What is NVIDIA Relighting?
NVIDIA Relighting is a specialized video-processing model for changing the apparent illumination of subjects and scenes after footage has been recorded. Instead of generating a new video from a text prompt, it takes video and applies virtual lighting effects while retaining the original scene context.
NVIDIA identifies the model as Relighting, with the model ID ai4m-relighting. It belongs to NVIDIA AI for Media and is delivered as a containerized NVIDIA NIM service rather than as a conventional conversational or token-based model API.
The practical result is similar to placing a subject under a different virtual studio light. A production team can use an HDR environment map to define the lighting environment, adjust how strongly the foreground and background are affected, and composite the result against the original video, a custom image, or an HDR-based background.
How the relighting pipeline works
Relighting combines several video-processing stages. The system analyzes incoming frames, separates foreground subjects from the background, projects HDR lighting onto the relevant image regions, and composites the processed elements into the output video. HDR, or high dynamic range, environment maps contain lighting information from a surrounding scene and provide the model with a richer description of the desired illumination than a simple color filter.
Users can select from five built-in HDR lighting presets or provide custom HDR environment map files. The available controls include foreground gain, background gain, blur strength, specular highlights, output resolution, bitrate, output encoding, and IDR intervals. These settings make the service more suitable for production pipelines than a fixed visual filter, although the quality and behavior of the final result still depend on the input footage, selected lighting environment, and GPU configuration.
Inputs, outputs, and supported controls
The primary input is video. Optional inputs include background imagery and HDR environment maps. The primary output is relit video, so the model should be evaluated as a video-effects system rather than as a general multimodal assistant.
| Category | Verified information |
|---|---|
| Input | Video, with optional background images and HDR environment maps |
| Output | Relit video |
| Lighting sources | Five built-in HDR presets or custom HDR environment maps |
| Background options | Source video, custom image, or HDR projection |
| Modes | Streaming and transactional inference |
| Adjustable settings | Foreground and background gain, blur, specular highlights, bitrate, resolution, encoding, and IDR intervals |
NVIDIA's documentation does not specify a conventional context window, maximum token output, or text-output limit for this model. Those concepts are not the right way to describe a video transformation service. Practical limits instead depend on supported video formats, resolution, encoding settings, GPU memory, and the deployment configuration documented for the NIM container.
Deployment and performance
NVIDIA provides Relighting as a containerized NIM that uses NVIDIA software components including Triton Inference Server, TensorRT, CUDA, and DeepStream. Streaming mode is intended for progressive playback or live-style processing, while transactional mode is intended for inputs that are not streamable and can be handled as discrete processing jobs.
The documented hardware support covers NVIDIA Turing, Ampere, Ada, Hopper, and Blackwell GPUs with suitable Tensor Core and video encode/decode capabilities. This requirement is central to the product's positioning: Relighting is designed for NVIDIA GPU infrastructure, not for ordinary CPU-only deployment or a typical consumer web application.
In NVIDIA's published performance results, single-stream throughput ranges from approximately 21 frames per second at 1080p on an L4 to approximately 59 frames per second at 1080p on an RTX 5090 under the documented test conditions. These are provider-reported results, not a guarantee for every installation. Actual throughput can vary with the GPU, resolution, encoding options, concurrency, input characteristics, and other services sharing the system.
What Relighting can and cannot do
Relighting's strength is narrow specialization. It can apply configurable virtual illumination, preserve and process foreground subjects, composite backgrounds, and support streaming-oriented video workflows. It is suited to a pipeline where the input and output are both visual media rather than text.
It does not provide text generation, conversational reasoning, coding assistance, embeddings, speech synthesis, transcription, image generation, or general-purpose tool calling. Its documented purpose is not to understand a user's question or produce a creative video from a written prompt. The system performs a defined transformation on video supplied by a larger production workflow.
The database classification marks reasoning and coding capability at minimal levels, but these are editorial catalog scores rather than NVIDIA-published benchmarks. They should not be interpreted as evidence that Relighting is a reasoning or programming model. Similarly, its high speed score reflects its reported GPU-oriented video throughput, not a universal latency guarantee.
