Last week, Luma AI launched Ray3, a generated video model that is positioned as a creative partner who can “infer” and deliver via prompts on the ACES2065-1 EXR in 10, 12, and 16-bit formats. The model is live within Luma's Dream Machine and Adobe's Firefly apps, and is targeted at filmmakers, advertisers and game developers who need production-grade output. Let's dive into Luma ai ray3.
Ray3 moves beyond the quick interpretation of one shot by evaluating its own output during generation and improving the results. Luma frames this as a shift from slot machine style randomness to instructions and temporal consistency that keeps characters, movements and physics consistent over time.
New features in Luma ai ray3
Ray3 introduces a multimodal inference system designed for creative tasks. The model can plan the scene and follow annotated instructions to determine whether the results match the user's intent. This helps maintain the continuity and physical validity of the story throughout the shot.
At the format level, the Ray3 is the first model to generate video as true HDR in the professional ACES2065-1 EXR, supporting 10, 12, and 16-bit pipelines commonly used in high-end finishes. Luma adds native 1080p generation and neural upscalars to 4K to meet broadcast and studio requirements.
Preparing Native HDR and Pipelines
Ray3 writes directly to the ACES2065-1 EXR with a maximum accuracy of 16 bits. Colorists access deeper highlights and shadow details, plus robust round trips through grading and VFX. Luma also points out that SDR materials that are not camera captured or generated by AI can be converted to generated HDR for the expansion stage.
Draft mode and creative control
The new draft mode accelerates your ideas by creating test videos up to 10 times faster and saving your identity, motion and configuration when promoting your draft to the final high-quality render. Ray3 also brings up Upgraded Creative Tools: Image-to-Video animation, timing and transition keyframes, stretch, stretch, and seamless repetitive loops.
CNET reports practical timing during testing. Draft shots can be rendered in about 20 seconds, and high-grade to high fidelity output takes about 2-5 minutes.
Adobe Firefly Integration Luma AI ray3
Adobe is the first partner to ship Ray3 outside of Luma Ai's Dream Machine. Ray3 appears on Firefly's video modules and Firefly boards, syncing content with Creative Cloud to finish with Premiere Pro. Adobe says all firefly generations are eligible for content, and assets created with Firefly are not used to train generative models. As a launch, paid Firefly and Creative Cloud Pro customers will win unlimited Ray 3 generations for the first 14 days.
Early Partners and Target Users
Beyond independent creators and studios, launch partners include agencies such as Monks, Galeria and Strawberry Frog, along with Human Create in Japan's Dentsu Digital and Mena region, showing a push to scaled ads and branded content.
Industry skepticism and filmmaker adaptation
Despite technological advances, it is understandable that many industry experts remain skeptical of generative AI videos. Concerns range from originality and copyright questions to the potential erosion of traditional crafts. As we continue to report on AI innovations already impacting craft, members of CINED viewers have also expressed reservations about the authenticity and cinematic value of AI-driven images. However, platforms like Adobe Firefly integrate models such as the Ray3 into everyday creative tools, making it increasingly clear that professional filmmakers need to get used to these systems. Understanding how to leverage or critically evaluate AI videos is becoming an important part of staying relevant in a rapidly changing production environment.
availability Luma AI ray3
Luma AI Ray3 is available within Luma's Dream Machine Platform and Adobe Firefly. Adobe points out that for the first two weeks, starting September 18, 2025, Ray3 access will be limited to Fireflies and Dream Machines.
The combination of Ray3's inference, ACES EXR HDR, and Adobe's integration point to an era of more controllable pipeline-enabled AI video.
Will it be a go-to pre-introduction and finishing tool on a real-world set, or it remains a quick concept engine. How do you plan to test it?
Source link
