Magic Hour is a browser-based creative studio and REST API that runs video, image and audio models on a single credit balance, with free no-sign-up tiers on core tools and commercial rights attached to paid plans.
Magic Hour is a browser-based creative platform that runs video, image and audio models behind one account and one credit balance. Magic Hour AI, Inc. operates it, and the Terms of Service carry a revision date of September 5, 2026. It positions itself as access to the best AI models for video and images in one place, free to try with no sign up and nothing to install.
Core tools run without an account at all, so an AI video generator, an image generator or an image editor can be tested before any signup screen appears. Accounts add saved projects and account credits; paid plans add premium third-party models, higher resolutions and commercial rights.
The consolidation argument is the reason to consider it over single-purpose apps: create, edit and finish in one studio under one account and credit balance, instead of moving assets between specialist apps with separate accounts and billing. The headline figures of 3.8M+ creators and 2K+ teams are self-reported, and no independent verification of them turned up in the channels checked this round. One structural fact sets expectations: the platform routes to other companies' models rather than running an in-house engine, so quality, duration and resolution follow whichever model a workflow calls.
Three families share one credit balance. Video covers text-to-video, image-to-video, restyling, face swap, character replace, lip sync, talking photo, animation, subtitles, upscaling and dubbing. Images cover generation, prompt editing, upscaling, background removal, generative fill, clothes changing, head and body swap, headshots, memes, GIFs and QR codes. Audio covers voice generation, cloning, voice changing and music.
The library lists 26 models in total, 14 of them in the video group, spanning video, image generation and editing in one place. Each entry is labelled by the task it accepts — text to video, image to video, reference to video, text to image or image edit — so selection starts from the input you hold.
The named models are recognisable third-party systems: Sora 2 is presented as OpenAI's model for surreal concepts and viral clips, Veo 3.1 as Google's frontier model with premium realism, and Qwen Edit as a low-cost prompt editor for object removal and inpainting, alongside Kling, Seedance, Seedream, Nano Banana, LTX, MiniMax, Wan, Flux and Z-Image variants. Because the engines come from other vendors, each model's ceilings travel with it, which is why duration and resolution are stated per model rather than per plan.
Allowances are published per tool and differ by an order of magnitude. An AI image generator session without an account gives 10 free images per day, face swap gives 5 free photo swaps plus 3 video and 3 GIF swaps daily, and the face editor allows 20 free edits per day. Video is tighter: image-to-video without an account yields 3 free generations per day at 3 seconds each, exported at 480p with a watermark.
Watermarking follows the output type rather than the plan alone: image outputs already carry no watermark, including for free users, while free video and audio outputs may include one and paid plans remove it where applicable. Free users therefore get clean stills but branded video.
The image-to-video path is three steps in the browser:
The model chosen here, not the plan, decides the clip length and resolution available.
Face swap video takes an upload or a public link, then runs in five steps: upload the video or paste a link; upload a photo of the incoming face; select which faces to swap, all or individually; render; download. Selecting faces individually is what makes multi-subject footage workable.
Video billing is frame-based, so cost is predictable with one multiplication: multiply the seconds you want by the frame rate, starting from 12 FPS for animations, 24 FPS for face swap and lip sync, and 30 FPS for text to video and image-to-video. There is no preview before rendering, but a render in progress can be tracked or cancelled and cancelled renders return all credits, so early cancellation is free.
The strongest documented use is repeatable marketing output: create campaign assets at scale, using the studio for hands-on work or the API for repeatable image and video workflows, with personalisation built as one reusable pipeline. Teams rendering one creative against many products or locales are the fit.
Dubbing is the clearest single-workflow example: upload an existing clip plus a translated audio track and the lip movements are re-synced to the new language. That turns localisation into an audio problem rather than a production one.
On the commerce side the benefit is operational: the clothes changer helps businesses, especially small ones, refresh product images and videos for online stores without a traditional photoshoot. Paired with background removal and upscaling, the realistic use is updating an existing catalogue, not shooting a new one.
Plan positioning maps onto three profiles and tracks real capability differences: Creator is aimed at social content and everyday creation at $144 billed yearly, Pro at professional work and frequent production at $300 billed yearly, and Business at teams, agencies and high-volume content at $792 billed yearly. Concurrency, upload ceiling and export quality move with the tier.
Developers are a distinct audience: the web app carries the full suite of 100+ tools and gets new features first, while the API exposes core popular tools programmatically, with features added after web app validation.
Anyone whose output must exceed a specific tool's published ceiling will hit it regardless of spend, because several caps are fixed per tool rather than per plan. Anyone whose workflow depends on adult content is excluded outright by the usage rules. And anyone needing one accountable model with stable behaviour may prefer going direct to that vendor, since a router's lineup shifts as licences change.
Supported free tools run in a browser session with no account, no sign up and nothing to install; signing up adds account credits, rewards and saved projects. Support reaches everyone by email and live chat, and paid subscribers get priority responses typically within 3 hours during Pacific business hours.
The AI video API is a REST service with first-party clients: SDKs for Python, Node.js, Go and Rust, and an async workflow where jobs are submitted, monitored and returned through polling or webhooks. That suits batch generation but means any integration needs a completion path rather than a blocking call. API output lands in the same dashboard as browser work.
Resolution is not one plan-wide pixel size. On newer endpoints you pick a named resolution: Sora 2 is 720p only, while Kling 3.0 reaches 4K. Cost follows that choice rather than the tier — a Creator-tier request for Kling 3.0 at 1080p bills at 72 credits per second, the same job at 4K at 240 credits per second. Budget by model and resolution, not by subscription name.
