Wan 2.2 first–last frame video: mechanisms and practical controls

Create a silent video by supplying beginning and ending images, replacing the saved motion prompt, and starting with the retained defaults. The configured result is 640 × 640 pixels, 81 frames at 16 FPS and approximately 5.1 seconds of playback. This duration is calculated playback length, not generation time or an executed example. Keep batch size 1 for a single clip.

The named result is output:83, labeled Sweet Tea Preview Video. Its source is slot 0 of CreateVideo node 86, connected to the video input of terminal node 83. Intermediate frames and latents are not separate exposed outputs.

Both endpoint loaders are enabled by default, although their interface upload defaults are empty. The graph retains example filenames; their presence does not establish that those files are available in your runtime. Resolve the prerequisites below before generating.

The selected graph requires these four model files and the Sweet Tea node pack:

Component	Native binding and directory
umt5_xxl_fp8_e4m3fn_scaled.safetensors	CLIPLoader 72, clip_name; text_encoders; type=wan, device=default
wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors	UNETLoader 76, unet_name; diffusion_models; weight_dtype=default
wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors	UNETLoader 77, unet_name; diffusion_models; weight_dtype=default
wan_2.1_vae.safetensors	VAELoader 79, vae_name; vae
tea-time-labs/sweet-tea-nodes	Required node pack for the Sweet Tea preview path

All five components are mandatory and recorded as unresolved in preparation; this is a catalog state, not a report that your installation lacks them. Model provider versions are recorded as main, without component checksums. No node-pack version or commit is retained. Model filenames containing fp8 do not change the recorded UNETLoader weight_dtype setting from default, and do not establish a complete runtime precision configuration.

Graph metadata records comfy-core 0.3.45 on most core nodes and 0.3.48 on WanFirstLastFrameToVideo 81. These are retained node metadata, not a tested minimum runtime version. No scoped GPU or RAM recommendation, supported platform list, or comparable generation timing is supplied. The preparation record's 1024 × 1024 Standard configuration differs from the graph's 640 × 640 default and has no measured profile. Neither configuration establishes hardware fit.

LoadImage nodes 80 and 89 feed IMAGE slot 0 into start_image and end_image on WanFirstLastFrameToVideo 81. Their MASK slots are unused. CLIPLoader 72 supplies CLIPTextEncode nodes 90 and 78, whose outputs enter node 81 as positive and negative. Node 81 also receives vae from 79 and the requested width, height, length and latent batch size. Its slots 0 and 1 provide positive and negative conditioning to both samplers; slot 2 provides the initial latent to node 84. The optional clip_vision_start_image and clip_vision_end_image inputs are unconnected.

The retained source describes the images as constraints on the enabled boundaries. It does not supply the internal endpoint encoding, resizing or masking implementation, or establish pixel-exact copying. Uploaded alpha masks have no connected consumer, so they do not protect regions or preserve transparency. Motion and negative text condition both stages regardless of endpoint selection.

The task family is generate_video. The interface's inferred role and modality labels are not mechanisms: frames and fps are numeric controls, guidance ports bind to numeric cfg, and the initial sampler and shift controls do not supply source images. Use the actual graph bindings when adapting this recipe. Its empty compiled capability and rule lists do not negate the retained graph's documented video path.

The shape controls affect node 81; playback controls affect CreateVideo 86 after decoding.

Interface port	Native input	Starting value and accepted interface values
width	81 width	640 px; 16–16384 in increments of 16
height	81 height	640 px; 16–16384 in increments of 16
frames	81 length	81; 1, 5, 9 and so on, with declared maximum 16384
batch_size	81 batch_size	1; 1–4096
fps	86 fps	16; 1–120 in increments of 1

For frame count, the stated sequence is 1 + 4k for nonnegative integer k. Its largest value within the declared maximum is 16381, not 16384. These interface limits describe control ranges, not feasible hardware configurations. Larger latent batches do not create separate named video outputs; retain batch size 1 for a single clip.

