Wan 2.1 video transformation: mechanisms and practical controls

The single named result is output:114, labelled Sweet Tea Preview Video. CreateVideo 68 supplies its VIDEO output at slot 0 to terminal node 114. The recipe exposes no separate edge preview or batch-item files.

For the initial run, keep Reference image and Edge guidance bypassed and CausVid LoRA enabled. Retain 720 × 720 pixels, 81 frames, 16 FPS, latent batch size 1, VACE strength 1, 4 sampling steps and CFG 1. This is the documented starting configuration; no executed example or measured quality result is retained.

The task family is generate_video: existing frames condition a newly generated sequence. Some retained port classifications conflict with native input types—for example, prompt is classified as video and CFG as text. Use the actual bindings: prompt → node 6 text (STRING), frames → node 49 length (INT), CFG → node 3 cfg (FLOAT), strength_clip → node 107 strength_clip (FLOAT), and bit_depth → node 68 bit_depth (INT). Those port labels do not establish additional media inputs or preservation capabilities.

The retained preparation inventory requires four model files and the tea-time-labs/sweet-tea-nodes node pack. The node pack is a prerequisite resource, not an executable fallback recipe.

The graph selects these files and model categories:

• UNETLoader 106: wan2.1_vace_14B_fp16.safetensors in diffusion_models, with weight_dtype=default.
• CLIPLoader 110: umt5_xxl_fp16.safetensors in text_encoders, with type=wan and device=default.
• LoraLoader 107: Wan21_CausVid_14B_T2V_lora_rank32.safetensors in loras, used when CausVid is enabled.
• VAELoader 105: wan_2.1_vae.safetensors in vae.

These selectors are fixed in the exposed recipe. All four model requirements remain unresolved in preparation: artifact checksums, sizes and immutable provider versions are absent; the recorded provider version is main. The node-pack requirement is resolved, but no pinned pack commit is retained. Core node metadata records versions 0.3.34 and 0.3.38; these describe provenance and do not establish a minimum compatible runtime.

GPU capacity, system RAM, execution mode, supported platforms and comparable timing remain unspecified. Preparation's generic 1024 × 1024 workload differs from the graph's 720 × 720 defaults and is not a measured profile for this configuration. Interface dimension limits likewise do not establish hardware feasibility.

LoadVideo 145 feeds GetVideoComponents 144. Its IMAGE output at slot 0 reaches Canny 147 and then WanVaceToVideo 49 control_video. Canny is saved in mode 4; the retained interface explicitly identifies this as the default bypassed Edge guidance stage, which passes original frames through. Enabling the stage instead supplies Canny-processed source edges. The low and high thresholds default to 0.4 and 0.8; each accepts 0.01–0.99 in 0.01 steps and applies only while that stage is enabled.

LoadImage 134 slot 0 connects to node 49 reference_image. The interface defaults Reference image to bypassed even though the raw graph retains a sample filename and connection. Upload an image and enable the stage to add appearance guidance. The port role identity_reference does not establish identity preservation. Node 49 also has a real optional control_masks input in its schema, but no saved value, connection or exposed mask port. No mask polarity, protected region or original-layer compositing path is established here.

CLIPLoader 110 passes through LoraLoader 107 before CLIPTextEncode nodes 6 and 7 encode positive and negative text. Their conditioning enters node 49 alongside the VAE and visual inputs. Node 49 supplies positive conditioning at slot 0, negative conditioning at slot 1 and LATENT at slot 2 to KSampler 3. The model follows UNETLoader 106 → LoraLoader 107 → ModelSamplingSD3 48 → KSampler 3.

KSampler's LATENT passes to TrimVideoLatent 58. Its trim_amount is connected to node 49 slot 3; the saved widget value 0 is not an independent effective trim setting. Keep that connection. VAEDecode 8 decodes the trimmed latent using VAE 105. CreateVideo 68 assembles the images at its own FPS and bit depth. Its optional audio input exists but is unconnected; the source audio and FPS outputs are unused.

These connections explain how the supplied signals reach generation. Internal VACE mask handling, source-frame resizing or temporal resampling, exact output trimming and quantitative resource effects are not characterized by the retained evidence. Source conditioning does not establish unchanged pixels or guaranteed visual identity.

CausVid is enabled by default. LoraLoader 107 uses model strength approximately 0.3, saved as 0.30000000000000004, and text-encoder strength 1. Both accept −100 to 100 in 0.01 steps. Model strength scales the diffusion-model modification. Text-encoder strength only affects matching text-encoder weights supplied by the LoRA; their presence in this artifact is not established here. Neither strength applies while the stage is bypassed.

The enabled starting configuration uses KSampler 3 steps=4, cfg=1, sampler_name=uni_pc, scheduler=simple and denoise=1. Steps sets the denoising-step count, with a range of 1–10000. CFG controls prompt guidance, accepts 0–100 in 0.1 steps, and has a schema warning that excessively high values can impair quality. The retained evidence does not establish the magnitude of negative-prompt influence at CFG 1. Sampler selects the denoising algorithm and scheduler controls noise removal.

Denoise accepts 0–1 in 0.01 steps and acts on the VACE-created latent. The generic sampler tooltip about retaining initial-image structure does not establish original-video preservation in this graph. VACE strength, which accepts 0–1000 in 0.01 steps, controls visual conditioning separately. ModelSamplingSD3 48 applies shift before sampling, saved as 8.000000000000002, conventionally described as 8; its range is 0–100 in 0.01 steps. The retained evidence does not explain a specific visual effect for changing shift.

The interface documents seed −1 as a randomization sentinel, while its numeric constraint and the raw KSampler schema require a nonnegative value. The graph saves 654654950714624, with a separate non-runtime randomize widget. Use the interface's documented randomization behavior where supported; do not write −1 directly to the native node. A fixed nonnegative seed reuses noise initialization without establishing identical output across other configuration or runtime changes.

Bypassing CausVid does not change steps or CFG automatically. The supplied recommendation for that configuration is 20 steps and CFG 6. Bypass is an interface stage action, not a node input supported by set_input. The controls below change only the actual sampler inputs and must be used together after bypassing the stage.

Node 49 sets generated width, height, length and batch_size. Width and height default to 720 pixels and use multiples of 16. The interface permits 16–16384 pixels. The frames port maps to length, default 81, advancing from 1 in steps of four: 1, 5, 9 and so on; its interface maximum is 16384, so choose a value on the stated sequence within that bound. These limits do not establish usable hardware capacity. Latent batch size accepts 1–4096; keep 1 for the documented single-video use because separate batch-item files are not exposed.

CreateVideo 68 fps defaults to 16 and accepts 1–120 in steps of 1. Requested duration is frame count divided by output FPS: 81/16 is about 5.1 seconds. Changing FPS changes playback duration for the same frames; it does not transfer source FPS or generate extra frames. Source-frame temporal alignment is not documented, so matching source FPS alone does not establish matching motion timing.

Node 68 bit_depth defaults to 8 and also accepts 10. Its schema describes smoother gradients with less banding at 10-bit, with possible player and downstream compatibility limits. This setting concerns assembled-video representation, separately from model precision.

SweetTeaPreviewVideo 114 exposes transport settings: format=auto, with auto/mp4/mkv/webm choices; codec=auto, with auto/h264/av1 choices; and hidden crf=-1, with an interface range of −1–63 in steps of 1. Their ports are output_format_114, output_codec_114 and output_crf_114. Auto and −1 retain transport defaults. Supported container/codec combinations and exact CRF sentinel implementation are not established here, so the menu choices do not guarantee every combination. The terminal is described as a no-permanent-save output transport; retain the returned result as needed for your work.