Luma AI
Last reviewed
Sources
40 citations
Review status
Source-backed
Revision
v7 ยท 6,567 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
40 citations
Review status
Source-backed
Revision
v7 ยท 6,567 words
Add missing citations, update stale details, or suggest a clearer explanation.
Luma AI (operating as Luma Labs, Inc.) is a generative AI company headquartered in Palo Alto, California, that builds text-to-video, 3D, and image models, best known for its Dream Machine platform and its Ray series of video models. Founded in 2021 by Amit Jain, Alex Yu, and Alberto Taiuti, the company first gained attention for neural radiance field (NeRF) based 3D capture from smartphone video before expanding into video generation. In November 2025 Luma AI raised a $900 million Series C led by Saudi Arabia's HUMAIN at a valuation of more than $4 billion, the largest single round in its history, and partnered with HUMAIN on Project Halo, a planned 2-gigawatt AI supercluster in Saudi Arabia.[2][40]
Luma AI's flagship Ray series powers text-to-video, image-to-video, and video-to-video generation through the consumer-facing Dream Machine application and a developer API. Its September 2025 model, Ray3, was marketed as "the world's first reasoning video model and the first to generate high fidelity 16 bit HDR."[5] As of early 2026, Luma AI has raised over $1 billion in total funding, is valued at $4 billion, and serves more than 30 million registered users on its Dream Machine platform.[40]
During the first half of 2026, the company expanded beyond video generation into agentic creative software with Luma Agents, unified image understanding and generation with its Uni-1 model family, and open robotics research through the Open Physical AI Lab, while signing advertising-industry partners including Serviceplan Group and Publicis Groupe.[28][29][33][37]
Luma AI was co-founded in 2021 by Amit Jain, Alex Yu, and Alberto Taiuti. Jain and Taiuti both previously worked at Apple, where Jain led work on the Passthrough feature for the Apple Vision Pro[15] and the integration of the first LiDAR sensors for the iPhone. Taiuti also worked at Apple as an AR/VR Engineer and previously served as a Senior Autonomy Software Engineer at Skydio. Alex Yu was a researcher at UC Berkeley, where he studied neural rendering under Professor Angjoo Kanazawa and authored papers on real-time neural rendering of 3D scenes; he turned down PhD offers from Stanford University and MIT to start the company.
The company's initial product focused on NeRF-based 3D capture, allowing users to create photorealistic 3D scenes from smartphone video. The technology used deep learning networks to process 5D coordinates (location and view direction), outputting volume density and RGB radiance for each point in the scene. Users could walk around a subject while recording video with a phone, and the AI would analyze camera movement and scene geometry to reconstruct a full 3D representation.
Luma AI raised a $4.3 million seed round on October 30, 2021, led by Amplify Partners.
In March 2023, Luma AI raised $20 million in a Series A round led by Amplify Partners, with participation from NVentures (NVIDIA's venture arm) and General Catalyst. The round valued the company at $100 million post-money.
During this period, the company released its iOS app for 3D capture, which allowed anyone with an iPhone 11 or newer to create lifelike 3D models without specialized equipment such as LiDAR scanners. The app supported exporting to industry-standard formats including USDZ, glTF, and OBJ, making captures compatible with tools like Blender, Unity, and Unreal Engine.
In April 2023, Luma AI released a plugin for Unreal Engine, allowing users to import and render photorealistic NeRF captures directly inside the game engine. The plugin included tools for scene cleanup, cropping, culling, and relighting via analytical or sun lights.
In June 2023, Luma AI hired Matthew Tancik, a co-author of the original NeRF paper and co-founder of the nerfstudio framework.[1] Tancik joined to lead the applied research team. He had previously collaborated with co-founder Alex Yu on papers such as PlenOctrees and Plenoxels.
In November 2023, Luma AI launched Genie as a research preview, a text-to-3D generation tool.[1] Genie attracted tens of thousands of users on its first day.
