gflow-studio

English · Tiếng Việt

Generate images and videos with Google Flow from your terminal. gflow-studio drives Imagen and Nano Banana (images) and Veo / Omni Flash (videos) through your own Flow account, so a whole storyboard can run as one batch command instead of an afternoon of clicking.

gflow-studio is maintained by ABM.DũngTQ for an animation pipeline. It is built on ffroliva/gflow-cli, keeps every upstream feature (currently 0.91.0) and adds 1080p downloads, a hidden browser on macOS, safer batches and a video batch runner. The command is still gflow, so scripts written for gflow-cli keep working. The full list is under What this edition adds.

What it does · What this edition adds · Setup · Usage · Batches · Hiding the browser · Demo · When something fails · Documentation · Author and credits

⚠️ Read this before you install. gflow-studio is not affiliated with Google. It reverse-engineers Flow and drives a real Chrome window signed in to your Google account, so automation happens on your account and under Google's terms. Flow can change its pages at any time. Images are free; Veo video generation spends your Flow credits. Read the DISCLAIMER and docs/ACCOUNT_SAFETY.md.

What it does

Every command takes --profile to pick a Google account and --project to work inside an existing Flow project. Most commands take --json for scripts. gflow docs prints the bundled documentation offline.

Related MCP server: Google Flow Browser MCP

What this edition adds

On top of upstream 0.91.0:

• --upscale: keep the 1080p clip. After a video downloads, Flow's 1080p export (no extra credits) replaces the 720p file at the same path. If the export fails, the 720p clip is kept and the command prints a retry command. Works on video t2v, i2v, r2v, video batch and the MCP tool gflow_generate_video(upscale=True).

• gflow video batch. Runs 1–50 image-to-video rows from a JSON file in one browser session and one Flow project, with --resume, --jitter and --json. A clip that was generated and billed but failed to download is fetched again for free instead of being lost.

• A hidden browser on macOS. GFLOW_CLI_WINDOW_MODE=background runs Chrome as a background-only app: no Dock icon, and it does not take keyboard focus while you work.

• Batches stop when Google pushes back. gflow run, image batch and multi-prompt image t2i stop at the first WAF 403, rate limit or expired session, even with --continue-on-error, and exit with that error's code. A flagged account no longer receives the rest of the batch.

• Named references in gflow run. Give a reference image a name ({"name": "cuong", "path": "cuong.png"}) and write @cuong in the prompt. A row can take several references (refs) and its own output path.

• Automatic session renewal. If the Google session expired before anything was submitted, gflow renews it once and retries. It never retries after a generation was sent. Turn it off with GFLOW_CLI_AUTO_RENEW=false.

• Fixes. -o on another disk no longer fails after generation. Long image batches re-upload a reference that scrolled out of the project. The Chrome tab is brought forward only when it is actually hidden.

CHANGELOG.md § Unreleased has the details. The headless measurements behind the browser work are in docs/superpowers/spikes/2026-10-28-headless-generation.md.

Setup

Requirements

• A Google account with access to Google Flow. Video generation needs Flow credits.

• Google Chrome installed normally.

• uv, which installs Python 3.11+ for you.

• macOS, Windows or Linux with a desktop. Chrome must be able to open a window for the one-time sign-in.

1. Install

uv tool install "git+https://github.com/abm-dungtq/gflow-studio.git@main"
uvx --from "git+https://github.com/abm-dungtq/gflow-studio.git@main" playwright install chromium
gflow --version

To upgrade later, run the same uv tool install command with --force. The package on PyPI (uv tool install gflow-cli) is the upstream project and does not include these features.

2. Sign in once

gflow auth login --browser chrome --profile main

A Chrome window opens. Sign in to Google and wait until the Flow workspace appears; gflow saves the session in its own profile folder. --browser chrome matters: generation later reopens the profile as real Chrome. Repeat with another --profile name for each Google account.

3. Check

gflow auth status
gflow credits user --profile main
gflow doctor

None of these spend credits. gflow doctor exits 33 when it finds something to fix and tells you what.

4. Optional settings

Copy .env.template to .env in the folder you run gflow from, or to the gflow home folder for a machine-wide setting. Useful ones:

Every setting is documented in docs/CONFIGURATION.md. Only .env is read; a .env.local file is ignored.

Usage

Images (free)

gflow image t2i "a hot air balloon over Tokyo at sunrise" --aspect 16:9 -o out/balloon.png
gflow image t2i "a cat mom, flat 2D" --model nano2 -n 4 --out out/cats/
gflow image i2i "the same cockroach, now sitting on a keyboard" --ref refs/cuong.png -o out/k1.png

Models: nano2 (Nano Banana 2.1, default), nano2-lite, nano-pro, imagen4. Aspect ratios: 16:9, 4:3, 1:1, 3:4, 9:16. -n makes 1–4 variations.

