Running a Workflow Manually Through PatchWork¶

This is the foundation of generation — you, in PatchWork's web UI, clicking through the workflow node by node. Slower than automation, but you see every step and understand what's actually happening. Learn this first; the automation in Chapter 12 will then make sense.

What "manually" means¶

A workflow file (.nbflow) describes a graph of nodes — prompt nodes, image gen nodes, video gen nodes, approval gates. Manually running it means:

flowchart TB
    A[Open .nbflow in PatchWork] --> B[Configure G-Labs URL]
    B --> C[Click first image gen node]
    C --> D[Click 'Generate']
    D --> E[Wait for 4 candidates]
    E --> F[Click approval gate]
    F --> G[Pick best candidate]
    G --> H[Click first Veo node]
    H --> I[Click 'Generate']
    I --> J[Wait for 4 clips]
    J --> K[Pick best clip]
    K --> L[Repeat for each scene]

You're walking the graph by hand. Each gen takes 10-60 seconds. A 10-scene workflow takes 30-90 minutes of human-paced clicking.

Why learn this first¶

You can skip straight to the Generation Runner and have everything done in one command — but you'll be using a tool you don't understand. When something fails, you won't know where in the graph things went wrong or how to fix individual nodes.

Manual runs teach you:

What each node actually does
What order things run in
What an Approve gate is for (it lets you pick before downstream nodes consume the output)
What inputs each gen needs (prompt + references)
What the output looks like at each step

After 1-2 manual runs, the Generation Runner makes total sense. Before that, it's a black box.

Step 1: Open the workflow in PatchWork¶

Open https://patchwork-33m.pages.dev/ in a browser
Click Import Project
Select the .nbflow file you want to run

When you click New Project instead of Import, PatchWork shows the template chooser:

For running an existing workflow, always Import — don't pick a template. Templates are for building new workflows from scratch (which is what our pipeline does for you anyway). 4. The graph appears on the canvas — nodes connected by lines

If you have a multi-account workflow (multiple tabs), the tabs sit at the top of the canvas. Switch between them to work on each account's flow independently.

Step 2: Configure G-Labs URL¶

PatchWork needs to know where your local G-Labs backend is. The cloudflared tunnel URL changes every session (see G-Labs + cloudflared Setup).

In PatchWork, open Settings (in the sidebar)
Paste the current cloudflared tunnel URL into the G-Labs Server URL field
Save

PatchWork stores this in your browser's local storage. You only have to update it when the tunnel changes (every session).

Step 3: Run the first image gen¶

Find the first NanobananaAPI node in the graph. Click it to select.

A panel opens showing:

The prompt text feeding into it (from the connected Prompt node)
The reference images wired in (avatar reference, product photo)
The model and outputCount settings
A Generate button

Click Generate. PatchWork sends the request to G-Labs (via the tunnel). 10-30 seconds later, 4 candidate images appear in the node's gallery.

Step 4: Approve a candidate¶

Find the Approve node downstream of the image gen. Click it.

The Approve gallery shows the 4 candidates from the image gen. Click the one you want to keep. PatchWork remembers your pick.

The picked image now flows downstream — typically to a Veo3 node that needs the approved still as its start frame.

Step 5: Run the Veo3 node¶

Find the Veo3 node downstream of the Approve gate. Click it.

The panel shows:

The video prompt feeding in (from the Veo template / Dynamic Prompt)
The start frame (from your approved image gen)
The end frame (often the same image, or null)
Model and outputCount settings
A Generate button

Click Generate. Video gen takes longer than image gen — usually 30-90 seconds. After it finishes, 4 candidate clips appear in the node's gallery.

Step 6: Approve a clip¶

Click the Approve node downstream of the Veo3 node. Pick the best of the 4 clips.

That's one scene done. Move to the next.

Step 7: Repeat for every scene¶

For a 10-scene workflow:

10 image gen nodes to run
10 image Approve gates to click
10 Veo3 nodes to run
10 video Approve gates to click

That's 40 clicks across the workflow. At 30-60 seconds per generation, the full run is 30-60 minutes including thinking time.

You can:

Run them in order (Scene 01 → 02 → 03 → ...) — easiest to track
Run all image gens first, then all video gens (batch) — slightly faster if image gens fail and need re-rolls
Run multiple gens in parallel if PatchWork lets you (some installs do, some don't)

What happens if a generation fails¶

A failure shows up as an error in the node's panel:

"Connection refused" → G-Labs isn't running or tunnel is stale (see Setup checklist)
"API key invalid" → wrong key in G-Labs configuration
"Content policy violation" → prompt tripped a safety filter. Edit the prompt or re-roll the seed
"Timeout" → the gen took too long. Bump the timeout in node properties or just retry

Click Generate again to retry the same node. The error doesn't propagate — only the failed node retries.

Multi-account workflows¶

If your .nbflow has multiple tabs (one per account), you run each tab independently. The same 40-click process per tab. A 5-account, 10-scene workflow = 200 clicks if done fully manually.

This is where the Generation Runner starts to pay off (Chapter 12). But for your first few workflows, do at least one tab manually so you've felt every step.

What you learn¶

After your first manual run, you'll understand:

Why Approve gates exist — you control which candidate flows downstream
Why the graph is wired the way it is — image gens feed Veo gens; references feed image gens
What a node's properties panel shows — the prompt, the references, the settings
What can go wrong at each step — a failed gen, a missing reference, a stale URL

This understanding is the foundation for everything else in the wiki. The Generation Runner is a labor-saver, not a substitute.

Tracking your picks¶

PatchWork autosaves your picks. If you close the tab and come back later, your previous picks are still selected.

But the .nbflow file on your laptop doesn't update automatically — PatchWork stores picks in the browser. To persist your picks to the file, export the workflow:

File → Export (or Save) in PatchWork
PatchWork serializes the current state (including your picks) to a new .nbflow
Save it next to your source file

That exported .nbflow is what you'd hand to the next step (post-production, fan-out, etc.).

Common mistakes during manual runs¶

Forgetting to update the G-Labs URL: Tunnel URLs change every session. If you start a session with yesterday's URL set, every gen fails until you update it.
Running Veo before approving the image: Veo needs the approved still as its start frame. If you haven't clicked an approval in the image's Approve gate, Veo has nothing to use. PatchWork should flag this — but it's still easy to forget the order.
Running all 40 nodes without checking outputs: If image gen Scene 03 looks wrong, you don't want to waste Veo budget generating videos for it. Check the image's candidates before running the Veo node.
Not exporting after picks: You spent 90 minutes picking candidates. You close the tab. Your picks are still in browser local storage — but if you open the workflow elsewhere, the picks aren't there. Export to persist.

When manual makes sense vs. when to automate¶

Situation	Manual	Generation Runner
First time using the pipeline (learning)	YES	No (yet)
Validating a single scene	YES	No
Iterating on a prompt by hand	YES	No
Production workflow (multiple scenes, multiple accounts)	No	YES
Time pressure	No	YES
Need batched parallel generation	No	YES

After a few manual runs, you'll naturally graduate to the Generation Runner for production work.

When you're ready¶

→ Next: Reviewing in PatchWork — deeper detail on the candidate-picking flow, especially for multi-account workflows.