Reel

Make a still. Then make it move. Reel is a 24-hour rebuild of the core of higgsfield.ai: one workspace where you draft images, pick the frame you like, and animate it into a short video shot with camera moves you direct in words.

It is not a clone. It keeps the one loop people pay Higgsfield for and fixes the parts that got in the way during recon.

Live: https://higgsfiled-rebuild.vercel.app (Stripe in test mode). Video on the live site runs on Google Veo 3.1: describe a shot and get an 8 s, 720p clip with sound in about a minute. See Honest constraints for how the free setup differs.

Try it in 60 seconds

1. Start free, no account needed: one click and you're in the workspace with 30 credits.
2. Click a starter prompt (or write your own), add a camera move like #push-in, pick 16:9 or 9:16 and press Generate (20 credits). Veo 3.1 makes the clip from the prompt alone; Add start image if you want to set the first frame yourself.
3. Watch the honest progress bar: real elapsed time against the typical time. Cancel any time and the credits come back.
4. Open a finished clip and press Extend this shot to continue it from its exact last frame; Play full shot plays the linked clips as one.
5. Find everything later in Library (filter, search as you type). Top up on Credits (Stripe test card 4242 4242 4242 4242).

The product judgement story

What I studied first

Before any code I went through Higgsfield end to end and saved 42 screenshots and a map of flows, the core loop, and what was slow or confusing (recon/MAP.md). The loop users actually pay for is short: prompt → still → pick → animate → keep. Almost everything else on the site (22 top-nav items, ~30 models, community, effects, editing suites) sits around that loop, and a lot of it gets in its way.

What I built first, and why in that order

#	Milestone	Why it came here
1	A thin vertical slice: sign in → prompt → job → result in the feed	Proves the risky part first: async jobs against real model latency, stored outputs, the whole stack talking.
2	Money-safe job pipeline: atomic charge, cancel at any stage, automatic refunds on failure or timeout	Credits are the product's trust contract. One conditional SQL update charges; partial unique indexes make refunds and grants idempotent, so double webhooks or races can't double-charge or double-refund.
3	The loop itself: detail panel, Recreate, Use as reference, Animate with lineage, uploads	This is what people pay for, so it got built before anything decorative.
4	Camera moves as #tokens and style presets as visible words	The cheapest way to give direction without a camera-rig UI.
5	Library (filter, prefix search, trash with Undo) and Pricing with Stripe Checkout	Retention and the money loop, each verified end to end.
6	Design pass and phone width	Last on purpose: polish on a loop that already works.
7	Extend (longer shots from linked 5 s clips) and a generating backdrop	Asked for after using it: the free video model stops at 5 s and ~480p, so length comes from chaining clips frame-exactly, not from a bigger model.
8	Deploy and a live end-to-end test	Guest, email and Google sign-in, drafts, library, a test-card purchase and the Stripe webhook, all on the live URL. The test found two UX issues, fixed the same day (below).
9	Google Veo 3.1 behind the same adapter, and a video-only composer with text-to-video	Once a paid key was available, the free model's ~480p and 2 clips a day were the weakest part of the product. One new provider file swapped it for 8 s, 720p clips with sound, and proved the adapter was the right call.

What I fixed compared with the original

On Higgsfield (seen in recon)	In Reel
Four different generate forms across Image, Video, Cinema Studio, Effects	One composer with an Image / Video switch
"Turn to video" switches model and drops your prompt	Animate keeps the prompt, the start frame and the link back to the parent take
Just "Generating" for 25 s, no ETA; Cancel disappears once a job runs	Elapsed vs typical time, honest "taking longer than usual", Cancel in every state with refund
The default model is paywalled, and you find out after writing the prompt	Cost and what you'll have left on the Generate button; short credits are flagged before you press it
Choosing a preset hides the prompt	Presets write plain words into the prompt you can read and edit
~30 models, marketing names, prices that vary ~100×	Three choices labelled "best for" (fast draft, high quality, animate)
Pricing explained in model names ("≈ 60 Nano Banana Pro generations")	Packs explained as outcomes: "about 150 drafts, or 7 five-second videos"
Sign-up lands on the marketing page; empty tabs show onboarding	Straight into the workspace, real empty states with starter prompts
Delete with no way back	Trash with Undo
Clips are a fixed length	Extend a clip from its exact last frame and play the chain as one shot

Found in our own live test and fixed: the landing button led to a sign-in page that asked the same question again (now one click), and Generate showed nothing for ~8 s on a cold serverless start (the card now appears the instant you click).

What I cut, and why

Cut	Why
~25 of ~30 models	Model choice was the most confusing step; three curated options cover the use cases.
Subscriptions	Proration and renewals cost hours; one-off packs prove the money loop.
Camera body / lens / aperture rig	Camera moves give most of the value for a fraction of the UI.
Community, profiles, likes, contests	A network effect plus moderation, and it doesn't help a solo creator's loop.
Characters / AI influencer / Soul ID, video edit suite, Motion Control, effects	Each is its own model pipeline or fine-tuning job.
Projects and folders	A flat library with fast search is enough at this scale.
Promo banners, academy, enterprise plans	Marketing, not product.