On January 9, 2024, Luma AI raised $43 million in a Series B round led exclusively by Andreessen Horowitz (a16z).[1][3] Alongside the funding announcement, Luma released Genie 1.0 for general availability.[1]
a16z's investment thesis cited Luma's position in generative 3D, writing that "generative 3D models are on the cusp of their own breakthrough into the mainstream" following inflection points in 2D image generation in 2022 and computer vision in 2012.[1] The firm highlighted the strength of Luma's technical team, which by this point included Jiaming Song as Chief Scientist.[1] Song, who holds a Ph.D. from Stanford University under advisor Stefano Ermon, invented the DDIM (Denoising Diffusion Implicit Models) algorithm, which reduced the sampling steps required for diffusion models from roughly 1,000 iterations to 20-50.[14] DDIM was subsequently adopted in production systems including DALL-E 2, Imagen, and Stable Diffusion.[14]
On June 12, 2024, Luma AI launched Dream Machine, a text-to-video generation platform powered by its first Ray model. The tool allowed anyone to type a descriptive prompt and generate a short high-definition video clip in minutes. Dream Machine gained one million users within four days of launch[14], achieved entirely through word-of-mouth without a marketing budget.
The launch attracted wide attention on social media. Users created video versions of images generated with Midjourney, as well as moving recreations of famous artworks such as "Girl with a Pearl Earring" and popular internet memes including Doge and the distracted boyfriend meme. Director Ellenor Argyropoulos posted a Pixar-style animation of a girl in ancient Egypt that went viral.
Liam Connell, lead AI engineer at Boston Consulting Group, noted at the time that Dream Machine's video quality and realism stood out from competing models, with fewer obvious inconsistencies than alternatives.
On November 25, 2024, Luma AI introduced Photon, a text-to-image foundation model.[9] The Photon family included two variants: Photon (full quality) and Photon Flash (optimized for speed). Luma claimed the models were 800 percent faster and cheaper than comparable image generation services.[9] Photon generated ultra-high-quality 1080p images at a cost of 1.5 cents per image, while Photon Flash cost 0.4 cents per image.
Photon introduced several capabilities new to image generation at the time, including a large context window for visual generative models and consistent character generation from a single reference image (in beta). Alongside Photon, Luma launched subscription tiers for Dream Machine: Lite ($9.99/month), Plus ($29.99/month), Professional ($99.99/month), and Enterprise.[9]
On December 6, 2024, Luma AI raised $90 million in an initial Series C round with participation from new investors Amazon and AMD, alongside existing backers a16z, Amplify Partners, and Matrix Partners.[16][17] This brought total funding to approximately $157 million.[16] Other new participants in the round included Factorial Funds, LDV Capital, and Hanwha.[17]
On January 15, 2025, Luma AI released Ray2, its second-generation video model.[4] Ray2 was built on a new multi-modal architecture with 10x the compute power of Ray1.[4] The model was trained directly on video sequences, teaching it to understand motion as continuous flow rather than discrete snapshots.
Ray2 supported video generation at 540p, 720p, and 1080p resolution with 4K upscaling, producing clips of 5 to 10 seconds that could be extended to approximately 30 seconds. The model introduced structured prompting, seamless looping, camera movement controls (crane, tracking, dolly), keyframe control, style references, and color grading. Ray2 was available exclusively to paid Dream Machine subscribers, via the iOS app, and through the developer API.
