Wan 2.2 14B first-last-frame ComfyUI template (video_wan2_2_14B_flf2v): start and end image wiring, 20-step default and a bypassed 4-step LoRA group

Wan 2.2 14B first-last-frame (FLF2V) in ComfyUI: the official template Scope. This guide reads the official video_wan2_2_14B_flf2v.json in Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2 (SHA256 0325fe78...fe2b, 59,591 bytes). All 40 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 WanFirstLastFrameToVideo. All nodes are comfy-core, with no custom node packs. Most were saved at version 0.3.45, the two WanFirstLastFrameToVideo nodes at 0.3.48 and the two LoRA loaders at 0.3.49. Those are saved version strings, not a tested minimum. If that node is missing, update ComfyUI. - Files, as declared by the loaders and the Model links note (node 94, sizes as the note states them): - models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors (6.27 GB), from Comfy-Org/Wan_2.1_ComfyUI_repackaged. - models/diffusion_models/wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors and wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors (13.31 GB each), from Comfy-Org/Wan_2.2_ComfyUI_Repackaged. - models/vae/wan_2.1_vae.safetensors (242.1 MB). - Only for the 4-step group: models/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors and _low_noise.safetensors (1.14 GB each). There is no separate FLF model: first-last-frame reuses the 14B I2V expert pair. The default route needs about 33 GB of files, and the LoRAs add about 2.3 GB. No CLIP vision model is used: the clip_vision_start_image and clip_vision_end_image inputs of WanFirstLastFrameToVideo are left unconnected. 2. The two groups The file holds two complete copies of the pipeline. Each has its own loaders, prompts, images, sampler pair and SaveVideo: - "Default workflow" (active as saved): fp8_scaled without LoRA. - "Enable lightning lora version (Can Speed up the workflow but will sacrifice some quality)": every node in it is bypassed (mode 4) as saved. Notes 96 and 95 say the template "provides two workflows" and you "enable one as needed": box-select a group and press Ctrl + B, and disable the other group so both do not run. Note 95's wording describes disabling the LoRA group "after enabling the fp8_scaled group". As saved, the opposite is true: the fp8_scaled group is already active and the LoRA group is bypassed. 3. Default route (as saved) - Images: LoadImage 80 (example video_wan2_2_14B_flf2v_start_image.png) goes to start_image, and LoadImage 89 (example ..._end_image.png) goes to end_image, both on WanFirstLastFrameToVideo 81. They sit in the groups "Step 2.1 - Upload start_image" and "Step 2.2 - Upload end_image". - Prompts: positive CLIPTextEncode 90 and negative 78 both use CLIPLoader 72 (umt5_xxl_fp8_e4m3fn_scaled, type wan). The saved positive prompt is Chinese; in the lane's translation it describes an ice-crystal kitten waking and turning into a large furred beast. The negative is the usual Wan Chinese quality list. - Size and length: WanFirstLastFrameToVideo 81 holds width 640, height 640, length 81 and batch 1. The Video Size note (88) says the small size is for low VRAM and can be raised. - Sampling: UNETLoader 76 (high noise) feeds ModelSamplingSD3 73 (shift 8, saved as 8.000000000000002), then KSamplerAdvanced 84: add_noise enable, seed 984937593540091 set to randomize, 20 steps, CFG 4, euler, simple, steps 0 to 10, return_with_leftover_noise enable. UNETLoader 77 (low noise) feeds ModelSamplingSD3 74 (shift 8), then KSamplerAdvanced 87: add_noise disable, seed 0 fixed, 20 steps, CFG 4, euler, simple, steps 10 to 10000 (the end), leftover noise disable. Sampler 87 takes its latent from 84. Both take positive, negative and latent from node 81. - Output: VAEDecode 85 (wan_2.1_vae, VAELoader 79), then CreateVideo 86 at 16 fps, then SaveVideo 83 (prefix video/ComfyUI, format auto, codec auto). With 81 frames at 16 fps, that is about 5 seconds if all 81 frames are written; the frame count was not checked by running it. 4. 