Main strengths
- Purpose-built video relighting: It addresses a specific production problem that general-purpose language or image models do not directly solve.
- Configurable lighting: Built-in and custom HDR environment maps provide more control than a simple brightness, tint, or color-grade adjustment.
- Foreground and background processing: Separate controls and compositing options help integrate lighting changes into virtual production and media workflows.
- Streaming support: The streaming mode is relevant to live or progressive playback scenarios, while transactional mode supports non-streamable inputs.
- NVIDIA acceleration: TensorRT, CUDA, Triton, DeepStream, Tensor Cores, and hardware video encode/decode are integrated into the intended deployment environment.
- Production-oriented controls: Resolution, bitrate, encoding, blur, gain, specular highlights, and IDR settings expose operational controls that are useful in a real video pipeline.
Limitations and pricing
The largest limitation is infrastructure dependency. Running the model requires compatible NVIDIA GPU hardware, NVIDIA container tooling, an NGC API key, and appropriate video encode/decode support. That makes it less convenient than a hosted web editor or a general cloud media API, especially for small teams without an existing NVIDIA deployment stack.
Relighting is also narrower than generative video systems. It changes lighting and composites video, but the supplied research does not support claims that it can create arbitrary scenes, synthesize dialogue, generate music, or replace a full video-editing application. Users should expect to prepare suitable footage and manage the surrounding ingest, storage, encoding, and delivery workflow themselves.
No public per-request, per-video, or recurring usage price was identified in the supplied NVIDIA documentation. Therefore, there is no verified price amount or standard billing period to report. Deployment cost will depend on GPU infrastructure, software licensing or support arrangements, and operational usage, but those costs should not be presented as a fixed Relighting price without an applicable NVIDIA quotation or published offer.
Best use cases
Relighting is a strong fit when the main requirement is to alter illumination without replacing the entire scene. Suitable applications include:
- Virtual production and previsualization, where a subject needs to be tested under different lighting environments.
- Film, television, advertising, and other media-effects workflows that need repeatable lighting variations.
- Live or near-live video effects that benefit from streaming inference.
- Background compositing, including workflows that combine a foreground subject with a custom image or HDR projection.
- Lighting adaptation for footage recorded under conditions that do not match the intended final environment.
It may be especially useful when an organization already operates NVIDIA GPUs and wants to keep video processing inside its own data center, workstation, or production infrastructure.
When to choose NVIDIA Relighting
Choose NVIDIA Relighting when you need controlled, repeatable lighting changes on existing video and can support NVIDIA GPU-based deployment. Its main advantage over a general-purpose video-generation system is directness: the workflow begins with footage and targets a defined lighting transformation rather than asking a model to invent an entire sequence.
A hosted video editor may be more appropriate for occasional users who want a simple interface and do not have compatible NVIDIA hardware. A generative video model may be more appropriate when the goal is to create new scenes or substantially alter the content of a shot. A conventional color-grading or compositing tool may be preferable when artists need manual, frame-level control or when the lighting effect does not require neural foreground separation.
Relighting is therefore best understood as a specialized GPU video service. It offers meaningful control and documented real-time-oriented performance, but it is not a general-purpose AI assistant, a text-to-video model, or a low-friction consumer application.
Position in NVIDIA's catalog
Within NVIDIA's current AI for Media and NIM catalog, Relighting represents a focused video-effects workload. It uses the same broader NVIDIA deployment ecosystem as other specialized media services, but its function is specifically virtual illumination and compositing. The model's value comes from combining that narrow task with GPU acceleration, streaming support, and configurable production parameters.
For evaluation purposes, the most important verified facts are its video-only transformation role, HDR-based lighting controls, NVIDIA GPU requirements, two inference modes, and provider-reported 1080p performance figures. No verified information was supplied for a conventional context length, token limits, public usage pricing, or general-purpose reasoning and coding capabilities.