Videos (spend credits)

gflow video t2v "slow cinematic push-in on a sunlit forest clearing" --aspect 16:9 -o out/forest.mp4
gflow video i2v "the paper lantern sways gently, slow push in" \
  --initial-frame out/lantern.jpg --model veo-lite --aspect 16:9 --upscale -o out/lantern.mp4
gflow video upscale <media_id> --scale 1080p      # 1080p for a clip you already made

Models: omni-flash, veo-lite, veo-fast, veo-quality. --duration is 4, 6 or 8 seconds; 10 seconds is Omni Flash only. --end-frame sets the last frame for i2v.

Projects, characters and catalog

gflow project list --profile main          # projects gflow knows about
gflow character create --project <id> --name "Aria" --face-prompt "..." --body-prompt "..."
gflow data --help                          # query everything you generated

For AI agents (MCP)

gflow mcp setup          # register the stdio server with your agent
gflow mcp run --no-spend # a server without the generation tools

The CLI and the MCP server expose the same features. docs/MCP.md lists every tool.

Batches

An image storyboard: gflow run

storyboard.json:

{
  "prompts": [
    {"text": "@cuong stands on a desk at night, flat 2D", "aspect_ratio": "16:9",
     "refs": [{"name": "cuong", "path": "refs/cuong.png"}], "output": "gen/k1.jpg"},
    {"text": "@cuong jumps onto the keyboard", "aspect_ratio": "16:9",
     "refs": [{"name": "cuong", "path": "refs/cuong.png"}], "output": "gen/k2.jpg"}
  ]
}

gflow run --config storyboard.json --profile main --project <id>
gflow run --config storyboard.json --profile main --project <id> --resume

Each reference file is uploaded once per run. --resume skips rows whose output already exists, so a rerun after a failure only makes what is missing.

Image-to-video in one session: gflow video batch

videos.json:

{
  "project": "<existing-flow-project-id>",
  "videos": [
    {"start": "gen/k1.jpg", "prompt": "the cockroach looks around, slow push in",
     "model": "omni-flash", "duration": 4, "aspect": "16:9", "output": "clips/v1.mp4"},
    {"start": "gen/k2.jpg", "prompt_file": "prompts/v2.txt",
     "model": "omni-flash", "duration": 6, "aspect": "16:9", "output": "clips/v2.mp4"}
  ]
}

gflow video batch --config videos.json --profile main --upscale --json

A submitted row is never sent twice, because a clip Flow accepted may already be billed. If a row fails, rerun with --resume.

Hiding the browser

Google only serves Flow to a real browser, so gflow drives a real Chrome window. How much of it you see depends on the platform:

Headless real Chrome generated images 6 out of 6 times on one test account (macOS, 2026-10-28), with no 403 and no reCAPTCHA error. Its user agent says HeadlessChrome, and video, Windows and long-term account effects were not measured, so headed Chrome stays the supported default. Try headless on a test account first.

Demo

A real two-row gflow run on a test account (2026-10-29, about one minute, no credits):

$ gflow run --config run.json --profile thu-headless --project <project-id> --output-dir tmp/f7/out
  project: <project-id> (existing)
                                   gflow run
┏━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ # ┃ prompt                      ┃ ratio ┃ status ┃ detail                    ┃
┡━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ 0 │ a small red paper lantern   │ 16:9  │ OK     │ tmp/f7/out/prompt_0_0.jpg │
│   │ on a wooden table, soft     │       │        │                           │
│   │ light                       │       │        │                           │
│ 1 │ a green ceramic teapot on a │ 16:9  │ OK     │ tmp/f7/out/prompt_1_0.jpg │
│   │ windowsill, morning light   │       │        │                           │
└───┴─────────────────────────────┴───────┴────────┴───────────────────────────┘

2/2 succeeded · 0 failure(s) · 0 skipped

The first lantern image was then animated with gflow video i2v --model veo-lite --upscale on the same account. The result was a 4-second clip at 1920×1080, with "upscaled": "1080p" in the JSON output.

When something fails

gflow fails loudly with a stable exit code rather than resubmitting, so a broken run does not silently spend credits.

Every error is printed as RFC 9457 problem details with a remediation hint. KNOWN_ISSUES.md lists current problems and workarounds.

Documentation

docs/INDEX.md indexes everything.

Support the original project

gflow-cli is built by one person upstream, and every release there is verified against real Flow with real credits. If this tool saves you time, sponsor the original author:

Tiers are listed in docs/SPONSORS.md.

No sponsors yet. Be the first — every public sponsor is listed here.

Author and credits

• gflow-studio: ABM.DũngTQ · 0976202028 · github.com/abm-dungtq/gflow-studio

• Original project: gflow-cli by Flavio Oliva (ffroliva). The core of this tool is Flavio's work; fixes that apply upstream are worth sending there.

• License: MIT, © 2026 Flavio Oliva, kept unchanged in this fork. The license covers the code only. Google's own terms govern Flow and everything you generate with it. See the DISCLAIMER.