Higgsfield's own models, preset videos and community content can't be cloned at all; Reel uses public models and its own outputs (the landing-page clip was made in Reel).

With Veo 3.1 switched on, the composer is video-only: the Image tab goes away because Veo no longer needs a still to start from. Switch video back to the free model (VIDEO_PROVIDER=huggingface) and the original loop returns: draft a still for 1 credit, open it, press Animate.

Honest constraints

• Free models. The paid model API (fal) had no balance, so Reel runs on free tiers behind one provider adapter (src/lib/providers/): images from Pollinations (about one per 45 s per IP) or Cloudflare Workers AI, video from a public Hugging Face ZeroGPU Space running Wan 2.2 image-to-video (real 5 s clips, ~1–2.5 min, about 2 clips a day on a free token). Research behind that choice: reports/Free video generation options.md. Swapping to a paid provider is an env change, and that was proven: Google Veo 3.1 (Lite, through KIE) plugs into the same adapter with VIDEO_PROVIDER=kie and makes 8 s clips with sound in about 80 s (verified at 720p for about $0.15 a clip; 1080p is the default and costs about $0.18).
• Start images. The free video model only animates images, so the free setup keeps the still → animate loop. Veo 3.1 also works from the prompt alone, so with it on, video starts from text and a start image is optional.
• Video quality and length. The free Space renders ~480p at 16 fps and 5 s per clip (Veo 3.1 on the live site: 720p, 8 s, with sound), and the composer says which before you pay. A 30 s shot is six extended clips, which takes days on the free daily allowance; with a paid video provider the same Extend flow works unchanged.
• Guests can't buy credits: anonymous credits would be lost with the browser session, so the page asks them to sign in with email first.

What's in the box

• Next.js 15 (App Router, React 19, TypeScript, Tailwind v4), Supabase (Postgres, Auth, Storage, RLS), Stripe Checkout via its REST API.
• Auth: guest (anonymous), email magic link, Google (shown only once enabled in Supabase).
• Jobs: POST /api/generations charges and submits in one transaction; reads go through GET /api/generations, which also advances active jobs with their provider (works locally without public webhooks). Long single-call providers finish in after(); a timeout fails and refunds anything stuck.
• Credits: a ledger (credit_ledger) with one charge and one refund per generation and one grant per Stripe session, enforced by unique indexes. The balance always equals the ledger sum.
• Payments: credits are granted when the buyer returns from Checkout (the server re-reads the session from Stripe) and by a signature-checked webhook, whichever arrives first, exactly once.

Plans and tracking: replica/architecture.md, replica/schema.sql, replica/features.csv (feature matrix: 27 of 28 must-haves done and verified; the 28th, curated models, runs on free tiers), replica/build-log.md.

Run it locally

npm install
cp .env.example .env.local   # fill in the values below
npm run dev                  # http://localhost:3000

1. Create a Supabase project and run supabase/migrations/0001_init.sql in the SQL editor. In Auth → Providers, enable Anonymous sign-ins (and Google, optionally).
2. Fill .env.local (names in .env.example):
   • NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, SUPABASE_SERVICE_ROLE_KEY
   • Providers: IMAGE_PROVIDER (pollinations | cloudflare | fal | mock), VIDEO_PROVIDER (huggingface | kie | fal | mock), plus HF_TOKEN (a free read token) for the free video model, or KIE_API_KEY for Veo 3.1
   • STRIPE_SECRET_KEY (test mode). STRIPE_WEBHOOK_SECRET only once deployed.
3. A prompt containing fail-test fails on purpose (mock provider), to see the refund path.

Don't run npm run build while npm run dev is running: they share .next.

Deploy (Vercel)

1. Import the repo in Vercel and add the same environment variables, plus PUBLIC_URL=https://<your-domain>.
2. In Supabase → Auth → URL configuration, set the Site URL to the Vercel domain and add https://<your-domain>/auth/callback to the redirect URLs.
3. In Stripe (test mode) → Developers → Webhooks, add https://<your-domain>/api/webhooks/stripe for checkout.session.completed and checkout.session.async_payment_succeeded (payload style Snapshot: Checkout events aren't offered as thin events), then set its signing secret as STRIPE_WEBHOOK_SECRET and redeploy.
4. Google sign-in (optional): a Google OAuth web client with redirect URI https://<project>.supabase.co/auth/v1/callback, its ID and secret in Supabase → Auth → Providers → Google. The button appears on the next visit to sign-in.

/api/generations uses maxDuration = 300 so a free-tier video can finish in the background.