ComfyUI : intégrez vos objets 3D avec la bonne perspective ! NKD fSpy Camera

Tech summary (ComfyUI custom nodes for correct 3D insertion)

Goal

Place a 3D object into an existing image (e.g., an empty room) with correct perspective, then generate outputs to support accurate compositing (shadows, masks, reflections, relighting, etc.).

Core idea

Use a set of custom ComfyUI nodes from “NKD FX Tools” (shown as “Cofi UI NKD FX Tools” in the subtitles) to:

• Extract camera / perspective information from a reference image.
• Predict depth for the scene.
• Render/preview the 3D object correctly in the image context.
• Offer adjustable parameters for position, rotation, scale, lighting, occlusion, and focus.
• Output multiple image buffers useful for compositing.

---

Custom nodes highlighted

NKDFX Tools includes many nodes beyond the intro setup, such as:

• Relighting
• 3D preview
• Blur based on depth
• Perspective unwrap / rewrap
• Lens distortion
• Schematic/segmentation masks (“sced masks” in subtitles)

Main nodes used in the workflow

• NKD fSpy Camera: retrieves shooting angle / camera perspective from the input image.
• NKD Preview 3D: generates the object-in-image preview and related outputs.
• Auxil PR Processor (Depth / “deing V2”): computes a depth map from the background image (example resolution: 1024).
• A 3D loading node (the speaker prefers “3D and animation rather than advanced”) to load models outside a fixed folder (example shown: a model generated from Tripos Splat → model file).

---

Workflow steps (as described)

1. Input image

   • Start from a background like an empty room.

2. Retrieve camera perspective using NKD fSpy Viewer

   • Align the green ground, the horizon, and the two horizontal axes.
   • Use a magnifier to place/focus on lines precisely.
   • Save and close to store camera info.

3. Compute scene depth

   • Use Auxil PR processor (Depth V2) with an example resolution of 1024.
   • Connect the background image so the node outputs depth.

4. Preview and insert the 3D model

   • Use NKD Preview 3D:
     • Connect camera info, background image, and depth (so the system understands spatial context).
     • Provide the 3D model into the node’s 3D model inputs.
   • Outputs mentioned:
     • Final preview with only the object.
     • Preview mask for the object.
     • Depth visualization/output.

5. Adjust the object placement

   • Control position / rotation / scale.
   • Modify lighting:
     • direction, shadow toggle, hardness, strength, and contact parameters.

6. Optional realism aids

   • Occlusion based on image depth (for embedding into walls/ground).
   • Adjust focus.

7. Run

   • Produces a set of images intended for compositing workflows.

---

Compositing / model comparisons (Flux / Quen Edit)

After object placement and obtaining outputs (including depth-related data), the video compares how different edit/generation models perform when compositing:

• Editing models mentioned:
  • Flux 2
  • Quin Edit

Variants tested

• Flux 2: Kine 9B / 4B / dev
• Quen: Quen image / Quen Edit 2511

Key observations

• Flux 2 with only the composite image (no depth map)
  • Results look good; sometimes depth wasn’t necessary.
  • Minor loss of detail and slight discrepancies vs reference.
• Flux 2 4B
  • Performed surprisingly well on smaller configurations.
  • Reported as more precise with less loss than the 9B result.
• Flux 2 dev (heavier model)
  • Adds more detail and is very close to reference.
• Quen Edit 2511
  • Also excellent; speaker notes colors could be adjusted a bit.

---

Installation note

Install the custom nodes via the ComfyUI manager by searching NKDFX Tools, or manually:

• Clone into: ComfyUI/custom_nodes
• Restart ComfyUI

---

Main speaker / source

• “N codificator” (speaker credits this author for the custom nodes).