Video¶
A prompt, or a prompt and a photo, into a clip. Or your photo moving like a video you choose.
There are four video models. Install any of them:
| model | makes | needs |
|---|---|---|
| Wan 2.2 Turbo (6.25 GB) | up to 5 seconds, any shape up to 1280 × 704. The better clips; minutes per clip | Snapdragon 8 Elite or newer (alpha on 8 Gen 3: tell us if it runs) |
| Wan 2.2 14B (about 21 GB; 3.7 GB less beside Wan 2.2 Turbo) | your picture moving, or text to video through a picture first (Text to video (14B) below). The best motion and detail, and slow: about 9 minutes for 2 seconds at 480 × 832, 12 minutes for 1 second at 720 × 1280 (faces come out much sharper at the larger size) | Snapdragon 8 Gen 3 or newer, 12 GB of RAM (on an 8 Gen 3, up to 832 × 480) |
| MiniMax H3 (17.5 GB; 2.4 GB less beside Krea 2) | a clip with sound: speech, sound effects and music made with the picture. Up to 5.2 seconds at 384 × 384, 3 seconds at 832 × 480. About 6 minutes for 3 seconds at 384 × 384 (Text to video with sound below) | Snapdragon 8 Gen 3 or newer, 12 GB of RAM. Not licensed for the EU, the UK, South Korea or the US |
| Neodragon (about 8.6 GB) | 2 seconds at 1024 × 640 in about 25 seconds | Snapdragon 8 Elite or 8 Elite Gen 5 |
Video has its own models; only Text to video (14B) also uses your picture model, for its first frame. Models → Video lists them all, and Wan in the family menu has Wan's models on their own page.
Text to video¶
Nodes: Prompt → Video → Output.
- Flows → Text to video.
- Write the prompt: describe the scene and the motion.
waves rolling onto a beach at sunset, camera slowly panning right. - Run. With Wan 2.2, 5 seconds at 832 × 480 takes under 3 minutes (about 2:45) on an 8 Elite; the bar moves once per step, four steps, then the clip is decoded and encoded.
The video node's Model dropdown switches between Wan 2.2 Turbo (the default) and Neodragon (the fastest), and Wan 2.2 14B with a picture wired: the 14B animates a picture. For text to video on the 14B, use the next flow. Use on Neodragon in Models opens the flow on Neodragon.
Text to video (14B)¶
Nodes: Prompt → Image (your picture model) → Video (Wan 2.2 14B) → Output. Needs Wan 2.2 14B and an 8 Gen 3 or newer with 12 GB (an 8 Gen 3 goes up to 832 × 480).
- Flows → Text to video (14B), or Use on Wan 2.2 14B in Models (a picture model's Use does not offer it). The picture is made by the model in the top bar, in its own shape: the clip takes that shape (a 512 × 768 picture makes a portrait clip), and the video node's size sets how many pixels, 832 × 480 worth for a quick try or 1280 × 720 worth for sharp faces. 1 second by default.
- Write one prompt for the scene and its motion: it makes the picture and drives the clip.
- Run. The picture appears on its node first, in seconds, then the clip takes minutes. Not the picture you wanted? Stop the run and Run again. Happy with it? Lock its seed (the seed chip lists each node's seed): the next Run keeps that picture and only makes a new clip.
The clip loops on the node that made it; tap it to play it full screen. With Save on
(the default), an MP4 is also written to Movies/Nightmare.
Image to video¶
Nodes: Prompt and Image → Video → Output, the same flow with a photo wired in.
- Flows → Image to video.
- Choose the photo on the image node. Tap the video node's Crop preview: the frame takes the clip's shape; drag and pinch to choose the part that is animated. The photo is the clip's first frame.
- Describe what moves:
the camera slowly pushes in, the woman turns her head and smiles. Run.
Text to video with sound (MiniMax H3)¶
Nodes: Prompt → Video (MiniMax H3) → Output. Needs MiniMax H3 and an 8 Gen 3 or newer with 12 GB of RAM.
- Flows → Text to video (or Image to video, to start the clip from your photo), then set the video node's Model to MiniMax H3 (with sound), or Use on MiniMax H3 in Models. With a photo wired the clip starts on it: describe what the person in it does and says. A reference instead of a first frame: Flows → Reference to video (or Use on MiniMax H3 Reference in Models) opens Prompt, Image and Video (MiniMax H3) with the picture wired to the node's reference input: pick your photo on the Image page and the clip keeps its subject in a new scene you describe, as Picture 1 in the prompt (it opens on an example: the person on a beach at sunset, waving and saying one line, 384x672 for 1.6 s). In the advanced view you can wire any picture to reference yourself. It needs MiniMax H3 Reference (Models → Video, 14.4 GB on top of MiniMax H3) and makes one part per clip; a first frame and a reference cannot be wired together. It is 1.5 to 2.5 times slower than plain MiniMax H3 (a 2 GB LoRA rides on every run: a 1.6 s clip at 384x384 takes about 8 minutes on an 8 Elite against 3), and the node says so while a reference is wired; unwire it for a faster clip.