4-step LoRA route (bypassed as saved) - Models: UNETLoader 37 (high) feeds LoraLoaderModelOnly 91 (wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise, strength 1), then ModelSamplingSD3 54 (shift 5). UNETLoader 56 (low) feeds LoraLoaderModelOnly 92 (the _low_noise LoRA, strength 1), then ModelSamplingSD3 55 (shift 5). - Samplers: KSamplerAdvanced 57 runs 4 steps at CFG 1, euler, simple, steps 0 to 2 with leftover noise. KSamplerAdvanced 58 runs steps 2 to the end, without noise, seed 0 fixed. - Own inputs: CLIPLoader 38, prompts 6 and 7, VAELoader 39 and WanFirstLastFrameToVideo 67 (640x640, length 81) are separate from the default group. Editing the default prompt, images or size does not change this route. - Swapped example images: in this group, start_image comes from LoadImage 68, saved with video_wan2_2_14B_flf2v_end_image.png, and end_image comes from LoadImage 62, saved with ..._start_image.png. The group labels follow the wiring (node 68 is under "Step 2.1 - Upload start_image"), but the saved file names are reversed relative to the default group. Re-upload both images before running it. - Different example prompt: positive prompt 6 describes a bearded man in a Van Gogh self-portrait style transforming into an astronaut, not the default kitten prompt. The example image contents were not inspected, so which prompt matches them is not established. - Output: VAEDecode 8, then CreateVideo 60 at 16 fps, then SaveVideo 61, with the same prefix video/ComfyUI as the default group. 5. Run it 1) Put the four default files in the folders above. Add the two LoRAs only if you will use the 4-step route. 2) Open the template. To use the default route, keep the "Default workflow" group active. 3) Upload your first frame to LoadImage 80 and your last frame to LoadImage 89. 4) Edit positive prompt 90 and negative 78, and change width, height and length on node 81 if VRAM allows. 5) Run. You should get one video under video/ComfyUI that goes from the start image to the end image. For the 4-step route: select the LoRA group's nodes and press Ctrl + B to enable them, then select the Default workflow group and press Ctrl + B to bypass it. Upload your first frame to LoadImage 68 and your last frame to LoadImage 62, edit prompts 6 and 7, and run. 6. Failure cases and checks - Both groups enabled: both pipelines run, each through its own loaders. The run takes the time of both routes and writes two videos under the same prefix. Note 95 warns about this and suggests partial execution as an alternative. - Reversed motion in the 4-step route: if you enable it without re-uploading, the example end image becomes the first frame. - Edits that do nothing: the two routes do not share prompts, images or size. Edit the nodes in the group you run. - Missing LoRA files only matter for the 4-step route. Missing I2V expert files stop both routes. The template loads i2v, not t2v, experts. - Image shape: how WanFirstLastFrameToVideo fits images that do not match width and height is not in this file. Prepare both frames at the output aspect ratio. - Quality and speed trade-off: note 96 says the Lightning LoRA reduces generation time but loses video dynamics. - Saved versions: if loading fails on node 81, update ComfyUI. 0.3.48 is the saved version, not a known minimum. 7. Author figures (not tested here) Note 93 states, for an RTX 4090D 24 GB at 640x640: fp8_scaled 84% VRAM, about 536 s first and 513 s second generation; fp8_scaled + 4steps LoRA 89% VRAM, about 108 s and 71 s. These are the template author's figures, not a measurement by this lane. 8. How it differs from the other 14B templates At the same commit, the published Wan 2.2 14B T2V guide and the LoRA switch comparison, which also covers the I2V template, describe a single Enable Lightning LoRA switch inside a subgraph. FLF2V instead duplicates the whole pipeline into two bypass-toggled groups. Its default CFG is 4, where the T2V template uses 3.5. It uses the lightx2v 4-step v1 i2v LoRAs at shift 5. The step split stays half and half: 10 + 10 by default, and 2 + 2 with the LoRAs. 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_flf2v.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)."
    }
  ]
}