Skip to content

How it works · Working knowledge

Luma video with camera motion: mechanisms and practical controls

Receive one Generated video through output:8. Start by uploading a Start image or bypassing that stage, then replace the example prompt while keeping ray-2, 720p, and 5s.

Use this exact recipe for Luma video generation with optional first- and last-frame references and camera-motion concepts.

Input: textInput: imageOutput: videoluma

Before you start

  • Prompt: Describe the scene, subject movement, and desired appearance; replace the retained flower example.
  • Start image (optional): Upload a first-frame reference or bypass Start image, which is initially enabled.
  • End image (optional): Enable End frame and upload a final-frame reference if needed; this stage is initially bypassed.

Start with one video

The result is Generated video, exposed as output:8: one video, not separate outputs for individual camera concepts. The starting generation settings are ray-2, 720p, 5s, and Loop off. The main camera stage initially requests zoom_in and orbit_left. End frame and Additional camera motion are initially bypassed.

Start image is optional but initially enabled, so it needs an upload unless you bypass it. For text alone, bypass Start image and leave End frame disabled. The retained peach-colored dahlia prompt is an editable example, not evidence of an executed result. This is video generation, not a still-image editing or original-pixel-preservation workflow.

What should guide the video?

  • A starting image: Keep Start image enabled and supply an upload.Continue
  • Text alone: Bypass Start image and leave End frame disabled.Continue
  • Starting and ending references: Supply a Start image, enable End frame, and upload an End image; provider compatibility still applies.Continue

Prepare access and understand the transfer

Configured partner API access and sufficient credits are required. Generation sends the prompt and enabled reference images to Luma and involves a paid provider call. Supply only material you can send to that provider. The pack does not provide access-configuration instructions or prices.

The retained required node-pack component is tea-time-labs/sweet-tea-nodes; no exact pack version or checksum is supplied. This component is a prerequisite, not an executable fallback recipe. Execution mode, supported platforms, GPU capacity, system-memory capacity, and comparable run timing are not established. The preparation record's 1024 × 1024 workload label does not establish generated video dimensions; the retained generation resolution is 720p.

How references reach generation

Node 9 (LoadImage) supplies IMAGE slot 0 to node 5 (LumaImageToVideoNode) input first_image. Node 23 (LoadImage) connects IMAGE slot 0 to last_image, but End frame is initially bypassed (mode: 4). These references guide the first and final frames; the pack does not establish pixel-exact preservation. Both loaders also expose MASK outputs, but neither mask is connected.

The loaders' real image inputs are declared in schema facts even though their inputs objects contain no saved image value. Both widgets retain api_luma_i2v_input_image.png; that filename does not establish an available runtime upload.

Node 5 receives the prompt, model, resolution, duration, loop, seed, frame references, and connected camera concepts. Its VIDEO slot 0 feeds node 8 (SweetTeaPreviewVideo) input video. Model choices can change supported resolutions, durations, and reference combinations. The pack lists the baseline but not the complete model, resolution, or duration menus or compatibility matrix. Loop requests a looping video; it does not establish compatibility with every reference combination or guarantee a seamless result.

  1. Upload your Start image, or bypass Start image to omit the first-frame reference.

    Look for: The enabled loader has an uploaded image, or the optional reference is omitted by bypass.

  2. Leave End frame bypassed unless you need a final-frame reference. To use one, enable End frame and upload the End image.

    Look for: The final-frame reference is omitted or supplied through the enabled End frame stage.

  3. Replace the flower prompt with your scene, subject movement, and desired appearance. Start with ray-2, 720p, 5s, and Loop off; use only provider-supported combinations when changing these settings.

    Look for: The recipe is configured for your description; successful generation returns one Generated video.

    Outputs: output:8

Request looping
Sets the generation loop request to true. Use only a provider-supported model and reference combination; this does not enable or disable either image stage.

An enabled image stage has no usable reference.

Start image is enabled initially, while neither loader has a saved image input value.

Check: A retained widget filename is not evidence of an available runtime upload.

Upload the intended image or bypass the optional stage. Enable End frame only when supplying its end image.

A changed model or setting combination is rejected.

Supported resolution, duration, looping, and frame-reference combinations depend on the provider and model.