-
Prompt: it opens on a worked example (a barista speaking one line in a café) in the sections H3 was trained on. Edit it section by section; the hint under the box says what each one is for:
section what to write subject_definitions who is on camera and how they sound: S1 is a man in his late twenties with short brown hair. He has a clear, warm voice. Or N/A for nobody summary the clip in one or two sentences detailed_description the look and camera, then the shot: where it is, what happens, and the words said out loud as <d>[English] Hello! This was made on a phone.</d>(any language; about 3 words a second)overall_soundscape the sounds: rain on the window and distant thunder non_diegetic_music background music: soft piano, or N/A for none Write N/A in a section to leave it out. The length sentence ("The target video lasts ... seconds.") is set to the Seconds you pick. A plain one-line prompt works too: it becomes the scene, with no speech and no music. 3. Run. 3 seconds at 384 × 384 takes about 6 minutes on an 8 Elite; 832 × 480 about 13. The clip plays with its sound in the Output and in Results, and saves with it.
MiniMax H3 was trained on prompts in this shape, so the sections matter: one line leaves the sound to chance (music you did not ask for, a robotic voice). A prompt from ComfyUI's H3 prompt rewriter works as it is.
| setting | what it does |
|---|---|
| Width, Height | 384 to 1344 in 32 px steps, up to 768 × 1344 worth of pixels. 384 × 384 is the quickest. Nothing under 384: smaller clips lose their colours |
| Seconds | H3's own lengths: 0.9, 1.6, 2.3, 3.0, 3.8, 4.5 or 5.2, at 24 frames a second. A larger size fits fewer: on an 8 Elite about 5.2 s at 384 × 384, 3.0 s at 832 × 480 or 640 × 640, 2.3 s at 960 × 544, 1.6 s at 1152 × 640, 0.9 s at 768 × 1344; on an 8 Gen 3, 5.2 s at 384 × 384, 2.3 s at 832 × 480, 1.6 s at 640 × 640 or 544 × 960, 0.9 s at 768 × 1344. A new node opens on 832 × 480 for 2.3 s, which fits both. The node says when a size and length are too much |
| Smooth motion | twice the frames, as on the Wan nodes; the sound stays in sync |
| Parts | up to 10 parts, each starting on the last frame of the one before, as on the Wan nodes; the sound of each part joins with it. A plain line in Part prompts replaces only the scene, so the speaker's voice, the sounds and the music carry through |
| Seed | 0 picks a new one each run |
Long renders heat the phone: see Heat. MiniMax H3 takes LoRAs made for it (CivitAI base model "MiniMax H3"), from the node's LoRA list or a <lora:name:1> tag in the prompt; one stored as 32-bit floats is rewritten once to 16-bit, which made a 298 MB LoRA 4x faster. Each LoRA adds time to every step. Powered by MiniMax H3. Its licence (the MiniMax H3 Community License) does not cover the European Union, the United Kingdom, South Korea or the United States; anything you make with it and share publicly must say it was made by AI; and its clips may not be used to improve other AI models. The app shows these terms before the download and asks you to accept them.
Wan 2.2 settings¶
| setting | does |
|---|---|
| Width, Height | Any shape, in 32 px steps, up to 1280 × 704 worth of pixels: square, 4:3, 21:9, a phone screen. Past that, the other side shrinks to fit. Wan 2.2 is made for 1280 × 704: smaller sizes, 832 × 480 included, can look soft or distorted. 832 × 480 is quickest; 1280 × 704 takes about three times as long. On Wan 2.2 14B (its own 16 px steps, up to 1280 × 720 worth of pixels), 1280 × 720 is the size for faces and fine detail (about 12 minutes for 1 second); 832 × 480 is for trying an idea first (about 5 minutes), with soft faces. On an 8 Gen 3 the 14B goes up to 832 × 480 worth of pixels (720p there is untested). Match picture sets the photo's shape |
| Seed | 0 = a new clip every Run; type a seed shown on the node to get that clip back |
| Seconds | 0 to 5, at 24 frames a second (Wan 2.2 14B: 1 to 3 at 16 frames a second; 1 second at 720p is about 12 minutes, 3 can take an hour); 0 makes one still picture (about 15 s at 832 × 480 on an 8 Elite; wire last frame to the output). On a 12 GB phone, 1280 × 704 for 3 seconds or more renders in parts so it fits: slower, about 12 minutes for 5 seconds. On an 8 Gen 3 a new flow opens on 2 seconds, the longest that has worked there so far; longer clips and 1280 × 704 still run, with a warning in the run log |
| Smooth motion | Twice the frames: a new one between every pair, made on the NPU by IFRNet, so 16 frames a second become 32 and 24 become 48. Fast and smooth motion stops stuttering. It adds seconds, not minutes: about 0.2 s a new frame at 832 × 480 and 0.5 s at 720p on an 8 Elite. Up to 1280 × 736 or 736 × 1280; a larger clip stays at its own rate. Needs Smooth motion from Models → Tools (about 75 MB, Snapdragon 8 Gen 3 or newer); switching it on offers the download |
