Set up ComfyUI (once)

1. Install ComfyUI

Install ComfyUI v0.39.0 from its release page on github.com/Comfy-Org/ComfyUI (commit b0b743566f65daafc423b4fea8a2fbda94b3384a). On Windows, use the portable build: download ComfyUI_windows_portable_nvidia.7z, check its sha256 (3dca347842ed1c6c833105fafb1669284685a3e54df41578d5f94bcc898ae039), and extract it to a new folder. The MiniMax H3 video nodes are built in; there are no custom nodes to install for these guides.

Version: Use ComfyUI v0.39.0 (commit b0b743566f65), from the Comfy-Org/ComfyUI repository. It is the recommended release, not a tested minimum: our latest test ran v0.39.0 plus 45 later commits (14c55315), and newer releases may work but aren’t tested. Older ones lack nodes the downloads use: the Rebuild a shot control node needs 0.35.0 or later, the Qwen Image 2.1 nodes 0.37.0 or later, and plain 0.37.0 can turn the area you replace solid black. Install that release instead of updating: the update script pulls the newest code, not this release.

2. Download the models

From Hugging Face, from the links below, into the folders under ComfyUI/models/. Each link is pinned to one version of the file; after downloading, check that the file’s sha256 matches the table before you move it into place. Take only these .safetensors files: some of the same repositories hold other files you don’t need. The text encoder and the two VAEs are shared, so a second recipe only adds its own model and LoRA. Allow about 47 GB of disk space for Swap a person, about 68 GB for Add new footage, about 51 GB for the two Rebuild a shot guides, about 76 GB for all four.

Use the Comfy-Org copy of the Fun ControlNet patch, as linked above. The 13.5 GB file from the original authors loads without an error but puts its control layers in the wrong places, and the results come out wrong.

3. Pick the size your GPU fits

Each guide offers two downloads. The default is sized for a 16 GB card; the higher-quality one needs a 48 GB-class GPU or a cloud GPU. The table shows what the defaults took on an RTX 5070 Ti (16 GB). They use nearly all of a 16 GB card. Close Chrome, OBS, games and other heavy apps while it generates.

System memory matters too. It needs 32 GB of system RAM, plus a pagefile of about 32 GB or more (on Windows, a system-managed pagefile is fine if the disk has room). The video model (21 GB) and the text encoder (15.7 GB) do not fit in 16 GB of GPU memory, so ComfyUI keeps part of them in system memory and the pagefile: on the measured runs, system memory peaked at 31.2 of 32 GB (the character swap is the heaviest). With 16 GB of system RAM it is untested and likely to fail.

No suitable GPU? Rent one by the minute, for example an RTX PRO 6000 on RunPod, at about $0.20 per 5-second clip; it runs the higher-quality files too. Install ComfyUI there the same way. Keep pictures you upload to a cloud ComfyUI as JPEGs of about 1.2 MB or less.

4. Load a workflow

Download a workflow from its guide, then in ComfyUI choose File → Open and pick it, or drag the file onto the canvas. It loads as a fully wired graph. Each guide lists the few inputs to change; leave the sampler settings as they are.

Where your files go. Put each still or clip directly in your ComfyUI/input folder (not a subfolder), or use the node’s upload button. The Load Image and Load Video nodes only list files that sit directly in that folder. What ComfyUI makes is saved under ComfyUI/output/; to use it in another workflow, copy it into ComfyUI/input first. Give each file a name of its own, for example myshow-s1e4-start.png, so a new run never overwrites an old one.

5. Pictures

Each recipe starts from pictures: a character design for a swap, a first frame for an added scene. Use an image model that edits a picture from a reference. Three run locally in ComfyUI; pick one. Seedream 4 and Nano Banana also work, through ComfyUI’s partner nodes, as optional paid services. Start from a real still of the episode, not a description alone, so the color, grain and design match the show.

Our latest test used Qwen Image 2.1. Seedream 4 and Nano Banana are optional paid services, never required.

For a creature or object you’ll upgrade, make one picture you’re happy with, then 2–3 more views of the same individual (side, head close-up), edited from that first picture so the design stays the same. Cut each out on plain grey. Its color decides the color of the result, so pick the color in the picture, not in the prompt.

You’ll need to do this by hand

No download can do these, and an AI agent working from the Markdown file stops and asks you:

• Capture: marking the shot in the extension’s Creator tools, playing the episode and recording it with OBS.
• System settings: the GPU driver and the pagefile (on Windows, changing the pagefile needs an administrator).
• Installing software: ComfyUI, FFmpeg, OBS, HandBrake and the extension’s Creator edition.
• Choosing who is recast, approving the new person’s pictures, and every review of a mask or a take.
• Adding the edit pack in the extension and watching it on the episode.