Reel Lab — Project Overview

The single reference page for the Reel Lab video studio: every feature we have discussed, the build phases, the technology underneath, and the decisions still open. Use this to add or edit items instead of re-discussing them.

Open the Studio Open Staging GitHub Repository

Feature Goals

Build Phases

Tests Passing

~$0.16

Per Video Clip

What this is: a private video studio at vstudio.viewervelocity.com — feed it a photo, idea, or reference and it makes realistic short meme and ad videos with consistent characters, captions, voice, and sound. Low cost. Personal use, not public.

Working rule: one capability at a time. Get it working, review it, then add the next. Cost is tracked on everything. No automatic paid retries and no premium model fallbacks. We label honestly what is and is not production-ready.

How It Works — The Video Pipeline

Everything in the studio serves this pipeline. The interface follows the pipeline, not the model names.

1 · Gather

→

2 · Generate

→

3 · Review

→

4 · Finish

→

5 · Export

→

6 · Save & Reuse

1 · Gather

Pick the inputs: photo, reference video, character, voice, saved prompt. Everything reusable lives in the Library.

2 · Generate (the only paid step)

Starting image, video clip, voice. Each becomes a Job with a recorded cost.

3 · Review

Preview, then decide keep / retry / save. Nothing is charged for looking.

4 · Finish (free, local)

Captions, music, sound, graphics, trimming. Done with FFmpeg on the PC — no re-spend.

5 · Export (free, local)

Deliver in the shapes needed: portrait 9:16, square 1:1, landscape 16:9.

6 · Save & Reuse (free)

Anything worth keeping goes to the Library, tagged, ready to feed the next video.

Current Status — Honest Assessment

Capability	Status	Notes
Simple prompt or base image → small motion clip	Strong	Our best and most reliable output
Basic captions / text overlay	Works Locally	Existing local overlay; not proof that the caption options feature is complete
Caption options: full word-for-word and important-words-only	Not Completed	Early priority; both modes need a usable workflow and tests
Continuations (last frame → next clip)	Decent	Coffee and shower continuations worked
Consistent character (stills)	Decent	Reference-image approach; not trained characters yet
Cost tracking and persistence	Done	Live in the panel
Timelapse / cleanup transformations	Weak	Garage test: clutter floated and disappeared. Physics not there yet
Real movements (cart, grill, stove)	Mixed	Kling motion-control: ghosting, duplicate people
Video mimic / human replacement	Unreliable	Ghosts and artifacts. Wan Animate still untested
Talking head / short yap videos	Not Built	Research done; needs text-to-speech and lip-sync
Podcast	Not Built	Research done; do edited shot / reverse-shot first
Voices / text-to-speech	Not Built	Easy to add; cheap (pennies per clip)
Music / sound effects	Not Built	Local mix with FFmpeg; free
Motion graphics	Not Built	FFmpeg can do it; free. Later priority
Longer segments / long-form videos	Not Built	Later priority

Done Foundation

Private panel on a real domain. Upload (drag and drop) → starting image → 8-second video. Access-code sign-in. Cost per job plus running total. Prompts and photos persist across refresh. Refresh-jobs button. Sticky access code.

Active Library + Asset Vault (Phase 1)

Built and tested locally: save job outputs to our own never-expiring storage, organize with name, tags, and favorite, reuse as video inputs, sidebar navigation (Make / Library / Jobs / Settings). 49 of 49 tests pass.

Blocked on: enabling R2 storage on the Cloudflare account (a dashboard button only you can click). Until then the vault has no file storage behind it.

Key lesson: simple, restrained motion works well. Complex object physics and precise action-replication are the hard, unreliable parts. Play to the strength (simple motion plus good stills) and treat the hard ones as separate research with honest expectations.

Feature Wishlist — Everything We Have Discussed

Listing order is not a build order. Each is built and reviewed one bounded piece at a time.