| Parts | 1 to 10. More than 1 makes a longer clip in parts, each as long as Seconds and starting on the last frame of the part before, joined into one video. Each part takes as long as one clip. While it runs, the node plays the clip so far after every part (silent until the end) |
| Part prompts | With 2 or more parts: what happens in part 2 onwards, one line each. A blank or missing line repeats the part before; part 1 uses the Prompt node |
| Keep colours | With 2 or more parts, on by default: holds every part to the first frame's colours. Without it a long clip tends to turn darker and redder part by part (the model drifts a little each part and the next starts from there). Turn it off for a clip meant to change its light |
| LoRAs | A LoRA made for the model you use: Wan 2.2 Turbo (TI2V-5B) LoRAs on Turbo; Wan 14B LoRAs on Wan 2.2 14B. A single-file Wan 2.1 14B LoRA works on both of its halves; a Wan 2.2 high/low pair is told apart by its file names (high_noise, low_noise). See LoRAs |
Neodragon has Seed and Upscale (on: 2× to 1024 × 640; off: 512 × 320, a little faster).

Longer clips: one film¶
The simplest way is Parts on the video node (above): up to 10 parts, a prompt for each, one video.
The Wan video node also has a second output, last frame: the clip's final frame as a picture. In advanced view, wire it into another Wan video node's image and give that one its own prompt: it carries on from where the first clip ended.
The second node's video is then the whole film so far, the first clip and its own joined into one (the frame they share appears once). Chain as many as you like: the last node hands the Output the complete film. A clip that carries on a film takes the film's frame rate: smoothed if the film is, whatever its own Smooth motion switch says. Results keeps every part on its own as well, each with the picture it started from, grouped together beside the film.
On the canvas, a film longer than about 5 seconds loops faster than real time; full screen plays it at its own speed. The camera button beside share (full screen, on the canvas and in Results) saves the frame on screen to the gallery as a picture.
Motion transfer¶
Your picture moving like a video you choose: its pose, frame by frame.
Needs Wan 2.2 Fun Control: about 3.7 GB with Wan 2.2 Turbo installed (it borrows Turbo's text encoder and decoder), 7.3 GB without. Models → Video → Wan 2.2 Fun Control → Download.
Nodes: Prompt, Image (who moves) and Video (the moving clip) → Motion transfer → Output.
- Flows → Motion transfer.
- Choose your picture on the image node: it is who the clip shows (face, clothes, setting), and the clip starts on it. A full-body photo works best.
- Choose a video on the Video node: a person dancing, walking, waving. Its sound can come along.
- On the Motion transfer node, tap the Video preview and drag either end of Part of the video: the part you choose is the clip's length, up to 30 seconds. A long part is made in pieces, each starting a few frames before the one before ends and joined where the two match best.
- Pick a Mode (the ⓘ beside it says which to use) and Run.
| Mode | for |
|---|---|
| Pose | a real video of a person. The body, hands and face are read from every frame and your picture copies them. Needs the Pose Detector and Whole-body pose (Models, Tools) |
| As is | a video that already is a control video: a pose skeleton or a depth map made elsewhere. An ordinary video here will not work well |
| Setting | what it does |
|---|---|
| Frame rate | how often the video is read: 24 (Wan's own rate, best), 12 or 8 (fewer frames, so faster, but the motion moves in bigger jumps; turn on Smooth motion to fill them in) |
| Follow the person | on: the crop moves with the person on every frame, so a dancer who crosses the room stays large. Off: draw a box in the Video tab that cuts every frame the same way |
| Line up my picture | on: your picture is cropped so its person stands where the dancer starts, at the dancer's size, so the clip does not open on two people |
| Keep audio | on: the video's sound over the same part is copied into the clip |
| Smooth motion | twice the frames, as on every Wan node |
Motion transfer takes no LoRAs: <lora:…> tags in its prompt are skipped, and the run log says
so. It opens at 480 × 832, where the crop that follows the person keeps faces and hands readable.
On an 8 Elite 10 seconds at 24 fps take about 10 minutes, in three pieces; at 12 fps with Smooth motion about 5 minutes,
in one piece (up to 10 seconds). If the app is closed mid-way (other apps need the memory), Run again with the same
seed and it carries on from the last finished piece.