Luma AI's Ray2 model was also made available through Amazon Bedrock, making AWS the first cloud provider to offer Luma's video models to enterprise customers.[10] Under the partnership, AWS became the preferred compute partner for Luma AI, and Luma began training its foundation models on Amazon SageMaker HyperPod infrastructure.[18]
In March 2025, Luma released Ray2 Flash, a variant of Ray2 that the company described as 3x faster and 3x cheaper, bringing Ray2's text-to-video, image-to-video, audio, and control capabilities to all paid subscriber tiers.[19] The same month, Luma's research team published Inductive Moment Matching (IMM), a paper proposing a new class of generative models that sample in one or a few steps (see Research below).[20]
On May 15, 2025, Luma announced a strategic partnership with HUMAIN, the Saudi AI company backed by the kingdom's Public Investment Fund, spanning compute, go-to-market, and investment, and began deploying Ray2 on HUMAIN infrastructure.[21] In June 2025, the company released Modify Video, a video-to-video tool that restyles or retextures real footage (changing wardrobe, props, environments, or overall aesthetic) while preserving the original motion and camera dynamics, without requiring motion capture or green screens.[22] Modify with Instructions followed on August 12, 2025, adding natural-language video editing with object removal and swapping, virtual sets, character refinements, and restyling, with adjustable Adhere, Flex, and Reimagine strength levels.[24]
In July 2025, Luma opened Dream Lab LA, a Los Angeles studio intended as a creative incubator and training ground where filmmakers and studios learn to integrate AI into production workflows. Veteran creative executive Verena Puhm was named head of the studio and Jon Finger was hired as creative workflow executive.[23]
On November 19, 2025, at the U.S.-Saudi Investment Forum during Crown Prince Mohammed bin Salman's visit to Washington, D.C., Luma AI announced a $900 million Series C investment.[2] The round was led by HUMAIN, a company owned by Saudi Arabia's Public Investment Fund (PIF) that delivers global full-stack AI solutions, with participation from AMD Ventures and existing investors Andreessen Horowitz, Amplify Partners, and Matrix Partners.[2] The round valued Luma AI at over $4 billion, making it a unicorn.[40] Amit Jain, Luma's co-founder and CEO, said: "HUMAIN is the perfect partner for this next stage in Luma AI's explosive trajectory."[2] HUMAIN CEO Tareq Amin framed the deal as a value-chain bet, saying: "This investment underscores an important point in HUMAIN's philosophy: we are not only funding the next wave of AI, we're building the full value chain."[2]
As part of the deal, Luma AI became a customer of HUMAIN's Project Halo, a planned 2-gigawatt AI supercluster in Saudi Arabia that would be one of the world's largest compute infrastructure buildouts.[2] Luma AI stated it was developing technology "capable of efficiently training on peta-scale multimodal data," handling "1,000-10,000 times more information than the current frontier Large Language Models (LLMs)."[2] The partnership also included HUMAIN Create, an initiative to build sovereign AI models trained on Arabic and regional data, with Luma working on what it described as the world's first Arabic video model.[2]
On September 18, 2025, Luma AI released Ray3, which it described as "the world's first reasoning video model and the first to generate high fidelity 16 bit HDR."[5] Ray3 was more than twice the size of Ray2 and introduced several technical firsts. It generated physically accurate videos with improved fidelity, instruction following, and temporal coherence. Ray3 was the first video generation model capable of producing High Dynamic Range (HDR) video in professional ACES2065-1 EXR format across 10-, 12-, and 16-bit pipelines, the color encoding standard used in film and television finishing.[5] Luma said Ray3 "can actually think through what a user is asking, plan complex scenes, and judge whether its own output makes sense."[5]
Ray3 included reasoning-driven generation, improved video-to-video workflows with character reference and keyframe support, a Draft Mode for rapid iteration, and an HDR pipeline.[5] Adobe integrated Ray3 into Adobe Firefly and Firefly Boards, making Luma AI's first third-party distribution partner.[8] Users on paid Firefly or Creative Cloud Pro plans received unlimited Ray3 generations through October 1, 2025.[8] Content generated with Ray3 in Firefly could be synced to Creative Cloud accounts for editing in Premiere Pro and other Adobe applications.
