Flux workflow examples: build a matched set of images in one canvas

A Flux workflow turns a single input into a finished set of images instead of one image you regenerate five more times until something sticks. That is the real gap between prompting a model and building around it. Three templates in Picsart Flow show the gap clearly, and they line up at three different levels of structure. The first hands Flux a photo and one instruction. The second hands it a block of text and lets the model improvise every layout decision. The third splits the job into content, outline and design system, then runs them together.

Read in that order, the three templates stop being three unrelated builds and start being a ladder. Every rung trades away a little improvisation for a lot of consistency.

What changes when Flux runs inside a workflow

A single prompt produces a single result, and rerunning it produces a different one. Nothing carries over. The layout drifts, the palette shifts, and matching two outputs to each other becomes a manual job.

A workflow fixes the parts that should not move. The input node holds the source material once, and every output node reads from that same node rather than from a prompt you retyped. Change the source and all the outputs change with it. Change one output and the rest stay exactly where they were.

All three templates below run Flux 2 Pro, which handles the pieces this kind of build depends on:

• Output up to 4MP, at any aspect ratio down to 64×64
• Text rendering inside the image, so labels and headings are part of the generation
• Hex color matching, which is what makes a swatch-based layout possible at all
• Pose guidance for compositions built around a person
• Structured prompting, so a long instruction stays readable to the model

The three templates also run at three different shapes. The palette build outputs 9:16, the collage build outputs 16:9, and the infographic build outputs portrait. That is set per output node, not globally, so one workflow can feed several placements at once.

The three levels of structure

Each level asks for more preparation and gives back more control. Pick by how much the outputs need to match each other.

Level one: one image, one instruction

The Create Fashion Color Palette template is as small as a workflow gets. Two nodes sit side by side. An image node holds the upload, and a single output node holds the instruction. One edge connects them.

What makes it worth studying is that the instruction is pure layout direction. It does not describe a subject, a mood or a style. It describes where rectangles go, how they stack, and how their labels are typeset. Flux reads the colors out of the uploaded image and builds the overlay around them.

That separation is the useful part. Swap the photo and everything regenerates against the new colors while the layout holds its position. A run against twelve outfit images returns twelve palettes that stack the same way, which is what makes the output usable as a series rather than a one-off.

Fashion color palette overlay

Create a minimal fashion color palette layout over the uploaded photo. Keep the original photo as the background. Add three large horizontal rectangular swatches centered in the composition, stacked vertically with even spacing. The rectangles should stay in the center regardless of the photo. Use clean edges, no shadows, and small monospaced hex code labels in the bottom-left of each rectangle.

Copied

Two details in that instruction do most of the work. “The rectangles should stay in the center regardless of the photo” pins the composition against a model that would otherwise rearrange it per image. “Clean edges, no shadows” removes the depth treatment Flux tends to add on its own.

Level two: one text block, six interpretations

The Create Photo Essay Collage template runs on a single text node holding roughly 600 words of source material. That node fans out to six output nodes, and each one returns a full editorial spread: period images, headings, body columns and handwritten-style annotations, arranged as a finished page.

Nothing tells Flux how the page should be organized. There is no outline, no grid spec and no type direction. The model reads the text, decides what deserves to be a headline, picks which moments become images, and lays the result out six separate times.

The six spreads come back genuinely different from each other. Some lead with one hero image and a column of body text. Some split into a four-panel grid. That variance is the point of this level, and it is why the template suits the stage where the direction has not been chosen yet.

The sample text shipped with the template is a history overview, but the node has no opinion about subject matter. A brand story, a launch narrative, a research summary or a long-form case study all behave the same way. Paste, run, and read six directions back.

Level two is the right call when the goal is to see options. It is the wrong call when the outputs need to match, because six independent interpretations will not agree with each other on type, palette or grid.

Level three: content, outline and style as separate nodes

The Create Educational Infographic template answers the matching problem by splitting the input into three nodes instead of one.

• The source node holds the raw material, about 600 words of it, written as plain prose with no formatting
• The outline node holds roughly 1,400 words of pre-written structure, broken into numbered sections with the key points already bulleted under each one
• The style node holds around 100 words of design system: two typefaces with a rule for where each is used, plus a color direction limited to one primary, one accent and neutrals

All three fan out into the same bank of output nodes. The result is a grid of variants that share a visual language, because one style block reaches the whole set instead of being retyped per image. Each output combines columns of body text with charts, timelines and diagrams, and they agree with each other on type and palette.

The outline node is the piece that separates this build from level two. Structuring the content happens in text, before any image exists, which means the sequencing gets fixed while it is still cheap to change. Flux is left to handle layout and illustration rather than editorial judgment.

Reusable style and color direction

Use Helvetica Bold for all titles, section headings, and key statistics. Use Helvetica Regular for all body text, captions, labels, and references. Maintain a clear visual hierarchy with generous spacing, consistent typography, and excellent readability. Keep the overall style clean, modern, professional, and suitable for an academic research report. COLOR DIRECTION: Use a limited color palette with one primary color, one accent color, and neutral tones. Automatically select the most appropriate palette based on the presentation topic to reinforce its mood and subject matter. Maintain consistent colors across all slides while using accent colors only to emphasize important information, data, and key takeaways.

Copied

That block is worth lifting into any build that needs a consistent set. Nothing in it names a topic. Swap “academic research report” for the register the project needs, leave the rest, and it becomes a house style that travels between workflows.

How to build a Flux workflow from scratch

Step 1. Open a blank canvas

Start in the Flow editor and add a frame. The frame keeps related nodes grouped, which matters as soon as the build has more than a handful of them.

Step 2. Decide what stays fixed

Name the thing that should not change between outputs. That is the source, and it belongs in its own node. Everything else in the build reads from it.

Step 3. Separate the style from the content

Put design direction in a second text node rather than folding it into the content. Editing type and color without touching the source material is the payoff, and it arrives the first time a revision comes in.

Step 4. Write the outline before generating

Draft the structure as text, section by section. Reordering a bulleted outline takes seconds. Reordering ten generated images takes a full rerun.

Step 5. Set the ratio per output

Choose the aspect ratio on each output node. One workflow can produce a 9:16 crop and a 16:9 spread from identical inputs.

Step 6. Fan out, then prune

Add several output nodes and run them together. Reading six results side by side beats judging one in isolation, and the weak ones are simply deleted.

Start at the level the project needs

Match the rung to the job rather than climbing to the top by default. One changing image with a fixed treatment stays at level one. An open question about direction belongs at level two. A set that has to look related is worth the outline and the style block that level three asks for.

Open any of the three templates, replace the sample input with real material, and run it before changing anything else. The shape of the build teaches faster than the settings do.