Subjects, continuations and fashion clips

Generative video and sound explains how a generation is priced, run, measured and finished. This page takes three jobs that need more than one clip from one prompt:

the same subject in several shots, where a single image is not enough and each new shot risks a different face, a different product;
a clip that goes on, where the new seconds must start exactly where the old ones end;
a product worn and in motion, from a photo of the product alone.

Each one rests on the same rule: a model sees only the images it actually receives. A node can have ports for a last frame or six reference images, but many models take a single image and ignore the rest. The sections below say, node by node, which models use what — and the step fails, or warns, where it cannot.

Which images reach the model

Node	Images it can send	What the models do with them
`ai/image_to_video`	a first frame, a last frame, up to 6 reference images	the last frame is used by Seedance Pro (the default), Seedance 2.0 and 2.5, Kling 2.5 Turbo, 2.6, O3 and V3, MiniMax Hailuo-02, H3, H3 Max and H3 Max Turbo, Vidu Q3, Wan 2.2, LTX-2 19B, and FLUX 3 First-Last Frame and PixVerse C1 Transition (which require it); other models ignore it. Reference images are used only by Veo 3.1; the others animate the first frame alone
`ai/fashion_video`	the same as Image to Video	the same models, the same rules
`ai/reference_to_video`	up to 9 images, 3 videos, 3 audios, cited in the prompt	all of them, on the default model — this is the node for a consistent subject
`ai/virtual_tryon`	the person, the garment, and optionally a background, a pose reference and a blank product	all of them on the default model; a model that takes one image is refused before any charge (`MODEL_TAKES_ONE_IMAGE`)
`ai/video_extend`	the clip to continue	the whole clip

So: to keep a subject the same across shots, give it to Reference to Video, not to Image to Video’s reference port. To end a clip on a chosen frame, keep the default model or pick one from the list above.

The same subject in several shots

ai/reference_to_video makes a clip from a prompt and a set of references. Each reference is cited in the prompt by its kind and its place: @Image1 is the first connected image, @Image2 the second, @Video1 the first video, @Audio1 the first audio. The numbering follows the connected slots in order and skips empty ones — connect the mascot to the first slot and the product to the second, and write “@Image1 pours tea from @Image2”.

What keeps a subject the same:

The same reference images in every shot. Two shots with the same @Image1 share the face, the colours, the clothes; two shots from two prompts alone do not.
References made for it. A character on a plain background, whole and in front view; a product isolated, in three-quarter view. What is not in the reference, the model invents.
Say what must not change. “Keep the fox’s round glasses, mustard apron and claymation look, and the teapot’s speckled teal glaze and bamboo handle, exactly as in the references” — in every shot’s prompt.
Change the scene, not the subject. Setting, light, framing and action are the prompt’s job.

Limits worth knowing:

up to 9 images, 3 videos (together 2 to 15 seconds, 480p to 720p, at most 50 MB) and 3 audios (together at most 15 seconds, MP3 or WAV); audios need at least one image or video (INVALID_REFERENCE_COMBINATION);
no real people. The default model refuses references showing the likeness of a real person, with CONTENT_POLICY_VIOLATION, before any charge. Use a character, a mascot, an illustration, a product — or a synthetic person;
the duration is the model’s: 5 seconds on the default, 4 to 30 if you set it; a seed is sent and honoured on a best-effort basis.

The default model, Seedance 2.5 reference, costs 29 credits per second at 720p — 145 for five seconds — and about half that at 480p. Seedance 2.0 (25 per second, 0.6× with a video reference), Seedance 2.0 Fast (16) and Gemini Omni Flash 1.1 reference (about 6.3 per second at 720p, 2 at 360p, up to 19 at 4K; up to 10 images and 3 videos of at most 3 seconds each, no audio) MiniMax H3 Max reference (5 per second at 768p, images only; more than four images add up to 24 credits) Grok Imagine Video 1.5 reference (5 per second at 480p, 8.75 at 720p, 1 to 7 images cited as <IMAGE_0>, <IMAGE_1>, about 0.6 credit each) and PixVerse C1 reference (about 1.9 per second at 360p, 3.1 at 720p, 5.9 at 1080p, a quarter to a third more with audio; 1 to 7 images cited as @Image1, @Image2, the last one optionally the background with Background Reference) are the alternatives. A step with more images than its model takes stops before any charge (TOO_MANY_REFERENCE_IMAGES).

Lab: one mascot, one product, two shots

The lab’s workflow, left to right: the mascot and the product, both connected to Shot 1 and Shot 2 (Reference to Video) as reference_images_0 and reference_images_1; the two shots go into Merge Videos, which gives the story.

A small ceramics brand has a mascot, Pip the fox potter, and a best-selling teapot. It wants two short scenes for a campaign: Pip pouring tea in the studio, and Pip showing the teapot at a market at dusk. The two references are connected to both shots, and the prompts change only the scene.

Left to right: the mascot reference, a claymation red fox with round glasses and a mustard apron marked PIP; the teapot reference, speckled teal glaze and a bamboo handle; a frame of shot 1, Pip pouring tea from the same teapot at a wooden workbench in a pottery studio; a frame of shot 2, Pip holding the teapot up to the camera in front of string lights at dusk.

Two shots of 5.04 seconds at 1280×720, joined with a one-second fade into 9.1 seconds. The glasses, the apron with its name, the speckled glaze and the bamboo handle are the same in both scenes. 290 credits: 145 per shot.

node madoo-install-example.mjs lab-reference-consistency --api <your Madoo API URL> --key <key.json> --run --save pip

The mascot and the teapot are images made with Madoo’s text-to-image node for this example; they show no real person or product.