Three subscription tiers are published, each showing an annual rate against the monthly one.
| Plan | Annual billing | Monthly rate shown | Credits included |
|---|---|---|---|
| Creator | $144 billed yearly | $12 per month against $19 | 144,000 credits per year |
| Pro | $300 billed yearly | $25 per month against $39 | 300,000 credits per year |
| Business | $792 billed yearly | $66 per month against $99 | 840,000 credits per year |
What separates the tiers in daily use is not in that table: export quality steps from 1024px to 1472px to 4K, upload size from 2GB to 5GB to 10GB, and concurrent generations from 3 to 5 to unlimited. The monthly-billing view of the plan cards did not render in the page state captured this round, so no per-month credit grant is stated here.
The headline credit number means little alone, because the same balance stretches very differently by tool. On Creator, 144,000 annual credits cover 144,000 images in the face editor but only 1,440 in the influencer generator — a hundredfold spread inside one plan. Reading the per-tool row for your intended workflow is the only reliable way to size a subscription.
Credit packs top up without a subscription, running from $10 for 4,000 credits to $160 for 64,000 at a flat rate of $1 to 400 credits, with no watermark and no expiry. The boundary is explicit: packs carry no subscription benefits beyond watermark removal, so they do not grant commercial rights.
Unused subscription credits roll over to the next month with no expiration. Refunds are narrower: a subscription or renewal qualifies only within 7 days of payment and if fewer than 400 credits have been used since it, and only one approved refund per customer is allowed, with duplicate charges and refunds required by law excluded from that limit. Pricing excludes tax, and cancelling ends the subscription at the end of the billing cycle. Apple Pay, Google Pay, Alipay and WeChat Pay are accepted wallets and Klarna and Affirm are among the pay-later options, while PayPal is not accepted.
The comparison set splits between single-capability specialists and rival bundles. On the product-review platform that lists this tool, the alternatives shown beside it are mostly specialists: Midjourney rates 4.7 from 81 reviews on text-to-image, and RunwayML rates 4.4 from 19 reviews on video. Picking one means accepting separate accounts and billing — exactly the trade-off the consolidation pitch targets.
A consumer review platform surfaces a different set, showing what people view in the same session: Higgsfield AI at 4 from roughly 4,000 reviews, Pollo AI at 4.3 from roughly 5,000, OpenArt at 3.7 from roughly 1,000 and HeyGen at 3.9 from roughly 3,000. Those samples dwarf this product's own 13 reviews on the same platform, and ratings from different samples and platforms cannot be averaged — so the usable conclusion is not that one scores higher, but that the alternatives have far more public track record to inspect.
The decision rule follows: if one capability dominates your work, a specialist is safer; if your bottleneck is juggling several subscriptions to finish one asset, the shared balance is the differentiator.
Several caps belong to the tool rather than the plan: the headshot generator at 1344px, the clothes changer at 1472px, talking photo at 720p, and video-to-video and animation at 1472px, with video-to-video also requiring dimensions divisible by 64. Duration is set by the model — LTX 2.3 reaches 30 seconds, Seedance 12, Kling 2.5 10, Kling 3.0 15, Sora 2 60 and Veo 3.1 56 — and the Sora, Kling, Veo and Seedance families require a paid plan.
Individual tools carry their own gaps:
Third-party feedback is thin but not absent. A product-review platform shows 4.9 across 28 community reviews, and its summary of that corpus lists the recurring weak spots as consistency across variations, occasional lip-sync slips, unclear tool or mode selection and limited export options — with confusing credit usage flagged as a single complaint rather than a pattern.
A consumer review platform shows a smaller, far more polarised sample: 13 reviews averaging 3.2, split 46% five-star and 46% one-star — too few for a quality judgement, so treat it as directional only. Two vendor replies there matter more than the score. One clarifies a boundary the marketing pages blur: the 3-second limit applies to the free no-sign-up preview rather than the full tool. The other concedes that the FAQ had not been updated quickly enough after the free daily credit system changed, so free-tier terms have shifted at least once without the public pages keeping pace.
Ownership and commercial use are separate questions: you retain full ownership including copyright for everything you generate, but only paid users receive commercial usage rights and free output is limited to personal, non-commercial use. Users must be 18 or older, one account is allowed per user, and the usage rules forbid creating adult content, gore or abusive material as well as benchmarking or competitive research.
On data handling the pages checked this round do not tell one story. The help pages state that generations are private from other users and from Magic Hour, that uploads are deleted within one week, and — for the face editor specifically — that uploaded images are not used for training or shared with third parties. The Terms of Service reserve a broader right, allowing prompts, uploads and generated assets to be used to help develop and improve the company's products while excluding prompts and uploads from marketing use. None of the pages checked this round reconciles the two, so raise it with support before uploading sensitive material. The privacy policy is the oldest document of the set, revised May 31, 2023: it confirms that personal information is not sold and that data goes to servers in the United States, and does not address model training either way.
Yes, on supported core tools. They run in the browser with no account and nothing to install, with a daily allowance applied per tool. An account adds saved projects, credits and rewards, plus models that require sign-in.
Only on a paid plan. You keep ownership of what you generate, but commercial rights go to paid users only. A credit pack alone does not change that, since packs add nothing beyond watermark removal.
It depends on the model rather than the plan: Sora 2 reaches 60 seconds and Veo 3.1 reaches 56, while Kling 2.5 stops at 10 seconds. Without an account, image-to-video is limited to three-second clips at 480p.
A subscription or renewal is refundable within 7 days of payment if fewer than 400 credits have been used since it, and credit packs only if untouched. One approved refund per customer applies, apart from duplicate charges and refunds required by law.