#	Feature	What It Means	Status
1	Realistic short videos with text over them	The original ask; simple motion clips with overlaid text	Working
2	Quality at low cost	Never spend on auto-retries or premium models by default	Working
3	Consistent characters / avatars	The same person reusable across every video; reference images first, trained characters optional later	Partial
4	Talking head / yap videos	Short 8–60 second speaking videos with believable lip movement	Not Built
5	Timelapse / contractor transformations	Builds, cleanups, demolitions, interiors, movers — with a stable environment	Weak
6	One person doing real movements	Pushing a grocery cart, cooking at a grill, walking down the street filming herself	Mixed
7	Podcast	One or two people talking and interacting, natural feel	Not Built
8	Mimic a video / human replacement	Keep the reference action (like jump-roping) but replace the person with any character	Unreliable
9	Voices + script	Text-to-speech voiceover with scripts	Not Built
10	Motion graphics	Animated text, callouts, logos. Later priority; separate from basic captions	Not Built
11	Revisions	Fix a flaw (like a deformed arm) without redoing everything	Not Built
12	Private dashboard	See everything we have made with cost per job	Working
13	Caption options	Full word-for-word captions OR only important words and phrases highlighted. Early priority	Not Completed
14	Longer segments / long-form videos	Longer talking segments first, then complete long-form pieces. Later priority	Not Built
15	Background music / sound effects	Local mixing with FFmpeg; free	Not Built

New Idea URL to UGC Video discussed, not started

The idea: paste a product page URL — starting with Amazon — and the studio pulls the product images and description automatically, then creates a UGC-style (user generated content) video promoting the product. No manual upload of product photos or copy-pasting descriptions.

What it would do

Accept a product URL (Amazon first, other stores later).
Fetch product images, title, and description.
Build a short promotional script from the product details.
Generate a UGC-style video (casual, person-on-camera feel) promoting it.

What needs deciding before building

How to fetch Amazon data — Amazon blocks bots hard, so a scraping service or API may be needed.
Does this need voice and captions built first to feel like real UGC?
Where it sits in the build order (my lean: after captions, as its own Make workflow).

Reference tool: adskull.io — free URL-to-UGC-video tool (what this idea is modeled on).

Build Phases — One Piece At A Time

Each row is one finished, reviewable piece. Nothing starts until the previous one is reviewed (or explicitly shelved with a reason).

Phase	Piece	Why In This Order	Finish Line
0	Foundation	The base studio must exist first	Done Private panel on a real domain: photo → image → video, cost tracking, persistence
1	Library + Asset Vault	Fixes expired provider links; everything after depends on it	Active Save any job output from any browser; find it by name and tag; reuse it as a video input; it never expires
2	Captions Workflow	Early priority; feeds every video goal	Real speech video in → timed captions → correct text and timing → export a captioned copy with original audio kept
3	Versions: Re-run With Changes	The revisions goal; cheap once jobs carry lineage	From any job, change the prompt and re-run; both versions visible and linked
4	Projects + Export Variants	The repurpose engine (shape, caption mode)	Assemble clips and captions in one project; export portrait and landscape copies
5	Characters in the Library	Must precede talking head, podcast, and replacement	One saved character used across two different videos and visibly consistent
6	Voices in the Library	Must precede talking head and podcast	One saved voice used in two different videos
7	Capability Workflows, one at a time	The hard video abilities; each adds exactly one Make card	Real movements → talking head → podcast → timelapse → mimic; each with agreed test examples and an honest quality bar
8	Motion Graphics + Long-Form	Stated last priorities	Defined when selected

What we explicitly do not do: build all the screens first and fill them in. No greyed-out menu cards of future ideas. No dashboard, library, and editing suite bundled into one phase. The sidebar gains words only when the area behind them is real.

Where "URL to UGC Video" fits: not placed yet. It is a Make workflow that depends on product-data fetching, script writing, and likely voice and captions. Recommendation: treat it as its own phase after Phase 2 (captions), or fold it into Phase 7 if we want the hard video abilities first. Your call.

Tech Stack — Everything Under The Hood

Hosting & App

Cloudflare Worker named reel-lab on the seosemworks account. Serves the panel and proxies all AI calls.
Live address: vstudio.viewervelocity.com (a Worker route on the viewervelocity.com domain).
Staging: reel-lab-staging.seosemworks.workers.dev (Worker reel-lab-staging).
Panel: plain HTML and JavaScript (no framework) in app/public/index.html. Job history in browser storage.
API proxy: app/worker.js — upload, image, video, status query, and library endpoints. The AI key never leaves the server.

