Phr00t's Rapid AIO Wan model, how to configure it for Image to Video?

Beginners
1

I am using the WAN Rapid AIO from Phr00t/WAN2.2-14B-Rapid-AllInOne at main and can not figure out how to configure it for image to video.

Has anyone done this?

1 Like
2

It seems you can modify and use this workflow file.

Use Rapid AIO Mega-v12 with the Mega v3 workflow. Load the v12 checkpoint, feed your still into Start Frame, bypass End Frame, keep the VACE block active, run Steps=4 and CFG=1, sampler euler_a with beta scheduler, set Number of Frames, decode with VAEDecode (non-tiled), then combine frames to video. (Hugging Face)

Why this works

  • Mega does I2V, T2V, and first↔last. Its model card states: “MEGA I2V: bypass the end frame so the start frame is your I2V starting frame.” One AIO .safetensors includes model + CLIP + VAE; use a single Load Checkpoint. All AIO variants are intended for CFG=1 and 4 steps. For Mega-v12 the author recommends euler_a/beta. (Hugging Face)
  • The included Mega-v3 workflow JSON exposes nodes labeled Start Frame (Optional) and End Frame (Optional), a VACE node titled “Bypass for T2V, use for I2V”, a Number of Frames control, and a VAEDecode node. It also documents T2V=Strength 0, I2V=Strength 1 on the VACE block. (Hugging Face)
  • The Mega-v12 folder contains a marker file .use_mega_v3_workflow that tells you to load the v3 workflow with v12. (Hugging Face)

Step-by-step (beginner-safe)

1) Put the model in the right place

  • Copy wan2.2-rapid-mega-aio-v12.safetensors into:

    ComfyUI/models/checkpoints/

  • In ComfyUI, add Load Checkpoint and select that file. You do not need separate VAE or CLIP with AIO. Start with CFG=1 and Steps=4 as the default recipe. (Hugging Face)

2) Load the correct workflow

  • Open the repo’s Mega-v3/Rapid-AIO-Mega.json (drag-and-drop into ComfyUI). v12 explicitly signals to use the Mega-v3 workflow. (Hugging Face)

3) Wire for Image→Video

  • Provide Start Frame: Load your still image into the node titled Start Frame (Optional). (Hugging Face)
  • Bypass End Frame: Disable or disconnect the End Frame (Optional) branch so it does not feed the graph. This is the I2V switch. (Hugging Face)
  • Keep VACE enabled: The node titled “Bypass for T2V, use for I2V” should remain in the path for I2V. Its helper text shows I2V strength = 1. (Hugging Face)

4) Minimal settings that work

  • Sampler: euler_a sampler with beta scheduler (author’s pick for Mega-v12). (Hugging Face)
  • CFG: 1. Steps: 4. Leave negative prompt blank; the JSON even names the node “Negative Prompt (leave blank cuz 1 CFG)”. (Hugging Face)
  • Frames: Set the Number of Frames node (e.g., 65 or 81) to control clip length. (Hugging Face)
  • Preview vs export: The model card notes you can swap the preview for VideoCombine to render a file. (Hugging Face)

5) Run and export

  • Run the graph. If you used Preview Image, you’ll see frames; add VideoCombine to output a video file at your target fps. (Hugging Face)

Sanity test preset

  • Start Frame: your still image
  • Sampler: euler_a + beta
  • CFG: 1
  • Steps: 4
  • Number of Frames: 65
  • Decode: VAEDecode (non-tiled)
  • Export: VideoCombine at ~24 fps
    This mirrors the author’s intended “short sampler” workflow. (Hugging Face)

Common pitfalls and fixes

  • Flicker or color “blinking”: Known issue if you decode with VAE Decode (Tiled). Use the standard VAEDecode node instead; the ComfyUI issue shows tiled decode causes Wan video blinking while non-tiled is fine. (GitHub)
  • Missing nodes or templates: Update ComfyUI to the latest version. Official docs say to use Workflow → Browse Templates → Video; if a template is missing or nodes are red, you’re likely not on the latest nightly. (ComfyUI)
  • Wrong workflow for v12: Mega-v12 ships a .use_mega_v3_workflow marker. If your graph doesn’t match screenshots, load Mega-v3. (Hugging Face)

Extra context and alternates

  • What “bypass End Frame” means: In ComfyUI you disconnect or disable the End Frame path so only the Start Frame image conditions the video. The VACE node stays active for I2V. The JSON labels make this explicit. (Hugging Face)

  • Official ComfyUI I2V baseline (non-AIO): If you want a plain reference graph, load Wan2.2 14B I2V from ComfyUI’s template browser. It shows a standard setup using separate diffusion, VAE, and text encoder files. Good for debugging outside AIO. (ComfyUI)

  • CLI semantics if you compare against the repo: For Wan2.2 I2V, --size is an area hint and the aspect ratio follows your input image. The repo provides working i2v-A14B commands and documents this rule. (GitHub)

  • Rules recap from the model card:

    • Mega supports I2V/T2V/First↔Last, switched by bypassing Start/End branches as described.
    • AIO = one checkpoint with VAE + CLIP + model.
    • Defaults: CFG=1, Steps=4.
    • Mega-v12: euler_a/beta recommended. (Hugging Face)

Sources you can rely on

  • Rapid AIO model card: capabilities, I2V wiring, sampler and defaults, and VideoCombine note. (Hugging Face)
  • Mega-v3 workflow JSON: Start/End frame nodes, VACE block “Bypass for T2V, use for I2V”, Number of Frames, VAEDecode, and strength hints. Drag-and-dropable. (Hugging Face)
  • Mega-v12 marker: .use_mega_v3_workflow confirms which workflow to load. (Hugging Face)
  • ComfyUI official Wan 2.2 guide: templates, update notes, and stock I2V workflow. (ComfyUI)
  • ComfyUI issue on tiled VAE flicker: explains why non-tiled decode avoids blinking. (GitHub)
3

Thank you. I have seen some improvement. It is still super experimental though.

1 Like