In October 2025, Luma published a Ray3 evaluation report stating that third-party evaluations placed the model ahead of competing video generators on most measured dimensions, including motion artifacts (where Ray3 exhibited artifacts less often and less severely than any other tested model) and temporal consistency.[25]
On December 2, 2025, Luma AI opened its first international office in London, United Kingdom. The company appointed Jason Day, a former executive at Monks and WPP, as Head of EMEA to lead international expansion across creative, advertising, gaming, and entertainment sectors.[7] Luma AI planned to hire approximately 200 employees in London by 2026, representing about 40% of its workforce.[7][13] The company chose London in part because of proximity to institutions like DeepMind and the strength of the UK's AI research community.[13]
In December 2025, Luma AI released Ray3 Modify, a model designed for hybrid workflows combining real-world footage with AI generation.[11] Ray3 Modify introduced four capabilities: Keyframe Control, Character Reference, enhanced Modify Video, and Start-and-End Frame control.[11] The model allowed creative teams to provide a start frame and end frame, then generate transitional footage between them. It could retain an actor's original motion, timing, eye line, and emotional delivery while transforming the surrounding scene, costumes, or environment.[11] The release came on December 18, 2025; TechCrunch noted that Ray3 Modify built on the Modify Video capabilities Luma had first introduced in June 2025.[26]
On January 26, 2026, Luma AI launched Ray3.14, an update that delivered native 1080p generation, 4x faster generation speeds, and per-second pricing that was 3x cheaper than Ray3.[6] The model applied the reasoning engine from Ray3 more powerfully to animation and professional video workflows, producing higher detail adherence and the strongest temporal stability Luma had achieved. Modify Video support was extended to 18-second clips.[6]
On February 11, 2026, Luma AI announced plans to open a Riyadh, Saudi Arabia office to accelerate the HUMAIN Create initiative.[12] The company also named Publicis Groupe Middle East as its preferred AI creative partner across the MENA region.[12] Through the partnership, Publicis would integrate Luma AI's generative video and multimodal AI technologies into creative production workflows for brands across the Middle East and North Africa.
In early 2026, Luma mounted a sustained push into the advertising industry. On February 2, 2026, the company announced The Luma Dream Brief, a global competition offering $1 million to any team that wins a Gold Lion at the June 2026 Cannes Lions festival with work created using Luma AI; submissions ran through March 22, 2026.[32] On February 19, 2026, Serviceplan Group, Europe's largest independent agency network with more than 6,500 employees across 43 locations, agreed to deploy Luma's creative AI across its global operations, with Luma serving as its AI technology partner.[33]
On March 5, 2026, Luma launched Luma Agents, AI agents that handle creative workflows end to end across text, image, video, and audio while maintaining context across assets, collaborators, and iterations. The agents orchestrate Luma's own models alongside third-party systems including Google's Veo 3 and Nano Banana Pro, ByteDance's Seedream, and ElevenLabs voice models. Early customers included Publicis Groupe, Serviceplan, Adidas, Mazda, and HUMAIN, and TechCrunch reported that one customer converted a $15 million annual advertising campaign into localized ads for multiple countries in 40 hours for under $20,000.[28]
Later in March 2026, Luma released Uni-1, its first Unified Intelligence model and its first model to combine image understanding and image generation in a single architecture.[29] Uni-1 is a decoder-only autoregressive transformer in which text and image tokens share a single sequence, allowing the model to perform structured reasoning before and during image synthesis rather than denoising images from noise as diffusion models do.[29] According to Luma, Uni-1 ranked first on human-preference Elo ratings for overall quality, style and editing, and reference-based generation, led the RISEBench reasoning benchmark ahead of Google's Nano Banana 2 and OpenAI's GPT Image 1.5, and cost roughly 10 to 30 percent less than comparable models at high resolution.[29][30]
The advertising campaign culminated in April 2026: Luma submitted 21 finalist entries from the Dream Brief competition to Cannes Lions after the initiative drew roughly 400 submissions in under eight weeks[34]; Johannesburg agency Boundless used Luma Agents to deliver Mazda's first AI-produced commercial, a heritage film supporting the MX-5's South African relaunch that went from concept to approved campaign in under two weeks[35]; and Luma joined Jon Erwin's Wonder Project to launch Innovative Dreams, a production services company, R&D lab, and post-production and visual effects firm backed by Amazon Web Services that combines performance capture, virtual production, and generative AI in what the companies call Realtime Hybrid Filmmaking. Innovative Dreams' first project, The Old Stories: Moses, a three-part companion special to House of David starring Ben Kingsley and O-T Fagbenle, was produced with Amazon MGM Studios for Wonder's subscription service on Prime Video.[36]
