AI creative production market map 2026: 63 products by workflow

A video model, an app that hosts several models, an editing system and an API provider solve different parts of production. The visual map separates those buying jobs. We checked current official pages and documentation on September 28–29, 2026. Magic Hour publishes this map and appears in it. The 63 products are a purposive sample, not the whole industry.

Figure: 63 selected products across eight buyer jobs and four production stages. A company can serve several jobs; placement shows one useful starting point. Open the overview or a readable stage view: create · generate · adapt · finish. Product sources and selection limits appear below.

The eight-job market map

Each linked product name opens its official page. These are examples, not every active company. Use the table as an accessible source directory and decision checklist; the image shows how the jobs fit together.

1. Creative workspaces: coordinate several creation steps

Start here when a job moves from concept to image, video, revision and export. Magic Hour, Higgsfield, Magnific, OpenArt, Krea, Canva, Leonardo AI, InVideo, Luma and Pika are distinct examples. Magnific is the current brand for the former Freepik creative suite; this map does not count the old brand separately. Being in one category does not imply equal features, models or quality.

Compare the input you already have, available model/version for that task, revision control, the exported file and how retries are billed. For a task-level shortlist, use our best AI video generators guide.

2. Image and visual inputs: design the starting frame

Midjourney, ChatGPT, Gemini, Black Forest Labs, Recraft, Ideogram, Adobe Firefly and Photoroom illustrate still-image creation or editing. They are here because a reference frame, product asset or visual direction often determines the video that follows—not because an image model is interchangeable with a video editor. Compare editability, text rendering, subject consistency and usable export rights for the image you will animate.

3. Video generation: make the shot itself

Runway, Kling AI, Google Flow, Hailuo AI, PixVerse, Pollo AI, Vidu, Dreamina and Moonvalley offer different routes from a prompt, image or reference to a clip. Some are broader workspaces; this is their starting buyer job in this map. For a real project, hold shot length, aspect ratio, input assets, motion request and success criteria constant before comparing output and cost.

Do not compare a model announcement with a finished production workflow. The same model can appear in several apps, and one app can expose several models. Record the exact model and version, then assess editing and export in the platform actually used.

4. Avatars and talking video: put a speaker on screen

HeyGen, Synthesia, Tavus, Hedra, D-ID, Colossyan and AI Studios illustrate recorded presenters, talking characters and interactive agents. A training video and a live conversational agent are different deliverables. Check likeness permission, voice source, languages, revision control, latency where relevant, and the final channel format.

5. Product ads and social: turn an offer into variants

Creatify, Pippit, Arcads, TopView, AdCreative.ai, Fliki and Pictory illustrate different ways to turn a product, script or campaign brief into ad or social-video drafts. Evaluate your own inputs, brand rules, claim review, aspect ratios and approval workflow. Generating many variants is not evidence that an ad will perform.

For concrete examples and shot planning, see our product-video guide.

6. Face swap: transform approved existing media

Reface, DeepSwap, Remaker, Akool, FaceFusion and Swapface are identity-transformation examples rather than general footage generators. Compare photo versus video support, single- and multi-face handling, local versus hosted processing, export controls, consent, retention and rights. Our eight-platform documentation audit provides a deeper source trail and flags advertised workflows we did not verify.

7. Editing and enhancement: finish footage you already have

CapCut, VEED, Descript, OpusClip, Kapwing, Captions, Topaz Labs and Submagic cover different finishing bottlenecks: cuts, captions, speech cleanup, short-form repurposing, upscaling and export. Compare them against the footage you actually have. A better generator cannot replace precise timeline edits or a required delivery format.

8. APIs and inference: build generation into a product

fal, Runware, Replicate, Baseten, Kie.ai, WaveSpeedAI, Together AI and Modal illustrate model access and inference infrastructure. Compare the exact endpoint, model version, input constraints, asynchronous job handling, output retention and usage billing. Magic Hour also has an API for supported creation workflows, so a brand can span this and a creative-workspace job. Our text-to-video API guide explains the integration questions in more detail.

How to use this map without buying the wrong category

1. Define one finished job. Write down the input, output, duration, resolution, required edits, delivery channel and rights. 'AI video' is too broad to evaluate.

2. Choose the first bottleneck. Need new footage? Begin with scene generation or a creative workspace. Already have footage? Start with editing or face swap. Need repeatable application calls? Start with API documentation. Need a spokesperson or ad variants? Use the corresponding lane.

3. Compare a reproducible brief. Use the same permitted source assets and success criteria in each candidate. Record every attempt, failed output and exported file, then compare cost per accepted result rather than subscription price alone. Our AI video pricing guide separates platform plans from model-specific generation costs.

4. Keep a source trail. Save the official page, date, plan, model/version and settings behind a decision. A provider can change pricing, limits or available models without changing the broad category it occupies.

Three buying paths that cross the lanes

A solo creator needs a social clip. Start in a creative workspace or scene generator to make footage, then use an editor if the final job requires precise cuts, captions or music timing. Compare one actual brief and export in the channel's aspect ratio. Buying a presenter subscription first would solve a different problem unless the clip needs an on-camera speaker.

An ecommerce team needs ten ad variants. Begin with product-ad tools if catalog images, hooks and rapid variant review are the main work. Keep an editor in the workflow for human approval, claims and platform-safe final files. If a generated scene is the bottleneck, add a workspace or shot generator; judge the whole pipeline on approved variants, not raw generations.

A developer needs generation inside an app. Start from exact API references, supported job states and returned file lifetime. An infrastructure provider may expose several models; a creative workspace may expose a narrower but integrated workflow through its own API. Prototype with a real input and record failures, retries, delivery time and cost before comparing volumes.

Method and limits

The source workbook contains 208 usable competitor entity rows and 45,352 attribute-claim rows. None of those attribute claims was approved for direct publication, so we used the workbook only to discover candidates. This edition selects 62 other products plus Magic Hour across eight buyer jobs. We also checked relevant official pages outside the workbook. Selection is purposive, not random or exhaustive; category counts sum to 63 placements, not industry shares. We exclude story development, music, 3D/worlds and distribution from this video-production-focused edition. We count Magnific once under its current brand rather than also listing old Freepik branding. Retired Sora is not shown as a current product; its historical search interest does not establish current availability. See OpenAI's official Sora discontinuation notice for the dated web, app and API shutdowns.

We checked current first-party pages or documentation on September 28–29, 2026, then assigned one editorial starting job per product. Fifty-four official pages returned HTTP 200 to our direct check; nine blocked that client, so we reviewed their official indexed pages or documentation separately. A product page can establish a documented offering, not output quality, conversion, uptime, exact account access or market leadership. We did not run a standardized output test across all 63. Omission is not a negative verdict, and the map should be refreshed as products change.

The linked product names in the table form the source trail for every placement. Underlying models such as Veo, Seedance or Wan are deliberately not counted as separate apps when the buyer is choosing a workspace or editor; a model and the product that exposes it are different purchase surfaces. Check the actual model/version available inside each platform before making a decision.