Exploded view round trip: mechanisms and practical controls

The intended result is an outward-then-return video exposed as Sweet Tea Preview Video, port output:381. Configure the destination image before tuning motion: the destination—not just the outward motion prompt—needs to depict the separated object.

Preparation lists the required tea-time-labs/sweet-tea-nodes node pack, without a pinned version. The retained effects inventory flags Gemini nodes 395 and 397 and ByteDance nodes 244 and 232 for external data transfer and provider-priced calls. That assessment is partial and its applicability is unverified; it is not a complete characterization of all nodes. Confirm provider access, billing and permission to send the image before running. Exact access setup, prices and platform support are not supplied.

No scoped GPU/RAM recommendation or comparable run timing is retained. Preparation's 1024 × 1024 workload is metadata, not a promise about the generated video dimensions. Recorded core-node versions are 0.14.1 for LoadImage and 0.16.4 for the video and combination nodes; these records do not establish a minimum compatible installation version.

LoadImage 29 sends IMAGE slot 0 to GeminiNanoBanana2V2 395 through model.images.image_1. Node 395 requests a static photorealistic recreation. Its IMAGE slot 0 then feeds node 397's model.images.image_1: the destination is generated from the recreation, not directly from the upload. A prompt asking for an exact recreation does not establish unchanged pixels.

Both image nodes save model=Nano Banana 2 (Gemini 3.1 Flash Image), model.resolution=2K, model.aspect_ratio=auto, response_modalities=IMAGE, temperature=1, top_p=0.95 and seed=42. Starting-frame thinking is MINIMAL; destination thinking is HIGH. Those are the two thinking choices, but no timing or quality advantage is established. Both saved system prompts prioritize image generation and permit invention when visual instructions are vague.

Image aspect-ratio choices are auto, 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, 1:4, 4:1, 8:1, 1:8. Here auto follows the input image. Only the saved 2K resolution is established, not a full resolution choice list. The model selector also lists Nano Banana 2 Lite; the retained material does not establish how all dynamic settings behave after switching models.

Temperature accepts 0–2 and top_p 0–1, each in 0.01 increments. Lower temperature reduces randomness; lower top_p narrows nucleus sampling. These controls do not enforce geometry or preserve labels. Starting-frame controls are advanced; destination controls are core and independent.

All four interface seed defaults are -1, but the native schemas require nonnegative values and the saved widgets show randomize. Translation of the interface sentinel is undocumented. Both saved image seeds are 42; their schema permits nonnegative integers up to the recorded maximum 18446744073709552000. Use a modest nonnegative integer for a direct native edit rather than copying -1. A fixed Gemini seed offers best-effort repeatability, not determinism.

Some retained port metadata labels image-stage prompts and settings as video inputs. The graph and native schema establish their actual types: is text, is an integer, and these nodes' connected slot 0 products are images. Similarly, the audio-generation controls on the video nodes are booleans, not uploaded audio inputs. This is a video-generation recipe with image-generation stages, not a masked regional edit or a pixel-preserving composite.

ByteDanceFirstLastFrameNode 244 takes first_frame from 395 slot 0 and last_frame from 397 slot 0. Node 232 swaps those endpoints. The return is a separate generation, not reversed outward footage. The outward prompt requests separation along assembly seams; the return prompt requests components sliding back into place. These are requested motions, not a physical simulation.

Both nodes save model=seedance-1-5-pro-251215, duration=4 seconds, resolution=720p, aspect_ratio=adaptive, camera_fixed=false, watermark=false and generate_audio=true. Duration accepts integers 3–12. Resolution choices are 480p, 720p, 1080p; aspect ratios are adaptive, 16:9, 4:3, 1:1, 3:4, 9:16, 21:9. Keep both clips' resolution and aspect ratio matched for combination. Four seconds each implies approximately eight seconds combined, not a measured execution result. Return controls are advanced: changing a core outward control does not change its return counterpart.

camera_fixed=true appends a fixed-camera instruction; it does not guarantee a stationary camera. Each watermark boolean requests whether to add the provider's watermark to that clip. Both motion prompts request no dialogue, narration or voiceover, but that wording is not an audio-off switch. The schema says generate_audio is ignored except for Seedance 1.5 Pro; the selected model is the listed versioned 1.5 Pro option. Other model choices are seedance-1-0-pro-250528 and seedance-1-0-lite-i2v-250428; do not carry the audio-generation expectation over to them.

Saved seeds are 1561859130 outward and 642612484 return. Native video seeds accept integers 0–2147483647. The interface defaults of -1 differ from those saved inputs; their translation and the widgets' randomization behavior are not documented sufficiently to promise repeatable video.

GetVideoComponents 392 extracts the outward video from 244; 393 extracts the return video from 232. BatchImagesNode 259 receives 392 IMAGE slot 0 at images.image0, then 393 IMAGE slot 0 at images.image1. Its combined IMAGE output feeds CreateVideo 380 images.

CreateVideo's audio is connected only to 392 AUDIO slot 1, and fps only to 392 FLOAT slot 2. Return audio and frame rate are not connected. The fps widget displays 30, but the executable fps input is a connection—not a fixed 30 fps setting. A scalar control cannot overwrite that connection. No audio concatenation or policy for extending outward audio across the return portion is retained, nor is a return-frame-rate reconciliation policy documented.

Node 380 VIDEO slot 0 feeds SweetTeaPreviewVideo 381 through video, exposed as Sweet Tea Preview Video, output:381. The transport is described as no-permanent-save; do not assume a permanently saved artifact. Intermediate endpoint images and clips are graph products, not separately exposed output ports.

Bit depth defaults to 8; 10 is the other supported value. The schema describes smoother gradients at 10-bit but warns that some players and downstream nodes may not support it. This creation-stage setting does not establish that an automatically chosen transport codec will preserve 10-bit output. Transport saves format=auto, codec=auto, crf=-1. File type choices are auto, mp4, mkv, webm; codec choices are auto, h264, av1. CRF is hidden, with integer steps from -1 to 63. The retained material does not define auto selection, the -1 CRF sentinel, compatible codec/container combinations or numerical quality consequences.