From analysis to generation: a deep walkthrough of the TrendRemix workflow
A step-by-step walkthrough of what happens between uploading a reference video and rendering a new one — including the canvas model, async tasks, and where credits get spent.
This article walks through every step of the TrendRemix workflow from the moment you drag a video into the workspace until the moment a new video shows up on the canvas. It is the article we wish we had when we onboarded the first ten users.
Step 0 — Create a project
Every workflow lives inside a project. A project is the container for your uploaded sources, your analysis output, your chat messages, your canvas, and your generated assets. We make a new project the moment you open /create — you do not have to click anything.
Step 1 — Upload the reference video
Drag a mp4 / mov / m4v / webm file onto the workspace, or paste a URL. The file goes to Cloudflare R2 under your project's namespace. Maximum 50 MB per file. Once the upload completes, an analysis task is queued and a placeholder node appears on the canvas.
Step 2 — The two-phase analysis
Analysis runs as a background task that survives page reloads. Phase A (Observation) extracts grounded facts: assets, segments, on-screen text layers, speaker characteristics, pacing. Phase B (Synthesis) turns those facts into a recreation blueprint. Each phase costs credits and is checkpointed — if Synthesis fails, Observation is reused on retry without re-charging.
While the task runs, the placeholder node on the canvas shows a progress indicator. Server-Sent Events (SSE) push status updates to your browser in real time. You can navigate away, come back, and the canvas resumes from the latest state.
Step 3 — Read the canvas
When analysis finishes, the canvas explodes into a dozen new nodes: the strategy card, the emotion-curve chart, the segment table, the selling-point map, the mechanics card, the timeline, and the global prompts (one for image, one for video). Each node is editable. Each node has a stable ID. Each node feeds the next generation.
Step 4 — Chat to rewrite
The chat panel on the left is where you direct the remix. Tell the AI which audience to target, which selling points to swap, which CTA to use. Every instruction either updates an existing canvas node or branches a new one. The agent has the full canvas in its context, so you do not need to re-explain the structure each turn.
Step 5 — Generate
Type "render this" or use the generation control on any canvas shot node. The system bundles the analyzed blueprint plus your accumulated chat instructions into a generation request and sends it to Nano Banana (for images) or Seedance (for video). The task is async — the canvas shows a placeholder, you can keep chatting, and the result lands when ready.
Generated assets are first-class canvas citizens. You can branch them, re-prompt them, or chain them into another generation.
Step 6 — Branch and version
Almost every remix worth shipping comes from a branching session, not a single linear pass. Right-click a shot node and select "Branch" to fork the canvas at that point. Use branching to A/B test alternate hooks, alternate CTAs, alternate brands — without losing the parent version.
Step 7 — Export
Generated assets can be downloaded individually or exported as a project bundle. The export captures the canvas state, the analysis JSON, and all generated media — so you can hand the bundle to a video editor or archive it for compliance.
What gets charged
- Video analysis: roughly 5-10 credits per video, charged on task creation, refunded if the task fails before completing
- Image generation: 1-3 credits per image, depending on quality band
- Video generation: 10-30 credits per video, depending on duration and model tier
- Storage and SSE bandwidth: free
Where to go next
Open the workbench, upload a reference, and watch the workflow unfold. The first analysis takes about a minute. The first remix takes about five. After that, the loop tightens fast.