On May 5, 2026, Luma released the Uni-1.1 API, exposing Uni-1's image generation and natural-language editing through a REST interface that supports up to nine reference images per request, with launch partners including Envato, Comfy, Runware, Flora, Krea, Magnific, Fal, and LovArt.[31]
On June 1, 2026, the company announced the Open Physical AI Lab, an open science initiative aimed at solving generalization in physical AI. The lab invites outside robotics teams and researchers to test whether Luma's world models trained on video, image, and 3D data can help machines generalize beyond narrow tasks, and CEO Amit Jain framed the open approach as a way to prevent physical AI infrastructure from being controlled by one or a few companies.[37]
On June 9, 2026, Luma introduced Ray3.2, an update to the Ray3 family developed with input from entertainment, advertising, and gaming creatives. Ray3.2 added frame-level control with up to 16 keyframes per clip, native 1080p generation at up to 20 seconds, HDR generation with 16-bit EXR export, performance tracking for up to 8 faces simultaneously, and integrated reframing tools, and shipped alongside a developer API.[38]
Dream Machine is Luma AI's consumer-facing web platform and iOS application for AI video generation. Launched in June 2024, it allows users to generate video from text prompts, images, or existing video clips. The platform supports text-to-video, image-to-video, video-to-video, and video extension workflows.
| Feature | Description |
|---|---|
| Text-to-video | Generate video clips from natural language descriptions |
| Image-to-video | Animate a static image into a video sequence |
| Video-to-video | Modify existing footage with AI-driven transformations |
| Start/End frame | Provide beginning and ending frames; the model generates transitional footage |
| Video extension | Extend an existing clip while maintaining motion coherence |
| Camera controls | Specify camera movements such as crane, tracking, dolly, and pan |
| Keyframe control | Set keyframes to guide the model's generation across a clip |
| Character reference | Provide a reference image to maintain character consistency |
| Style reference | Apply a visual style from a reference image |
| HDR output | Generate video in High Dynamic Range using ACES2065-1 EXR format (Ray3+) |
| Looping | Create seamlessly looping video clips |
| 4K upscaling | Upscale generated video from native resolution to 4K |
As of early 2026, Dream Machine has over 30 million registered users and generated approximately $21.2 million in revenue during 2025. Sacra, a private-markets research firm, had earlier estimated that Luma reached $8 million in annualized revenue as of December 2024, shortly after Dream Machine subscriptions launched.[16]
| Plan | Monthly Price | Features |
|---|---|---|
| Free | $0 | Limited generations per day, watermarked output |
| Lite | $9.99 | Increased credits, commercial use rights |
| Plus | $29.99 | More credits, faster generation |
| Unlimited | $94.99 | Unlimited generations, priority access |
| Enterprise | Custom | Custom volume, dedicated support, SLAs |
The Ray models are Luma AI's core video generation models, powering the Dream Machine platform and API.
| Model | Release Date | Key Capabilities |
|---|---|---|
| Ray 1 | June 2024 | First-generation video model; text-to-video and image-to-video at launch |
| Ray 1.6 | 2024 | Improved visual quality and coherence over Ray 1 |
| Ray 2 | January 15, 2025 | New multi-modal architecture with 10x compute; 540p/720p/1080p with 4K upscale; extended clips up to 30 seconds; camera controls, looping, style reference |
| Ray 2 Flash | March 2025 | Speed-optimized Ray2 variant; 3x faster and 3x cheaper; available to all paid tiers [19] |
| Ray 3 | September 18, 2025 | First reasoning video model and first to generate high-fidelity 16-bit HDR; 2x+ size of Ray2; native 1080p; HDR in ACES2065-1 EXR; character reference, keyframes, Draft Mode [5] |
| Ray 3 Modify | December 2025 | Hybrid AI/live-action workflows; start/end frame control; actor performance preservation; scene transformation |
| Ray 3.14 | January 26, 2026 | Native 1080p; 4x faster; 3x cheaper per-second pricing; strongest temporal stability; 18s Modify Video support |
| Ray 3.2 | June 9, 2026 | Frame-level control with up to 16 keyframes; 1080p at up to 20 seconds; 16-bit EXR export; performance tracking for up to 8 faces; integrated reframing [38] |
Ray2's transformer architecture processes text prompts by decomposing descriptions into scene components, motion dynamics, lighting conditions, and camera behavior. It then generates video by predicting the temporal evolution of the scene frame by frame, maintaining consistency with physical properties learned during training. The model was trained directly on video sequences rather than still images, which Luma credits with its ability to produce natural, continuous motion.