AI Generation (RunningHub)

RunningHub API (runninghub.ai) — one paid gateway for all generation models.
Wan 2.2 — 8-second video clips. Portrait 720×1280 and landscape 1280×720.
Nano Banana 2 (economy) — starting images and character stills.
Kling motion-control — real-movement tests (unreliable so far).
Cost is read from the provider response on every job and shown in the panel.

Storage

Cloudflare KV (reel-lab-kv) — asset metadata: names, tags, favorites, lineage, cost.
Cloudflare R2 — the actual image and video bytes in the vault. Not yet enabled — waiting on the account setting.
Saved assets get copied into our own storage the moment you save them, because provider links expire.

Local Finishing (free)

FFmpeg on this PC — captions, music and sound mixing, voice, cropping, export.
Python scripts — the original pipeline (run_test.py, custom scenes, test runners) for work not yet in the panel.
Finishing never re-spends: only the Generate step costs money.

Code, Tests & Secrets

GitHub: makbot2000/reel-lab (private). main = production, staging = staging.
Tests: Node's built-in test runner, 8 suites / 49 tests, including real-Chrome browser flows. Run before every deploy.
Access: one access code (the APP_TOKEN secret) typed once in the panel.
Secrets: the RunningHub key and access code live only as Worker secrets — never in code or chat.

Known Format Limits

Square video output is not supported by the current video model — shown honestly in the interface. Square images and square exports of existing video are fine.
Generated videos are silent — sound is added locally.
ETA shows "unavailable" until we have real completion history to estimate from. No invented countdowns.

Costs — Observed Prices

Item	Cost	Notes
Image (Nano Banana 2 economy)	~$0.027 each	Starting images and character stills
Wan 2.2 vertical video	~$0.16 per 8-second clip	Our workhorse
Kling motion-control	~$0.06 per second	Untested here for quality
Captions / music / voice mixing	Free	Done locally with FFmpeg
Text-to-speech voice	~pennies per clip	When voices are built

Cost rules: no premium models. No paid retries without asking. Every job shows its cost. The human is the cost guard.

Deploy & Workflow

How Changes Go Live

Work and commit on the staging branch.
Deploy staging and test there.
After review, merge staging into main.
Deploy production (vstudio.viewervelocity.com).

Tests run before every deploy. Production is never deployed from unreviewed work.

Deploy Commands (from the app folder)

# staging first
npx wrangler deploy --config wrangler.staging.toml

# production only after review
npx wrangler deploy

# regression tests before any deploy
node --test --test-concurrency=1 \
  tests/*.test.cjs tests/*.test.mjs

Where Everything Lives

Project folder: X:\Dropbox\hermes_projects\reel-lab\
Source of truth for goals and status: PLAN.md in that folder
Studio map and build order: INTERFACE-PLAN.md
Code: app\ (worker, panel, tests)
Generated media: stays on Dropbox only — never in git

Standing Rules

Never resubmit a job on timeout — poll for status instead.
Never auto-retry a paid job or fall back to a premium model.
Paid generation stays behind an explicit click with the cost shown.
Finishing work (captions, music, trimming) is always free and local.
Test with clearly labeled fixtures; no paid generation just to grab a test sample.

Open Decisions

Decision	Context	Status
Enable R2 storage on the Cloudflare account	The vault needs file storage. This is a dashboard setting only you can switch on. Everything else in Phase 1 is built and tested	Waiting On You
Review Phase 1 (Library) on staging	Once R2 is on: deploy staging, save real media end to end, then your review before production	Next Up
A real short video with speech	Needed to test the captions workflow. Generated clips are silent, so they cannot prove transcription or timing	Waiting On You
Where "URL to UGC Video" sits in the build order	My lean: its own phase after captions. Alternative: Phase 7 if the hard video abilities matter more	Your Call
How to fetch Amazon product data	Amazon blocks bots; likely needs a scraping service or product API. This shapes what the UGC feature can pull automatically	Your Call
Organizing style for the Library	Tags plus favorites (my lean — simple), or folders, or both	Your Call
Cost guardrail in Settings	Warn before any single job over a set amount? If yes, what amount?	Your Call

How to use this page: when a new idea or change comes up, tell me to add it here — the Feature Wishlist, Build Phases, Tech Stack, and Open Decisions tables all grow with the conversation, so nothing gets repeated or lost.

Reel Lab project overview · vstudio.viewervelocity.com · GitHub repository · source of truth: PLAN.md + INTERFACE-PLAN.md in the project folder