Check: The draft contains no exhaustive compatibility matrix.

Return to ray-2, 720p, 5s, Loop off and omit the optional end reference while resolving the rejected combination. This retained starting configuration is not an execution guarantee.

Choose camera concepts and keep their stages active

Node 24 (LumaConceptsNode) sends LUMA_CONCEPTS slot 0 to node 5 input luma_concepts. Its independent inputs concept1–concept4 start as zoom_in, orbit_left, None, and None. None leaves a slot unused. These labels request camera concepts, not a measured trajectory or guaranteed camera path.

Node 27, another LumaConceptsNode, connects slot 0 to node 24 input luma_concepts. Its four additional slots all start as None, and this stage is initially bypassed (mode: 4). Additional concepts affect generation only when both Additional camera motion and the main Camera motion stage are enabled. Bypassing the main stage omits both sets.

Stage bypass is a UI operation here, not a native set_input delta. Editing a concept slot does not enable its stage. The retained schema does not enumerate the complete concept menu or explain ordering and conflict resolution, so the graph does not establish a sequential camera choreography.

Remove the second camera concept
Omits orbit_left from slot 2. Starting from the retained configuration, zoom_in remains in slot 1 and slots 3–4 remain unused; stage states are unchanged.

How much explicit camera guidance do you need?

  • Keep the starting combination: Leave Camera motion enabled with zoom_in and orbit_left.
  • Simplify the combination: Set unwanted slots to None; removing orbit_left leaves the retained zoom_in request.
  • Omit all camera concepts: Bypass Camera motion; both concept sets are omitted, without guaranteeing a static camera.
  • Use additional slots: Enable both camera stages and select non-None additional concepts from the available UI choices.

Additional camera concepts appear to have no effect.

The additional stage is bypassed by default and depends on the main camera stage being enabled.

Check: Changing node 27's slot values alone does not enable either stage.

Enable both camera stages and choose a non-None additional slot. If the chain is active, the pack still establishes concept requests rather than guaranteed visible motion.

Separate reruns, generation, and output encoding

Node 5 saves native seed 63119199850581. The Resource interface instead defaults to -1, described as randomizing the rerun seed each run. Native schema facts specify a nonnegative integer, while the interface itself also lists a minimum of 0. Treat -1 as the documented Resource convention, not a validated direct native-input value; its translation is not shown in the pack. The retained randomize widget is marked non-runtime. A fixed seed controls whether the API node reruns; actual video results remain nondeterministic regardless of seed.

Node 8 (SweetTeaPreviewVideo) receives the already-generated video and returns Generated video through output:8. Its original terminal class is recorded as SaveVideo, but the selected graph's actual class is SweetTeaPreviewVideo. File type maps to format, offering auto, mp4, mkv, and webm. Codec maps to codec, offering auto, h264, and av1. Both default to auto. Quality (CRF) maps to crf, is hidden in the retained interface, and defaults to -1 for automatic quality selection; its interface range is -1 through 63 in steps of 1.

These controls affect output encoding, not Luma's scene, motion, model, or generation quality. The pack does not establish every container/codec combination or numerical quality and file-size effects. Keep automatic selection unless you have a delivery requirement.

Select MP4 output container
Requests an MP4 container without changing generation. Codec remains independently configured; the pack does not establish every encoding combination.
Restore automatic codec selection
Lets the output stage select the codec; it does not change the generated scene or camera requests.

The same seed produces a different video.

The seed governs rerunning, not deterministic provider generation.

Check: Node 5 explicitly describes results as nondeterministic regardless of seed.

Keep the delivered video you intend to use; do not rely on seed reuse to reconstruct it.

Changing codec or CRF does not fix unwanted scene content or motion.

These settings operate at output encoding rather than generation.

Check: Node 8 receives the video already produced by node 5.

Adjust the prompt, references, or camera concepts for content changes; use encoding controls for delivery requirements. A further generation can involve another paid provider call.

Where this approach stops

  • An exhaustive provider model, reference, and setting compatibility matrix
  • A complete camera-concept menu or guaranteed camera trajectory
  • Deterministic reproduction from a fixed seed
  • Measured timing or GPU and system-memory recommendations

Sources