Genie is Luma AI's text-to-3D generation tool. It converts natural language prompts into 3D models in under 10 seconds. The system follows a two-stage process: first generating four low-resolution previews in approximately 10 seconds, then allowing users to refine their chosen variant into a high-quality model with PBR (Physically Based Rendering) textures and optimized mesh topology.
Genie generates quad meshes and materials at configurable polygon counts, exporting in standard formats compatible with professional 3D pipelines. It is accessible via web, iOS, and Discord. The tool targets developers, 3D artists, game designers, and VR/AR creators who need rapid 3D asset prototyping.
Genie was first released as a research preview in November 2023 and reached general availability with version 1.0 in January 2024.[1]
Photon is Luma AI's text-to-image foundation model, released on November 25, 2024.[9] The model family includes Photon (full quality at 1.5 cents per 2MP 1080p image) and Photon Flash (speed-optimized at 0.4 cents per image). Luma claimed Photon was 800% faster and cheaper than comparable models at launch.[9]
Photon supports a large context window for visual generative models and can generate consistent characters from a single input reference image. The model is available through the Dream Machine web interface and via API.
Uni-1, released in March 2026, is the first model in Luma's Unified Intelligence family and marked an architectural departure for the company's image generation: instead of stitching a generator to a separate reasoning system at inference time, Uni-1 is a single decoder-only autoregressive transformer that processes interleaved text and image tokens, so instruction understanding, reasoning, and pixel generation run in one forward pass.[29] The model supports text-to-image generation, natural-language image editing, and reference-based generation with multiple reference inputs. At 2K resolution, Luma priced Uni-1 API generation at approximately $0.09 per text-to-image output, below comparable Google and OpenAI image models at the time.[29][30] The Uni-1.1 API, released May 5, 2026, added Build and Scale pricing tiers, output in standard aspect ratios, and generation times of roughly 31 seconds per image.[31]
Luma Agents, launched March 5, 2026, sit above the model layer: they plan and execute multi-step creative briefs across Dream Machine, Uni-1, the Ray models, and selected third-party models, and can critique and revise their own outputs.[28]
Luma AI's original product was a 3D capture tool built on neural radiance field technology. The iOS app allows users to capture 3D scenes by walking around a subject while recording video on an iPhone (model 11 or newer). The AI processes the video to reconstruct 3D geometry, lighting, and material properties.
Key capabilities of the 3D capture system include:
The Unreal Engine plugin (released April 2023) allows photorealistic NeRF and Gaussian Splatting captures to be imported, cropped, relit, and combined with traditional Unreal Engine VFX and real-time graphics.
Luma AI provides a developer API for integrating video and image generation into third-party applications. API credits are purchased separately from Dream Machine subscription credits.
