🪐 👓 Stereoscopic 3D Generator

Stereo 3D is a niche of a niche - anaglyph questions show up maybe six times a year. Be honest about why you'd install this: you own red-cyan glasses, or a VR viewer, or you want a cross-eye image that looks like it's floating.

SaturnStereo3D takes one 2D image and fakes a camera rig out of it, using monocular depth estimation - a grayscale map where brightness encodes distance. The community settled on Depth Anything V2 for that back in 2024 and nothing has displaced it since: V2 Large for ControlNet, V2 Small when you want speed. Small is the weight this node reaches for.

How it actually works

This is DIBR - depth-image-based rendering, the same idea behind every "living photo" parallax tool, minus the animation. Each pixel shifts horizontally by (depth - convergence_plane) * max_shift, where max_shift is depth_intensity / 100 * 3.5% of image width; the image is then warped twice through that field.

Three things fall out of that math:

• Intensity is subtle by design. At the default 25 you shift pixels by under 1% of the width; at 100 you cap at 3.5%. Push past what the depth map supports and you get tearing and smeared edges, not more depth - the wall everyone hits with depth-based parallax.
• Occlusions aren't inpainted. Where the warp reveals pixels that weren't in the source, the node grabs the nearest border pixel - no LaMa, no diffusion fill. Hard vertical edges smear sideways when you crank it.
• Convergence is your focus knob. Depths equal to convergence_plane don't move at all - that's the screen surface. Raise it and more pops out; lower it and more sinks away.

With no depth_map wired in, the node runs Depth Anything V2 Small itself: letterbox to 518×518, ImageNet normalization, p2/p98 normalization back to full resolution. The model is ~94 MB and downloads on first run into ComfyUI/models/depth/depth-anything-v2-small-518.onnx. Predictions are cached in-process, so a repeat run on the same image is fast.

The inputs that matter

image and mode come first. mode picks the format: Cross-Eye, Parallel (Wall-Eyed), Red-Cyan Anaglyph, Side-by-Side (Parallel), Top-Bottom (Over-Under), or a single eye. Cross-eye swaps the panels - the left half is the right eye's view - which is why the wrong choice from that dropdown makes your eyes hurt instead of converge.

Then it's depth_intensity, convergence_plane, depth_gamma (default 0.50; below 1.0 fattens the near plane so more reads as close), and invert_depth if your map is upside down. image_zoom, pan_horizontal and pan_vertical crop into the frame before disparity is computed - how you make a subject sit proud of its background. gap_spacing and outer_border apply only to the side-by-side and over-under layouts - framing, and your "shrink it so crossing your eyes is easier" knob. alignment_dots drops white focus dots above each panel so you can lock in.

Four outputs: stereogram, left_eye, right_eye, and depth_map - the normalized map it used. This isn't an output node, so nothing shows until stereogram hits a Preview or Save Image. The depth_map output is quietly the most useful: a ControlNet-ready depth map for the rest of your graph.

One quirk: preview_zoom has a tooltip about scaling the canvas preview, but nothing in the shipped 2.5.0 code reads it. Moving it won't do anything yet.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/KOFiblto/ComfyUI-SaturnNodes

Or search "SaturnNodes" in ComfyUI Manager, then restart. requirements.txt is just Pillow, piexif and numpy - no install.py, no runtime pip calls.

For this node specifically, add the one it needs:

pip install onnxruntime          # or onnxruntime-gpu for CUDA

Where people get burned

No onnxruntime means no real depth. It's commented out in requirements.txt as optional, and the failure mode is nasty: instead of erroring, the node prints a console notice and generates a pseudo-depth map - a vertical gradient blended with luminance, brighter at the bottom. Your stereogram comes out vaguely three-dimensional and completely wrong, and if you never open the console you'll spend an evening blaming your sliders. If the effect reads as a flat tilt rather than a scene, that's what happened.

Air-gapped installs. The download tries Hugging Face, then a mirror, and raises Failed to download Depth Anything V2 Small ONNX model if both fail. Pre-place the file at ComfyUI/models/depth/depth-anything-v2-small-518.onnx - the >10 MB size check means a truncated file gets fetched again.

Your own depth map must be white = near. That's Depth Anything's convention and what the node assumes; ControlNet preprocessors differ. If the scene comes out inside-out, that's invert_depth, not a broken node.

Depth is computed at 518×518, then warped onto your render. Fine around 1024 pixels; past that the depth edges are softer than your image edges, and no amount of intensity fixes it.

Every tweak is a re-queue. No live preview in this node, so budget a few iterations rather than expecting a slider to update on canvas.