Wan 2.2 14B T2V ComfyUI template (video_wan2_2_14B_t2v): six files, Enable Lightning LoRA switch, 20 or 4 steps and duration-driven length

Wan 2.2 14B text-to-video in ComfyUI: set up and run the official template Scope. This guide reads the official video_wan2_2_14B_t2v.json in Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2 (SHA256 dada2276...0207, 81,344 bytes). The template is unchanged from the earlier 3c2199b revision used by the published template indexes. Every value below was traced link by link in the JSON. Nothing was run and no weights were downloaded, so these are the settings the graph wires, not observed results. 1. Prerequisites - ComfyUI with subgraph support and the core ComfySwitchNode, ComfyMathExpression and Primitive nodes. Every node in the graph is a comfy-core node; no custom node pack is needed. The saved node metadata names comfy-core 0.3.45/0.3.49 for the Wan nodes and 0.18.1 for the switch, math and primitive nodes, and the file was saved by frontend 1.42.8. These are saved strings, not a tested minimum: if a node shows as missing, update ComfyUI first. A missing node named with a UUID is the embedded subgraph, not a custom node. - Six model files. The template's Model Links note points to the Hugging Face repos Comfy-Org/Wan_2.2_ComfyUI_Repackaged and Comfy-Org/Wan_2.1_ComfyUI_repackaged (the 2.1 repo supplies only the text encoder) and states these sizes: models/diffusion_models/wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors (13.31 GB) models/diffusion_models/wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors (13.31 GB) models/loras/wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors (1.14 GB) models/loras/wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors (1.14 GB) models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors (6.27 GB) models/vae/wan_2.1_vae.safetensors (242.1 MB) These are the T2V files. Do not load the I2V high/low models or the I2V v1 LoRAs that the 14B image-to-video template uses. 2. The controls you edit Everything you normally change is promoted onto the root node "Text to Video(Wan2.2)" (node 128): prompt, width (640), height (640), duration in seconds (5), the high- and low-noise model and LoRA names, clip_name, vae_name and Enable Lightning LoRA (saved off). The values are stored in the inner nodes of subgraph cb63cc72-d82e-4c9b-988d-6807f538d6ab. - Length is computed, not typed. EmptyHunyuanLatentVideo's length input is linked from a math node, floor(a * b) + 1, where a is duration and b is the FPS primitive (16). Five seconds therefore gives 81 frames by arithmetic. Change duration on the wrapper. Editing the saved length widget on the latent node has no effect while that link exists. - Width and height on the latent node are linked from the wrapper too. - FPS (16) and the negative prompt (the stock Chinese Wan negative prompt) are not promoted. Open the subgraph to change them. The same FPS primitive also feeds CreateVideo. 3. What actually reaches the two samplers Two KSamplerAdvanced nodes split the work: node 81 runs the high-noise expert and node 78 the low-noise expert. Both use euler with the simple scheduler, and each model passes through ModelSamplingSD3 with shift 5.0. Steps, CFG and the handoff step come from ComfySwitchNode nodes that all read the single Enable Lightning LoRA boolean (node 129). - Lightning off (the saved default): 20 steps, CFG 3.5. The high-noise sampler runs steps 0-10, adds noise and returns the leftover noise; the low-noise sampler runs steps 10-20 without adding noise. The model switches pass the raw high and low UNETs. - Lightning on: 4 steps, CFG 1. High-noise runs 0-2 and low-noise runs 2-4, and each model goes through its v1.1 lightx2v LoRA at strength 1.0. Do not copy the sampler widgets. Both samplers store 4 steps and CFG 1 in their own widgets, but those inputs are linked, so the switch values above win. - Seed: the high-noise sampler stores noise_seed 923510416338945 with control randomize, so each run changes it unless you set it to fixed. The low-noise sampler adds no noise. To reproduce a run, record the seed and use fixed. 4. Run it 1) Put the six files in the folders above. 2) Open the template from the Template Library or load the JSON. 3) On the wrapper, check the two model names, the two LoRA names, umt5 and the VAE. 4) Write the prompt, then set width, height and duration. 5) Choose Lightning on (4 steps) or off (20 steps). 6) Run. SaveVideo writes one video with the filename prefix video/ComfyUI, using the auto format and codec, at 16 fps; with the saved settings that is 81 frames at 640x640 by arithmetic. 5. Failure cases and checks - The Comfy docs Wan 2.2 page, as captured on 2026-10-28, lists only four T2V downloads (both diffusion models, the VAE and the text encoder). The template note and graph also name the two v1.1 lightx2v LoRAs, which the Lightning path loads. Install them before switching Lightning on. The graph does not show whether a runtime skips the unselected LoRA branch when Lightning is off, so do not assume you can delete the LoRA loaders or files. - That docs page's step 5 says to change size and frame count on EmptyHunyuanLatentVideo. In this graph those inputs are linked, so use the wrapper's width, height and duration instead. - The docs page's community section links a lightx2v Wan2.2-T2V-A14B 4-step rank-64 V1 folder, while this template names the Comfy-repackaged v1.1 high/low pair. If you swap LoRAs, change both files together and re-check the 4-step settings. - Mixed variants: the T2V and I2V files share folder names. A T2V graph with I2V weights, or with the I2V v1 LoRAs, is a different, untested setup. - Speed and memory: the template's VRAM Usage note is an author table for an RTX 4090D 24 GB at 640x640. It reports about 84% VRAM and roughly 536 s first run / 513 s second run without the LoRAs, and about 89% VRAM and roughly 108 s / 71 s with the 4-step LoRAs. These are author statements, not measurements here, and they do not establish that the template fits a 12 GB card. 6. Related records The Wan 2.2 14B image-to-video template uses the same switch structure (20 steps, CFG 3.5, handoff at 10, or 4/1/2) with I2V weights; see "Wan2.2 14B I2V ComfyUI: compatible files, tutorial conflicts and actual linked sampler defaults". For the repackaged file map and the 5B template, see "Wan 2.2 in ComfyUI: repackaged file locations and official templates" and "ComfyUI Wan 2.2 5B template (video_wan2_2_5B_ti2v)". Evidence and rights. Evidence status: source_inspected, not_run, unreviewed. This is shared-founder desk curation by the tutorial lane. The workflow repository declares the MIT licence at this commit; that covers this analysis of the template, not the model weights, LoRAs, text encoder, VAE, prompts or generated videos, which keep their own terms.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/video_wan2_2_14B_t2v.json",
      "note": "Exact pinned template; node byte ranges and excerpt hashes are in the source pins."
    },
    {
      "url": "https://docs.comfy.org/tutorials/video/wan/wan2_2.md",
      "note": "Existing native capture of the Comfy docs Wan 2.2 page (2026-10-28); T2V download list, steps and community LoRA link compared with the graph."
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/LICENSE",
      "note": "Repository MIT licence at the same commit (existing native source)."
    }
  ]
}