| API Plan | Monthly Credits | Features |
|---|---|---|
| Free | Limited | Basic text-to-video |
| Creator | $10 free credits | Unlimited task speed, email support, file-to-URL conversion |
| Pro | $60 free credits | All task types, live support, custom invoicing |
| Enterprise | Custom | Dedicated infrastructure, SLAs, volume pricing |
The API supports text-to-video, image-to-video, video extension, end frame specification, and watermark removal. The Ray2 model is also available through Amazon Bedrock, allowing AWS customers to access Luma's models through a managed API.[10]
| Name | Role | Background |
|---|---|---|
| Amit Jain | Co-founder, CEO | Former Apple Systems and ML Engineer; led Passthrough for Apple Vision Pro[15] and LiDAR integration for iPhone; B.S. in Mathematics and Computer Science from Missouri Valley College |
| Alex Yu | Co-founder, CTO | UC Berkeley researcher; studied NeRF neural rendering under Professor Angjoo Kanazawa; turned down PhD offers from Stanford and MIT to start Luma; authored papers on real-time neural rendering and single-image 3D generation [1] |
| Alberto Taiuti | Co-founder, CTO | Former Apple AR/VR Engineer (2 years); former Senior Autonomy Software Engineer at Skydio; Abertay University alumnus |
| Jiaming Song | Chief Scientist | Ph.D. in Computer Science from Stanford (advisor: Stefano Ermon); inventor of DDIM[14]; former Principal Research Scientist at NVIDIA Deep Imagination group; joined Luma AI in July 2023 |
| Matthew Tancik | Applied Research Lead | Co-author of the original NeRF paper[1]; co-founder of nerfstudio; joined from UC Berkeley in June 2023 |
| Jason Day | Head of EMEA | Former executive at Monks and WPP; appointed December 2025 to lead London office and international expansion [7] |
As of February 2026, Luma AI has approximately 246 employees at its Palo Alto headquarters, with plans to grow to over 400 employees globally through London and Riyadh office expansions.
Luma AI has raised over $1.07 billion across six funding rounds.
| Round | Date | Amount | Lead Investor(s) | Notable Participants | Post-money Valuation |
|---|---|---|---|---|---|
| Seed | October 30, 2021 | $4.3M | Amplify Partners | - | - |
| Series A | March 20, 2023 | $20M | Amplify Partners | NVentures, General Catalyst | $100M |
| Series B | January 9, 2024 | $43M[3] | Andreessen Horowitz (a16z) | - | - |
| Series C (initial) | December 6, 2024 | $90M[16][17] | - | Amazon, AMD, a16z, Amplify Partners, Matrix Partners | - |
| Series C (extended) | November 19, 2025 | $900M[2] | HUMAIN (PIF) | AMD Ventures, a16z, Amplify Partners, Matrix Partners | $4B+ |
The company has attracted a total of 17 investors, including Amazon, General Catalyst, NVentures, Amplify Partners, a16z, AMD Ventures, and HUMAIN.
Luma AI and AWS formed a strategic partnership in late 2024.[18] AWS became the preferred compute partner for Luma AI, and Luma's Ray2 model was integrated into Amazon Bedrock.[10] Luma trains its foundation models on Amazon SageMaker HyperPod and plans to bring models to AWS Trainium and Inferentia chips for lower-cost training and inference.[18]
The partnership was unveiled at AWS re:Invent on December 4, 2024, where Luma previewed Ray2 ahead of its January 2025 release.[18] In April 2026, AWS also backed Innovative Dreams, the hybrid AI-filmmaking production services company that Luma launched with Wonder Project.[36]
Adobe integrated Luma AI's Ray3 model into Adobe Firefly and Firefly Boards in September 2025, making Adobe the first partner to launch Ray3 outside of Luma AI's Dream Machine platform.[8] This was Luma's first third-party distribution partnership. Ray2 was also subsequently added to Firefly in April 2025.[39] The integration allows users to generate video within Adobe's creative tools and sync content to Premiere Pro for editing.
Through the $900 million Series C investment in November 2025, Luma AI became a customer of HUMAIN's Project Halo, a 2-gigawatt AI supercluster in Saudi Arabia.[2] The partnership includes development of Arabic-native foundation models under the HUMAIN Create initiative and Luma AI's plans to open a Riyadh office.[12]
The relationship predates the investment: Luma and HUMAIN first announced a strategic partnership on May 15, 2025, covering compute capacity, go-to-market collaboration, and investment, with Luma deploying Ray2 on HUMAIN infrastructure to serve creative customers.[21]
In February 2026, Publicis Groupe Middle East became Luma AI's preferred AI creative partner across the MENA region, integrating Luma's generative video technologies into advertising and brand production workflows.[12]
Luma AI operates in the AI video generation market alongside several other companies.