A clip that goes on

ai/video_extend continues a clip past its last frame and returns only the new footage, 5 to 20 seconds of it. The clip you give it should be an MP4 under 15 seconds and 50 MB. Then:

Join it after the original with video/merge, the original in videos_0, the new footage in videos_1.
Join without a transition. Merge Videos fades between clips by default, for one second: on a continuation the fade eats a second of each side and blurs the very frame where the shot continues. Set transition to none.
Let the merge set the size. The model may answer at a slightly different size — 800×1088 for a 720×960 clip in the lab — and Merge Videos brings every clip to the first one’s.
Say how the shot goes on, not what the clip is: “the camera slowly pulls back and reveals the frame”.

The only model, FLUX 3 extend, costs 26 credits per second of new footage at 720p — 130 for five seconds — and about 1.3 times that at 1080p. The original’s length is not billed.

Lab: a clip that goes on

The lab’s workflow, left to right: Clip to continue; Continue the shot (Video Extend), giving the continuation; Original, then the new footage (Merge Videos), giving the whole clip; Measure the whole clip, giving the clip facts.

The five-second animation of Girl with a Pearl Earring from the first lab goes on for five more seconds: the camera pulls back and reveals the painting’s gilded frame on the museum wall.

The joint, frame by frame: the original at 4.95 seconds, the continuation at 5.05 seconds, the same picture, and the continuation at 5.6 seconds, the camera pulling back and the gilded frame appearing at the edges.

Ten frames of the whole clip, one per second: six close on the painting alive, then the camera pulls back and the painting appears in its gilded frame on a dark green wall, smaller and smaller.

10.04 seconds at 720×960, joined without a transition: the five new seconds arrived at 800×1088 and Merge Videos brought them to the original’s size. The continuation starts on the original’s last frame. 130 credits for the five new seconds; the clip facts are the final measurement.

node madoo-install-example.mjs lab-video-extend --api <your Madoo API URL> --key <key.json> --run --save extended

A product worn, and in motion

Two nodes turn a product photo into a fashion clip:

ai/virtual_tryon puts the product on a person. It takes the person’s photo (image 1) and the product’s (image 2), and keeps the person’s face and body as they are. Optional inputs:
- a background image, or else one of 41 environments, from a white studio to a Parisian café;
- a pose reference, whose pose only is taken, or else one of 15 poses;
- a blank product with an effect, to change a design’s finish.
It returns one image, or up to four variants. The default model, Nano Banana 2 edit, costs 5 credits per image and takes up to 14 images. A model that takes a single image would dress nobody, so the step stops before any charge (MODEL_TAKES_ONE_IMAGE).
ai/fashion_video animates it. It is Image to Video with presets written for products worn:

Product	Presets
Clothing	Model Walking, Model Turning, Fabric Close-up, Catwalk Strut, Fabric Flow, Pose Transition
Shoes	Shoe Walk, Shoe Close-up, Shoe Lifestyle, Shoe Turntable (no person)
Bags and accessories	Bag Showcase, Bag Detail, Accessory Lifestyle
Jewellery	Jewelry Sparkle, Necklace Reveal, Bracelet & Watch, Earring Turn, Ring Close-up

A custom prompt is added after the preset’s; a connected prompt replaces both. Its default model is Seedance Pro, at 8 credits per second: 40 for five seconds.

The try-on image keeps the aspect ratio you give it — square unless you set one — and the fashion clip follows it: set aspect_ratio on the try-on to 3_4 or 9_16 for a full-length, vertical clip.

Lab: from a product photo to a fashion clip

The lab’s workflow, left to right: the model and the garment go into Dress the model (Virtual Try-On), which gives the look; the look goes into A slow turn (Fashion Video), which gives the fashion clip.

A clothing shop has a flat photo of a linen overshirt and a studio photo of a model in a white t-shirt. The try-on puts the jacket on, worn open; Fashion Video’s Model Turning preset turns the model slowly on the spot.

Left to right: the model in a white t-shirt and black trousers; the mustard linen overshirt laid flat, with wooden buttons and two chest pockets; the model wearing it open over the t-shirt, the same face, pose and trousers.

Five frames of the clip, one per second: the model faces the camera, turns to her left, shows her back — the jacket’s back and hem — and turns back to face the camera.

The jacket keeps its colour, its buttons and its pockets through the whole turn, back included. 45 credits: 5 for the try-on, 40 for five seconds of video.

node madoo-install-example.mjs lab-tryon-fashion --api <your Madoo API URL> --key <key.json> --run --save look

The model and the jacket are images made with Madoo’s text-to-image node for this example; the model is not a real person.

When a step fails

Code	Meaning	Fix
`CONTENT_POLICY_VIOLATION`	the model refused a reference or a prompt — on Reference to Video, a real person’s likeness	use a character, a product or a synthetic person
`MODEL_TAKES_ONE_IMAGE`	Try-On or Image Fusion on a model that receives only one image	keep the node’s default model
`INVALID_REFERENCE_COMBINATION`	audio references without an image or a video	connect at least one image or video
`MISSING_PROMPT`	no prompt, preset or custom prompt	write one, or choose a preset
`PAYLOAD_VALIDATION`	the model needs something the step did not give — FLUX 3 First-Last Frame or PixVerse C1 Transition without a last frame	connect it, or choose another model
`RESULT_DOWNLOAD_ERROR`	the provider finished but its file could not be fetched; nothing is charged	run again