Kandinsky 5 text to video: mechanisms and practical controls

Open Kandinsky 5 text to video, version 0.1.7, and begin with the supplied configuration. Its named output is Generated video, port output:11: one video, not separate frame or batch-member outputs. Defaults request 768 × 512 pixels, 121 frames and 24 FPS—approximately five seconds—with MP4/H.264 encoding and CRF -1. This recipe selects the fixed Kandinsky 5 Lite 5-second model; frame count is not a model-duration selector.

The retained positive prompt describes a woman talking off-camera beside a stained-glass window at sunset. It is an editable example, not an executed-result record. No image or audio upload is required. Confirm the prerequisites in the next section before running; the pack does not establish that they are installed in your environment.

The preparation record lists five required components, all required together. The graph identifies these model selections and directories:

• kandinsky5lite_t2v_sft_5s.safetensors in diffusion_models, selected by 20 (UNETLoader.unet_name), with weight_dtype=default.
• qwen_2.5_vl_7b_fp8_scaled.safetensors and clip_l.safetensors in text_encoders, selected by 23 (DualCLIPLoader.clip_name1 and clip_name2). Keep type=kandinsky5; the saved device is default.
• hunyuan_video_vae_bf16.safetensors in vae, selected by 25 (VAELoader.vae_name).
• The required tea-time-labs/sweet-tea-nodes node pack. The graph's terminal class is SweetTeaPreviewVideo.

The Qwen encoder's preparation resource identity is unresolved, so it has no prerequisite resource link here. It is still required and has a retained filename and model URL in the graph. Model components list provider version main, but no component checksums are supplied; the node pack has no retained commit or version. Resource links do not establish immutable installed artifacts.

Node 25 retains its VAE filename through a primary-source recovery marked as requiring independent review. Its connections to conditioning and decoding are unchanged. Keep the retained filename; this pack does not establish compatibility with a substitute VAE or supply a completed independent review.

There is no scoped GPU/RAM recommendation, comparable run timing, or supported-platform list. Preparation's generic 1024 × 1024 workload is not the graph's 768 × 512 default and is not a measured profile. Node metadata records core versions 0.3.68 and 0.3.71 on some nodes, not a tested minimum runtime version for the complete recipe. Use a runtime that provides the listed classes and inputs.

Node 23 supplies the encoder object to both CLIPTextEncode nodes: positive text at 22 and negative text at 24. Their slot-0 outputs enter 21 (Kandinsky5ImageToVideo) as positive and negative. Node 25 supplies its VAE to both 21.vae and 17.vae.

Despite node 21's class name, its real optional start_image input has no saved value or connection. The selected configuration is text-to-video from an empty video latent, not a connected image-animation workflow. Node 21 sends conditioning slots 0 and 1 to 18.positive and 18.negative, and latent slot 2 to 18.latent_image. Its other LATENT output, slot 3, is unused; its purpose is not established here.

The model path is 20 (UNETLoader) → 19 (ModelSamplingSD3, shift 5) → 18 (KSampler.model). Sampling returns latent slot 0 to 17.samples. VAEDecode at 17 turns those samples into images, which feed 16 (CreateVideo.images). Node 16 assembles the frames at the selected FPS and sends VIDEO slot 0 to 11.video (SweetTeaPreviewVideo). The interface exposes that path as Generated video, port output:11.

CreateVideo.audio is a real optional input but has no saved value or connection. The selected output is silent. Optional native inputs do not by themselves provide upload controls or supported image/audio adaptations. The interface's frames port is classified as a video source in its port metadata, but its actual binding is the integer 21.length; it is a frame-count control, not a video upload.

These connections explain where conditioning, sampling, decoding, playback and encoding act. The retained execution-effects inventory characterizes none of the 11 nodes; there are no compiled mechanism bindings or adaptations. That absence does not erase the graph or its real controls, but it does not establish additional motion, quality or resource-use guarantees.

Width and height bind to 21.width and 21.height: defaults 768 and 512, minimum 16, in multiples of 16. Together they set spatial dimensions and aspect ratio. Larger dimensions increase computation and memory requirements; no numerical estimate is retained.

Frames binds to 21.length: default 121, minimum 1, increments of 4. Use the 4n + 1 pattern. More generated frames increase work and memory demand. Changing frame count does not select a different model-duration variant, and the schema's permitted values do not establish equal quality at other lengths.

FPS binds to 16.fps: default 24, range 1–120, step 1. Approximate playback duration is frames divided by FPS: 121/24 is about 5.04 seconds. Changing FPS plays the same generated frames faster or slower; it does not create extra frames.

21.batch_size defaults to 1, with schema range 1–4096. It controls latent sequences generated together, but this path has one video output and no separate batch-member outputs. Keep 1 for a single clip; the large schema maximum is not a capacity recommendation.

For a first run, keep the supplied shape, frame count, FPS and batch size. After prerequisites are available and prompts are chosen, run the recipe and inspect Generated video before changing advanced settings.

Sampling defaults at 18 are 50 steps, CFG 5, euler_ancestral, beta, and denoise 1. Steps are denoising iterations (1–10000); more steps increase runtime, without a retained quality or timing curve. CFG is prompt-guidance strength (0–100, step 0.1). Higher values emphasize conditioning, but excessive guidance can harm quality. Negative text at 24.text defaults to empty; its influence depends on CFG, so unwanted traits are not guaranteed to disappear.

Sampler binds to 18.sampler_name; scheduler binds to 18.scheduler. They select the denoising algorithm and distribution of noise levels respectively. Alternative choices come from the runtime, not a retained complete option list. Sampling shift binds to 19.shift (default 5, range 0–100, step 0.01). The retained description identifies a noise-schedule adjustment but gives no directional tuning rule; keep 5 initially rather than treating it as a known motion-strength control.

Denoise binds to 18.denoise (0–1, step 0.01). Keep 1 initially. Lowering it changes sampling here; it cannot preserve an uploaded image because none is connected.

Seed has two retained representations: the interface advertises -1 as a randomization sentinel, while native 18.seed requires a nonnegative integer and the graph saves 102704842425568. The saved widget contains randomize, but that widget index is explicitly non-runtime. Proxy metadata mentions control_after_generate; there is no executable input by that name on node 18. The pack does not show the interface-to-native randomization translation. Use the interface's documented behavior when working in that UI, but do not write -1 directly into the native seed input.

For comparisons, use a fixed nonnegative seed, retain the same model and other settings, and change one control at a time. A fixed seed controls starting noise; it does not promise identical output across changed settings or runtimes. These are comparison instructions, not recorded experiments.

Node 11 exposes format, codec and crf through ports output_format_11, output_codec_11 and output_crf_11. Container options are auto, mp4, mkv, webm; codec options are auto, h264, av1. Defaults are mp4 and h264. Other combinations depend on output transport support; listed options do not establish every pairing as supported.

CRF defaults to -1, meaning the transport's encoder default; explicit values are 0–63. The actual default quality and codec-specific behavior are not retained. These controls affect encoding of the returned file, not scene conditioning, generated motion or frame count.

16.bit_depth is a real native input saved as 8, although it is not an exposed recipe interface port. Its schema also permits 10. The tooltip describes smoother gradients with less banding at 10-bit, with possible player or downstream-node incompatibility. It does not establish end-to-end 10-bit support through node 11 and the selected transport. Keep 8 for the baseline.

The decoder at 17 is VAEDecode, not VAEDecodeTiled. It has no tile-size input. The retained tiled-decode implementation concerns a different node class and decode-stage work only; applying it here would require a graph change, not a scalar control adjustment. No executable tiled-decode adaptation or fallback recipe is supplied.