| Company | Product | Strengths | Typical Use Cases |
|---|---|---|---|
| Luma AI | Dream Machine (Ray models) | Cinematic realism, natural motion, physics simulation, fast generation, HDR support | Motion-intensive content, cinematic storytelling, product videos |
| Runway | Gen-3 / Gen-4 | Motion Brush, Director Mode, Camera Control, fine-grained editing | Brand advertising, 4K master files, precise frame control |
| OpenAI | Sora | ChatGPT integration, synchronized dialogue and sound, long-form generation | Realistic scenes, imaginative storytelling, dialogue-driven content |
| Pika Labs | Pika | Fast rendering, low cost, stylistic expression, social-first | Social media content, meme animations, dance clips |
| Google DeepMind | Veo | High physical realism, long-form video, integration with Google tools | Professional video production |
| Kuaishou | Kling | Strong motion quality, competitive pricing | General video generation |
Luma AI holds an estimated 15-20% market share in AI video generation as of late 2025, positioning it between Pika's experimental tools and Runway's professional offerings. Dream Machine's lower entry price ($9.99 for the Lite plan) compared to competitors like Runway ($15 for Standard) has contributed to rapid user growth.
In image generation, Luma's Uni-1 competes directly with Google's Nano Banana family and OpenAI's GPT Image models, with Luma positioning reasoning quality and lower per-image cost as its differentiators.[29][30]
Luma AI's tools are used across several industries:
Luma AI's video models are built on a multi-modal transformer architecture. When processing a text prompt, the model decomposes the description into separate components: scene layout, motion dynamics, lighting conditions, and camera behavior. The model then generates video by predicting the temporal evolution of the scene, maintaining frame-to-frame consistency with physical properties learned during training.
Key technical characteristics of the Ray model family include:
For 3D capture, Luma AI uses both NeRF (neural radiance field) representations and Gaussian Splatting. NeRF represents scenes as continuous volumetric functions, while Gaussian Splatting uses explicit 3D Gaussian primitives for faster rendering. Both methods can be exported for use in game engines and 3D software.
Luma frames its long-term agenda as building multimodal world models rather than larger language models. Jain has described the goal as a "universal imagination engine" trained on multimodal data approaching 1.2 quintillion tokens across tens of petabytes, arguing that video, audio, and image data carry far more information about the physical world than text.[15] The company says its training systems are designed for peta-scale multimodal data, 1,000 to 10,000 times more information than frontier LLMs process.[2] With Uni-1, Luma also departed from diffusion-based image generation in favor of a decoder-only autoregressive transformer that reasons over interleaved text and image tokens before rendering pixels.[29]
Luma's research team, led by Chief Scientist Jiaming Song, publishes work on efficient generative modeling. In March 2025, Linqi Zhou, Stefano Ermon, and Song released Inductive Moment Matching (IMM), which proposes a class of generative models trained in a single stage, without pre-trained initialization, that sample in one or a few steps. Unlike consistency models, IMM guarantees distribution-level convergence; the method reached 1.99 FID on ImageNet-256x256 using 8 inference steps and a state-of-the-art two-step FID of 1.98 on CIFAR-10.[20]
In November 2025, Zhou, Mathias Parger, Ayaan Haque, and Song released Terminal Velocity Matching (TVM), a generalization of flow matching that models transitions between any two diffusion timesteps and regularizes behavior at the terminal time, providing an upper bound on the 2-Wasserstein distance between model and data distributions. TVM achieved 3.29 FID on ImageNet-256x256 with a single network evaluation and 1.99 FID with four evaluations, matching diffusion-quality output at a fraction of the inference cost; the paper was published as a conference paper at ICLR 2026 and the ImageNet code was open-sourced.[27]
The Open Physical AI Lab, announced June 1, 2026, extends this research agenda from generative media toward robotics, applying Luma's world models and evaluation frameworks to the generalization problem in physical AI in collaboration with outside researchers, academia, and hardware partners.[37]