Playback duration is frame count divided by FPS. Changing FPS changes the timing of the generated frames without adding frames. Changing length changes the requested generated sequence. Neither shape nor frame-count changes have numerical memory or speed estimates in the retained pack. The adaptations below are unexecuted control examples, not measured configurations.

The high-noise model travels from UNETLoader 76 through ModelSamplingSD3 73 to KSamplerAdvanced 84. The low-noise model travels from 77 through ModelSamplingSD3 74 to KSamplerAdvanced 87. Node 84 adds initial noise, samples from step 0 to the handoff at 10 in a 20-step schedule, and returns a latent with leftover noise. Node 87 receives that latent, adds no fresh noise, resumes at step 10 and completes the schedule. Its saved end_at_step=10000 does not make this a 10000-step run: its total steps value is 20. Refinement is part of the complete generation, not a separately requested finishing operation.

Interface port	Native binding	Retained default
initial_steps, refinement_steps	84/87 steps	20 each; total schedule length for each sampler
initial_end_step, refinement_start_step	84 end_at_step, 87 start_at_step	10 each; shared handoff boundary
initial_sampler, refinement_sampler	84/87 sampler_name	euler each
initial_scheduler, refinement_scheduler	84/87 scheduler	simple each
initial_guidance, refinement_guidance	84/87 cfg	4 each
initial_shift, refinement_shift	73/74 shift	8.000000000000002 and 8
seed	84 noise_seed	Interface -1; saved graph 984937593540091

For a deliberate schedule edit, keep both total step counts equal, match the schedulers, and make the initial handoff equal the refinement start within that schedule. Two steps=20 values describe two portions of a coordinated schedule, not two independent full 20-step passes. Steps accept 1–10000 and handoff inputs 0–10000; their individual ranges do not enforce the required relationship. Restore 20/20 and 10/10 if an experiment leaves the intended handoff unclear.

Guidance is independently adjustable for each stage, from 0–100 in increments of 0.1. Shift is applied independently to each model before sampling, from 0–100 in increments of 0.01. Sampler and scheduler alternatives are exposed by the interface, but the retained sources do not rank them or establish an optimal guidance, shift or handoff. Matching total steps and schedulers follows the retained configuration guidance; it is not proof that every experimental setting has equivalent behavior.

VAEDecode 85 receives samples from sampler 87 and vae from loader 79, producing IMAGE frames at slot 0. CreateVideo 86 receives those frames through images and produces VIDEO at slot 0. SweetTeaPreviewVideo 83 receives that VIDEO through video and returns the named result output:83. No audio is connected or exposed as an input in this recipe.

Interface port	Native binding	Default and choices
bit_depth	86 bit_depth	8; 8 or 10
output_format_83	83 format	auto; auto, mp4, mkv, webm
output_codec_83	83 codec	auto; auto, h264, av1
output_crf_83	83 crf	-1; -1–63 in increments of 1

The native bit-depth tooltip says 10-bit keeps smoother gradients with less banding, but some players and downstream nodes may not support it. This supports trying 10-bit video creation when gradients matter; it does not guarantee an end-to-end 10-bit result for every encoding choice. Format selects the container, codec selects encoding, and CRF -1 retains automatic quality selection. CRF interpretation depends on the codec. No complete compatibility matrix, automatic-selection algorithm or codec-specific quality recommendation is retained.

These choices affect video creation and encoding after sampling. They do not repair endpoint conditioning or change the generated motion. Start with automatic format and codec, CRF -1 and 8-bit unless delivery requirements call for another supported choice.

Node 85 is VAEDecode, not VAEDecodeTiled, and exposes no tile controls. The retained general tiled-decode source describes a separate decode-stage mechanism; it does not create a tile_size input in this revision or supply a selected replacement recipe. Tiling would also leave model loading and denoising unchanged, so it cannot be treated as a remedy for every memory failure.