Wan 2.2 14B Fun Inpaint ComfyUI template (video_wan2_2_14B_fun_inpaint): start and end frames, active 4-step LoRA group and bypassed 20-step group

Wan 2.2 14B Fun Inpaint in ComfyUI: the official template Scope. This guide reads the official video_wan2_2_14B_fun_inpaint.json in Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2 (SHA256 9233091e...8a93, 53,940 bytes). All 38 nodes and 44 links were traced statically. The file has no subgraphs and no switch nodes. Nothing was run, and no model, image or linked page was downloaded. 1. Prerequisites - A ComfyUI build with the core node WanFunInpaintToVideo. Every node is comfy-core: most were saved at version 0.3.45, and WanFunInpaintToVideo, LoadImage and the two LoRA loaders at 0.3.49. Those are saved version strings, not a tested minimum. The graph uses no custom node pack. - Files, as declared by the loaders and the Model Links note (node 155, sizes as stated there): - models/diffusion_models/wan2.2_fun_inpaint_high_noise_14B_fp8_scaled.safetensors and wan2.2_fun_inpaint_low_noise_14B_fp8_scaled.safetensors (13.31 GB each), from Comfy-Org/Wan_2.2_ComfyUI_Repackaged. - models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors (6.27 GB), from Comfy-Org/Wan_2.1_ComfyUI_repackaged. - models/vae/wan_2.1_vae.safetensors (242.1 MB). - models/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors and _low_noise.safetensors (1.14 GB each). These are the i2v Lightning LoRAs, applied here to the Fun Inpaint experts. They are needed for the group that is active as saved. The 4-step route needs about 35.4 GB of files. The 20-step route needs about 33 GB, with no LoRAs. No CLIP vision model is used: the clip_vision_output input of WanFunInpaintToVideo is left unconnected in both groups. 2. The two groups, and which one is active The file holds two complete copies of the pipeline: - "Wan2.2_fun_Inp fp8_scaled + 4 steps LoRA" (active as saved): 18 pipeline nodes, all active. - "Wan2.2_fun_Inp fp8_scaled" (bypassed as saved): 16 nodes, every one in mode 4. This is the reverse of the Wan 2.2 14B Fun Control and first-last-frame templates, whose 20-step group is the active one and whose LoRA group is bypassed. Note 159 says the Lightning LoRA loses video dynamics but reduces generation time, and that you enable one workflow as needed. Note 156 says to box-select a group and press Ctrl + B to enable it, and to disable the fp8_scaled + 4steps LoRA group after enabling the fp8_scaled group so both do not run, or to use partial execution. That wording matches what the file needs: to use the 20-step group, you must also bypass the LoRA group. The groups do not share loaders, prompts, images or size. Each has its own copy of the Step 1 to Step 4 sub-groups. 3. Start and end images (both groups) LoadImage 110 (example video_wan2_2_14B_fun_inpaint_start_image.png) feeds start_image on WanFunInpaintToVideo 111 (link 192). LoadImage 112 (example video_wan2_2_14B_fun_inpaint_end_image.png) feeds end_image on the same node (link 191). In the bypassed group, LoadImage 140 and 147 hold the same two example file names and feed WanFunInpaintToVideo 148. Unlike the first-last-frame template's LoRA group, the file names are not swapped between groups. This file always supplies both images. What the node does when an image is missing, or when an image does not match the output size, is not in the file; prepare both frames at the output aspect ratio. 4. Size and length WanFunInpaintToVideo 111 holds width 640, height 640, length 81 and batch size 1 (widget names from the node's serialized connectable widgets, in that order). Node 157 says the video is set smaller by default for low VRAM and can be raised if you have enough. 81 frames at CreateVideo's 16 fps is about 5.06 s. 81 has the 4n + 1 form that the Wan 2.2 reference code is documented to use (see the frame-count guide). Whether WanFunInpaintToVideo enforces or rounds length is not in this file. 5. Active route: 4-step LoRA - Prompts: positive CLIPTextEncode 99 ("A dreamy scene where a little cat is sleeping. Zoom in, and the cat opens its eyes, looks up, and blinks. In Q-style, with ice crystals.") and negative 91 (the usual Wan Chinese quality list). Both use CLIPLoader 90 (umt5_xxl_fp8_e4m3fn_scaled, type wan). VAELoader 92 loads wan_2.1_vae. - High noise: UNETLoader 101 feeds LoraLoaderModelOnly 116 (wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise, strength 1), then ModelSamplingSD3 93 (shift 8), then KSamplerAdvanced 96: add_noise enable, seed 221824495232956 set to randomize, 4 steps, CFG 1, euler, simple, steps 0 to 2, leftover noise enable. - Low noise: UNETLoader 102 feeds LoraLoaderModelOnly 117 (the _low_noise LoRA, strength 1), then ModelSamplingSD3 94 (shift 8), then KSamplerAdvanced 95: no added noise, seed 0 fixed, 4 steps, CFG 1, euler, simple, steps 2 to 4. Sampler 95 takes its latent from 96. - Both samplers take positive and negative from WanFunInpaintToVideo 111, and sampler 96 takes its starting latent from it too (links 193 to 197). - Output: VAEDecode 97, then CreateVideo 100 at 16 fps, then SaveVideo 158 (prefix video/ComfyUI, format auto, codec auto). Expect one 640x640 video of 81 frames, about 5 s, running from the start image toward the end image. That is the template's purpose; no output was produced here. 6. Bypassed route: 20 steps, no LoRA - Models: UNETLoader 138 (high) feeds ModelSamplingSD3 146 (shift 8) and UNETLoader 139 (low) feeds ModelSamplingSD3 144 (shift 8). No LoRA loader is in this route. - Samplers: KSamplerAdvanced 149 runs 20 steps at CFG 3.5, euler, simple, steps 0 to 10 with leftover noise, seed 247225372043700 set to randomize. KSamplerAdvanced 150 runs steps 10 to the end without added noise. - Own inputs: CLIPLoader 136, prompts 141 and 142 (the same text as the active group), VAELoader 137, LoadImage 140 and 147, and WanFunInpaintToVideo 148 (640x640, length 81). Output goes through VAEDecode 151 and CreateVideo 152 (16 fps) to SaveVideo 153, with the same prefix video/ComfyUI. 7. Run it 1) Put the four base files in the folders above, plus both LoRAs for the route that is active as saved. 2) Open the template. 3) Load your first frame into LoadImage 110 and your last frame into LoadImage 112. 4) Edit prompts 99 and 91, and size and length on node 111. 5) Run. Expect one video under video/ComfyUI. For the 20-step route: select the bypassed group's nodes and press Ctrl + B to enable them, select the LoRA group and press Ctrl + B to bypass it, load your frames into LoadImage 140 and 147, edit prompts 141 and 142 and node 148, and run. Skip the two LoRA files if you will only use this route. 8. Failure cases and checks - Both groups enabled: both pipelines run through their own loaders, so the time of both routes is spent and two videos are written under the same prefix. - Edits that do nothing: the groups do not share prompts, images or size. Edit the nodes in the group you run. - The LoRA files are required by the route that is active as saved, which is the opposite of the Fun Control and first-last-frame templates. The Fun Inpaint experts are required by both routes. They are not the i2v, Fun Control or Fun Camera experts. - File-list note: the folder tree in note 155 lists the low-noise LoRA twice and omits the high-noise LoRA. The loaders (116, 117) and the file list above name both files, so use those. - The Tutorial link in note 155 has a line break before its closing parenthesis, so the link may not resolve as written. It was not fetched. - Lightning trade-off: note 159 says the 4-step LoRAs reduce time but lose video dynamics. - Frame and size handling: see sections 3 and 4. 9. How it differs from the other Wan 2.2 14B templates It takes the same two images as the first-last-frame template, but through WanFunInpaintToVideo with the Fun Inpaint expert pair instead of WanFirstLastFrameToVideo with the i2v pair. First-last-frame uses CFG 4 in its 20-step default and shift 5 in its LoRA group; this file uses CFG 3.5 and shift 8 in both. It follows the two-group bypass pattern of the Fun Control template, with the groups' roles reversed. The published T2V guide and the LoRA switch comparison describe one switch inside a subgraph instead. 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; that covers this analysis of the template, not the Wan models, LoRAs, example images or outputs, which keep their own terms.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/video_wan2_2_14B_fun_inpaint.json",
      "note": "Exact pinned template; node, link and group byte ranges and excerpt hashes are in the source pins."
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/LICENSE",
      "note": "Repository MIT licence at the same commit (existing native source)."
    }
  ]
}