Setup instructions for customized modules

This pipeline uses several modules that are not included in a standard ComfyUI installation. Before loading the workflow JSON, you need to install the relevant custom node packs via ComfyUI Manager, download the required model weights from HuggingFace, and prepare your own character reference images and a 3D humanoid mesh. The tables below list everything you need.

NON-STANDARD NODES — WHAT TO INSTALL

Node(s) in JSON	Package	Install via
CLIPTextEncodeFlux	Built into ComfyUI core since late 2024	Update ComfyUI to the current version — no extra install if up to date
ControlNetApply	ComfyUI-Advanced-ControlNet	" data-it="ComfyUI Manager → cerca "Advanced ControlNet"">ComfyUI Manager → search "Advanced ControlNet"
ApplyPuLIDFlux	ComfyUI-PuLID-Flux	" data-it="ComfyUI Manager → cerca "PuLID Flux"">ComfyUI Manager → search "PuLID Flux"
IPAdapterModelLoader IPAdapterApply	ComfyUI_IPAdapter_plus (cubiq)	" data-it="ComfyUI Manager → cerca "IPAdapter Plus"">ComfyUI Manager → search "IPAdapter Plus"
CLIPVisionLoader	Built into ComfyUI core	No install needed
Load3DMesh 3DPoseEditor OrbitCamera Render3DMeshToDepth	ComfyUI-3D-Pack	or "OrbitCamera"" data-it="ComfyUI Manager → cerca "3D Pack" o "OrbitCamera"">ComfyUI Manager → search "3D Pack" or "OrbitCamera"

MODEL FILES REQUIRED

All files go inside the ComfyUI models/ folder, in the indicated subfolder.

File in JSON	Download from	Subfolder
flux1-dev.safetensors	black-forest-labs/FLUX.1-dev ~24 GB — requires licence acceptance	models/unet/
t5xxl_fp16.safetensors	comfyanonymous/flux_text_encoders	models/clip/
clip_l.safetensors	comfyanonymous/flux_text_encoders	models/clip/
ae.safetensors	black-forest-labs/FLUX.1-dev same repo as flux1-dev	models/vae/
flux-depth-controlnet.safetensors	Shakker-Labs/FLUX.1-dev-ControlNet-Depth	models/controlnet/
ip-adapter_flux.safetensors	XLabs-AI/flux-ip-adapter	models/ipadapter/
sigclip_vision_384.safetensors	Comfy-Org/sigclip_vision_384	models/clip_vision/

FILES YOU MUST PROVIDE YOURSELF

The pipeline uses a 3D humanoid mesh to generate a per-shot depth map that ControlNet uses as spatial guidance (body position, scale, rough pose within the frame). Without the mesh file, the Render3DMeshToDepth nodes produce nothing and ControlNet depth guidance is silently disabled. The mesh does not need to be detailed: a generic neutral T-pose figure is sufficient. Recommended free sources: MakeHuman (export directly as OBJ, simplest option); Mixamo (T-pose FBX, then export as OBJ from Blender); or any neutral-pose humanoid OBJ from CGTrader or Sketchfab. Place the file at ComfyUI/input/generic_humanoid.obj — the Load3DMesh node reads from the input/ folder, the same folder used by LoadImage.

File in JSON	What it is	Notes	Preview
jim_turnaround.jpg	Full-body reference image of character Jim	Used by IP-Adapter for clothing/posture
barbara_turnaround.jpg	Full-body reference image of character Barbara	Used by IP-Adapter
jim_face_crop.png	Tight face crop of Jim (~512×512 px)	Used by PuLID for facial geometry
barbara_face_crop.png	Tight face crop of Barbara (~512×512 px)	Used by PuLID for facial geometry
jim_barbara_ref.jpg	Two-character reference — Jim and Barbara together	Visual reference for dual-character shots (shots 2 and 5)
generic_humanoid.obj	A 3D humanoid mesh in neutral T-pose	Used to generate per-shot depth maps for ControlNet. See the note above for recommended sources and where to place the file.

RECOMMENDED RUN SEQUENCE

1. Update ComfyUI to the current version before anything else (git pull in the ComfyUI folder).
2. Install all custom node packs via ComfyUI Manager. Restart ComfyUI after.
3. Download all model files and place them in the correct models/ subfolders (see table above).
4. Prepare your reference images — full-body turnarounds and tight face crops for both characters.
5. Export or download a neutral T-pose humanoid OBJ and place it at ComfyUI/input/generic_humanoid.obj.
6. Fix the class name if needed: verify that all six ControlNet nodes in the JSON read "class_type": "ControlNetApply". The JSON provided here already has this fix applied.
7. Load the JSON in ComfyUI via Load → workflow_api_v2.json.
8. Update the four LoadImage nodes with your actual character image files.
9. Update the Load3DMesh node with the path to your .obj file if the filename differs.
10. Queue — each of the 6 shots runs sequentially. At 30 steps on a 3090 expect roughly 3–5 minutes per shot.