ai-chatbot/docs/nanobanana-browser-poc.md

Nano Banana Browser POC

This repo includes a small browser-driven POC for inspecting Nano Banana's free-tool flow without trying to bypass Turnstile, login, or rate limits.

What it does

• Opens the Nano Banana page in a real browser
• Optionally prefills a public image URL
• Optionally clicks the primary action button
• Captures requests and responses for the relevant private endpoints
• Saves screenshots, console logs, downloads, and browser storage state

What it does not do

• It does not bypass Cloudflare Turnstile
• It does not bypass login
• It does not try to evade rate limits
• It is not a production API adapter

Commands

Remove background:

pnpm poc:nanobanana --mode remove-background --auto-submit

Upscale:

pnpm poc:nanobanana --mode upscale --auto-submit

Use a custom image URL:

pnpm poc:nanobanana \
  --mode upscale \
  --image-url "https://example.com/demo.jpg" \
  --auto-submit

Useful flags

• --locale en
• --headless
• --no-auto-fill
• --auto-submit
• --close-on-success
• --timeout-ms 300000
• --out-dir /absolute/path/to/output

Output

By default, artifacts are written to:

playwright/nanobanana-poc/<mode>-<timestamp>/

The folder includes:

• summary.json
• session-plan.json
• storage-state-initial.json
• storage-state-final.json
• requests/*.json
• responses/*.json
• logs/*.json
• screenshots such as 01-loaded.png and 99-final.png

Current endpoints observed

• /api/img/rmbg
• /api/img/upscale
• /api/img/result/{taskId}?type=img
• /api/img/image-safety
• /api/presigned-urls

Current findings (2026-05-21)

These findings were confirmed from the live frontend bundle and manual browser validation.

Remove background

Frontend schema:

• imageSource: "upload" or "url"
• imageUrl: optional string
• turnstileToken: optional string

Observed request shape:

{
  "imageSource": "url",
  "imageUrl": "https://...",
  "turnstileToken": "..."
}

Observed endpoint behavior:

• POST /api/img/rmbg
• Content-Type: application/json
• success path reads response.data as the final output image URL
• after success the page clears turnstileToken
• 403 with message === "nsfw" is handled as an NSFW block

Manual validation in the in-app browser produced a successful result and exposed a final image URL hosted on replicate.delivery, which suggests the private backend may fan out to third-party storage or processing before returning the final asset URL.

Image upscaler

Frontend schema:

• imageSource: "upload" or "url"
• imageUrl: optional string
• scale: number, min 1, max 8, default 4
• turnstileToken: optional string

Observed submit request shape:

{
  "imageUrl": "https://...",
  "scale": "4",
  "turnstileToken": "..."
}

Upload mode can also send an image field populated from the uploaded asset URL.

Observed endpoint behavior:

• POST /api/img/upscale
• Content-Type: application/json
• success path reads taskId from response.data.taskId or response.taskId
• the frontend then polls GET /api/img/result/{taskId}?type=img
• polling treats code === 1 or status === 1 as success
• final output is read from data.imgAfterSrc or imgAfterSrc
• polling treats code === 2 or status === 2 as failure
• the polling loop waits 2 seconds and retries up to 60 times
• 402 is handled as an upgrade/paywall condition
• 403 with message === "nsfw" is handled as an NSFW block

Turnstile

Observed Cloudflare site key:

• 0x4AAAAAAAh8CLI0IWTYOWMY

Observed frontend behavior:

• when turnstileToken is missing, the frontend renders or executes Turnstile before submitting
• after a successful request, the frontend clears turnstileToken
• this means the free-tool flow is not designed as a stable unattended API

Upload helper

Observed upload helper behavior:

• POST /api/presigned-urls with JSON like { "count": 1, "fileType": "<mime>" }
• response returns urls
• the browser PUTs the file to the returned presigned URL
• the frontend then derives a public asset URL in the form:

https://temp.videostudioai.com/<key>

Practical takeaway

This is enough to build an internal unofficial wrapper, but not enough to remove the core operational risk:

• the flow still depends on Turnstile
• Turnstile tokens are intentionally short-lived and cleared after use
• the frontend contract may change at any time
• upscaling still appears more likely to require session state than remove-background

Suggested use

Use this only as an internal validation tool to answer questions like:

• Does the browser session create anonymous or guest state?
• When does Turnstile appear?
• Which cookies or local storage values appear before the request?
• Does upscaling require an authenticated session?
• What request and polling shape does the frontend use today?