v0.4Seedance + Seed 2.0, wired into a loop

Stop guessing.
Start retaking.

Retake wraps BytePlus Seedance in a self-healing feedback loop. Generate a video, score it, and if it fails — rewrite the prompt and try again. Up to three takes, no human in the loop.

Get started — freeOpen playgroundno credit card · 50 retakes free
retake · playground — autonomous mode
generate
score
heal
Prompt · iter_01
composing prompt…
live
iter_01 · attempt
Prompt fit
Temporal
Physics
self-healing promptsCI for video modelsbatched retakes3-dim scoringBytePlus SeedanceSeed 2.0 evalregression baselinesparallel generation
self-healing promptsCI for video modelsbatched retakes3-dim scoringBytePlus SeedanceSeed 2.0 evalregression baselinesparallel generation
The two problems

Video AI is a slot machine. Retake makes it a lathe.

Two failure modes, one loop. A feedback signal where there was silence, and a regression net where there was none.

01 / slot machine
⚙ → 🎲 → ?

You prompt, it generates, you have no idea why it failed.

Today's video models are black boxes. Bad physics? Temporal jitter? Wrong subject? You get a video — not a reason. So you reroll, reword, and hope.

Retake scores every frame
02 / regression
v1.0 ✓ → v1.1 ✗

The prompt that worked yesterday silently breaks today.

Change a system prompt, swap a template, upgrade a model — and a dozen workflows regress without warning. There's no git blame for video quality.

Retake is CI for prompts
The loop

Generate. Score. Heal. Retake.

A single prompt becomes up to three takes, each informed by the score of the last.

01

Prompt

You write it once. Retake treats it as a seed, not a spell.

02

Generate

Seedance renders your video at 720p, 5 seconds, 24fps.

03

Score

Seed 2.0 grades it across prompt fit, temporal, and physics.

04

Retake

Below threshold? Prompt rewritten with the failure in mind, regenerated.

01

A playground that talks back.

Interactive mode gives you the steering wheel — see every score, add a remark, decide whether to heal. Autonomous mode runs the whole loop and streams progress live.

  • Three-dimensional scoring on every take
  • Inline preview of the healed prompt
  • Add personal remarks before a retake
  • Cancel autonomous runs mid-stream
running
iter_01 · 6.3
generating…
Prompt
8.2
Temporal
6.4
Physics
4.3
iter_02 · 8.9 ✓
generating…
Prompt
9.1
Temporal
8.7
Physics
8.9
healer →added "rain on glass" + "24fps" · temporal coherence restored
benchmark_v4.yaml · 12 prompts
11/12 pass
cyberpunk owl, rain, neon
7.2
8.9
✓ +1.7
astronaut gardening on mars
6.8
7.4
✓ +0.6
candle in hurricane wind
8.1
8.3
✓ +0.2
marble rolling uphill, slow-mo
7.9
5.2
✗ -2.7
golden retriever skateboarding
8.3
9.1
✓ +0.8
02

CI for your prompts.

Define a suite of benchmark prompts and thresholds. Run them before every template change. Retake grades each one against a saved baseline and gives you a clean pass/fail.

  • Per-prompt score thresholds
  • Baseline deltas with +/- indicators
  • Suite passes only if every prompt passes
  • Snapshot baselines on demand
03

Batch, in parallel, with the receipts.

Paste twenty prompts, walk away. Retake runs two concurrently, streams progress into a live grid, and hands you a CSV when done.

  • CSV upload or one-per-line paste
  • Live grid with status per card
  • Two prompts running concurrently
  • Export everything as CSV
batch_2604 · 14 of 20 complete
2 running
generating…
✓ 7.5
generating…
✓ 7.8
generating…
✓ 8.1
generating…
✓ 8.4
generating…
✓ 8.7
p_06
gen
p_07
p_08
14 / 202:41 remaining
The engine room

Two models, one loop.

We didn't train a model — we orchestrated two of the best ones. Generation and evaluation in a choreographed handoff.

Generation

BytePlus Seedance

State-of-the-art text-to-video. 720p, 5-second clips at 24fps. Retake handles prompt templating, parameter tuning, and retries.

resolution 1280×720
duration 5s @ 24fps
avg gen time 38s
Evaluation

Seed 2.0 Vision

A vision-language model that watches your video and scores it across prompt adherence, temporal consistency, and physical logic.

dimensions 3
scale 0 — 100
avg eval time 4s
Ready to retake?

Three
takes, one
that lands.

Create free accountTry the playground