger words.
• No music asks for voice, sound and ambience without a score.
• No voice appears when the selected character has a voice: the face still locks, the audio stays ambient.

The Avoid box

Avoid + opens a second box for what the model should not make — blurry hands, distorted fingers, warped text.

When it works Avoid applies on High, High · 720p, Extend, FFLF / Keyframes and Audio. Quick, Balanced and Standard run without guidance and ignore it. On Hailuo H3 there is no Avoid box: write refusals as plain sentences in the prompt, and only for what H3 adds unasked — camera drift and on-screen text.

Quality and Length on LTX

Each Quality chip shows its size and an estimate for this Mac. A chip reads unavailable when this Mac cannot run it, or names the download it needs.

Quick, Balanced and Standard run on the fast (distilled) model; High and High · 720p need the High add-on (the Q8 model).

Length: 3s, 5s (default), 7s, 10s — and 20s, which is offered at Quick only.

On the fast qualities, a Speed row offers Tuned or Fast draft — a faster schedule that changes the take. High has a STG — detail guidance slider.

With a trained character selected the chips become Q8 Draft (704×384), Q8 Pro (1024×576, best identity) and, on LTX-2.5, High and High · 720p. Trained faces hold best on Q8 Pro.

Sliding windows on LTX

For a clip longer than one pass, Advanced → Long clips → Windows renders it as a sequence: one first pass, then one Extend per later window, each continuing the kept tail of the one before. This needs the High add-on (the Q8 model) and a Mac that supports Extend.

Per-window prompts. Window 1 is always the main prompt above — there is no separate box for it. Every later window gets its own labelled, timed box (“Window 2 · 0:05–0:09”), built from the real render schedule; a box left blank holds the previous moment. Write the next beat of the action in each one, led by the movement (LTX renders whatever a prompt names first). Invariants — what must not change between windows (who is in frame, the light, the lens, the place) — go in the field below the boxes and are re-injected into every later window automatically.

This is the LTX equivalent of H3’s per-window prompts — same idea (one prompt per pass so the action doesn’t repeat), different engine and a different box layout, since the number of windows and their real timing come from LTX’s own window planner rather than a fixed 5-second grid.

Quality and Length on Hailuo H3

Length: 3s, 5s (the longest single pass), 10s and 15s. 10 s and 15 s are chained from 5-second windows at render time, so choose the length before you render — there is no Extend afterwards.

Turbo and Steps

• Speed: Standard / Turbo. Turbo is a 4-step adapter: much faster, and it overrules Steps. If it is not downloaded yet, the chip shows its size.
• Steps: Auto / 12 / 16 / 20. More steps, a longer wait (20 is the official reference recipe, about 2.4× Auto).
• On H3 installs whose runner has a single adapter slot, an Adapter row asks whether that slot goes to Turbo or My LoRA — they cannot both run. See LoRAs on H3.

Per-window prompts

A 10 s or 15 s H3 clip asks every 5-second window for the same prompt, which can repeat the action. Open Per-window prompts and turn on One line per window: 10 s gives two boxes, 15 s three. A box left empty uses the main prompt.

H3 first frame and orientation

Image mode works on H3 — the picture is the first frame. The Orientation row (top-level, next to the size chips) turns the tier’s canvas Portrait at the same cost.

Engines side by side

H3 and memory. The full H3 engine needs 60 GB. From 36 GB up, H3 runs on its compact Q8 engine, which Install Hailuo H3 in Pinokio’s Phosphene sidebar builds locally (about 5 minutes, ~22 GB, no extra download). Settings → Hailuo H3 model chooses between Automatic, Full and Compact. Below 36 GB, H3 is not offered — render on LTX. See Troubleshooting.

Upscale and export

After the render, the clip can be resized without cropping.

On Hailuo H3 (under After the render → Upscale):

• Native — as rendered.
• 720p fit / 1080p fit — scaled to fill the frame. A canvas close to 16:9 (H3 Draft and Standard, LTX Standard) loses a few pixels at the edges instead of getting black bars; one far from it (square, portrait into landscape) is still padded. 720p fit is the default. Below the sizes, Also run Upscale & Face Fix after the draft is optional and off by default. When it is ticked, the draft ships as rendered and a second job re-renders it at twice the size with LTX-2.5, keeping the face and the sound; the fixed clip lands next to the draft. It needs the 0.3 GB Upscale adapter, downloaded from the Models window, and takes about the draft’s time again.

On LTX (under After the render → Export): 720p fit (default), Native, or 2×. When it is not Native, Method chooses Fast (instant) or Sharp (a sharper upscaler, +30–90 s).

Any finished clip can also go through Upscale & Face Fix later — one click over the player, in its action overlay.

More controls

• Seed — -1 is random; reuse a seed to get the same take with a changed prompt.
• LoRAs — the add-ons picker. See LoRAs.
• Orientation (LTX) — 16:9, 9:16, 1:1 or 4:5.
• Advanced — width × height, duration and frames.
• After the render — Export/Upscale, and Open file when done.
• ⊞ Batch — paste many prompts and queue them all.
• Hide (the eye icon on the player toolbar) — removes a clip from the gallery, with an Undo in the toast. The Hidden (N) chip in the Outputs filter row shows everything hidden, so a clip hidden by mistake is one click away.