Seedream workflow examples: 4 templates on Picsart Flow

Seedream is an image model, and in these Flow templates it is doing a job most people assume belongs to a video model. It draws the frames that a second model then animates, which puts it at the planning stage of a video rather than the render stage.

That split explains why Seedream templates look so different from the rest of the catalog. Most of them hand you a panel sheet before they hand you a clip.

The four workflow examples below run that pattern across very different jobs, from a two-image restyle that produces no video at all to a vertical product spot built around a single cup.

How Seedream is mostly used in Picsart Flow

Seedream runs perfectly well on its own, and across Picsart it usually does, generating a finished image straight from a prompt. The Flow template catalog is where it tends to work alongside a second model instead, most often Seedance.

Those pairings are built as a relay. The image model establishes what a scene looks like and the video model inherits that decision, which puts most of the creative control in the first half of the chain.

The relay has a failure mode the templates plan for. A video model handed a grid of six panels will happily animate the grid itself, borders and all, so the prompts on the video node spell out that it must not.

The four examples below cover both kinds. One is a pure image job that finishes with stills. The other three hand off to a video model, so a board arrives first and a clip follows.

Which Seedream version shows up where

Open any of these templates and the model picker tells you exactly which Seedream is loaded, and the four do not all use the same one.

Seedream 4.0 runs the restyling job, where the model edits an image that already exists rather than inventing a scene. Seedream 5.0 Lite runs both multi-panel boards, the anime grid and the product board, where a single frame matters less than getting six of them consistent.

Seedream 5.0 Pro runs the hardest job in the set, the board that has to reconcile three separate reference images into one coherent scene. That is the only one of the four reaching for Pro.

Here is the whole set at a glance before the breakdowns.

AI style transfer: one reference, two models

The Style Extractor template is the outlier here, and it is the best place to start because nothing moves. You upload the image you want restyled and a second image carrying the style you want, then run the frame.

What sits between them is the interesting node. A style analyzer reads the reference and writes the look out as structured data, naming the palette down to individual hex values, the lighting type and contrast, the mood, the surface textures and the recurring geometry.

The restyle then works from that spec rather than from a vague instruction to make it look similar.

The two outputs are not two attempts at the same thing. One comes from Seedream 4.0 and the other from Gemini, both driven by the same analyzed style and the same instruction to change the aesthetic without changing the content.

Running them side by side turns a restyle into a comparison. The frame is set to a square 1:1 at 2160p.

Anime storyboard: a labeled six-shot grid

One portrait goes in and a 3×2 grid comes out, each cell labeled Shot 1 through Shot 6. The anime storyboard template then animates that sequence into a short scene, running Seedream 5.0 Lite at 16:9 and 1440p.

The prompt driving the board is almost entirely an identity lock. It lists what has to stay the same in every frame, item by item: facial identity, facial proportions, hairstyle, hair color, eye color, skin tone, age and overall likeness.

It closes by telling the model not to redesign, replace or reinterpret the character.

The video node takes a different line. It treats the board as narrative and pacing reference only, then asks for original cinematography on top of it, with its own camera angles and framing. The board decides what happens in each shot, not how the shot is composed.

Action previs: three references into one board

Previs exists for sequences too complicated to plan in your head. The sci-fi laundromat template takes three separate uploads, a person, a location and a creature, and resolves them into one 3×2 board of six landscape panels before any video exists. It runs Seedream 5.0 Pro at 16:9 and 1440p.

Three inputs is the most demanding ask in this set, and the prompt handles it by addressing each upload directly. Every reference gets its own token, so the prompt can say which image is the person, which is the location and which is the creature, and insist those are the only identity references allowed.

It then asks for identical identities, wardrobe, layout, props, lighting and continuity across every panel.

The video prompt is written as a timecoded shot list rather than a description. It opens on a wide shot in the first three seconds and specifies no dialogue, which keeps the attention on staging and impact. That is exactly what previs is for.

AI product video: a vertical spot with one consistent prop

Commercial work follows different rules, and the boba tea template shows why. A product photo goes in, gets its background removed, and only then feeds the board. Seedream 5.0 Lite builds the panels at 9:16 and 1440p.

The continuity anchor is a product rather than a character. The prompt tells the model to reference the cup, the straw and the pearls exactly as they appear in the uploaded photo, and fixes a single palette across all six panels. A cup that subtly changes shape between shots breaks a commercial in a way a shifting background never would.

The video prompt goes further than any other in the set by assigning each panel a duration. Frames one through five get roughly two seconds each, and the final frame is stretched to around three and a half so the clip resolves instead of cutting off. This is the only example built for a phone.

Running one, and changing it

Step 1. Open the closest template

Pick the example nearest your job and open it in Picsart Flow. The chain structure transfers even when your subject has nothing in common with the sample.

Step 2. Replace the inputs

Swap the sample assets for your own. These four differ a lot in what they ask for, from three separate reference images down to a source photo plus a style reference.

Step 3. Rewrite the text first

Edit the scene description and the shot notes before running anything. Words are the cheapest thing to iterate on, and they decide everything the image stage produces.

Step 4. Stop at the image stage

Generate the board or the stills and actually read them. Fixing one panel now costs a fraction of regenerating finished video later, and this is the whole reason the chain is split in two.

Step 5. Animate and export

Send the approved frames through to the video stage, then export. Exporting usually requires signing in. The Flow editor keeps earlier versions, so a board you liked better is always recoverable.