Latent Rotate (Portrait / Landscape) — ComfyUI Custom Node

A minimal, opinionated ComfyUI custom node that emits portrait (0°) and/or landscape (90°) latent variants in a single run.

This node is designed for deterministic orientation comparison and geometry-aware workflows, without redundant rotations, filename hacks, or UI friction.

---

Why this node exists

Diffusion models operate on a grid-aligned latent space.
Only two rotations preserve that structure meaningfully:

• 0° — native orientation (portrait)
• 90° — orthogonal orientation (landscape)

Rotations of 180° and 270° do not add new structural information and only increase compute cost and dataset noise.
This node intentionally excludes them.

This node was created to remove repetitive manual rotation steps in orientation-sensitive workflows.

---

Features

• Emits 0°, 90°, or both in a single execution
• Uses latent lists (no queue tricks, no batch hacks)
• Deterministic output ordering
• Rotation applied before sampling
• Metadata passthrough (rotation, orientation)
• Compatible with all samplers, VAEs, and SaveImage
• No filename manipulation or suffix logic
• Clean, predictable graph semantics

---

Node Inputs

latent

Input latent.

mode

Select which orientations to emit:

• Portrait Only → 0°
• Landscape Only → 90°
• Both → 0° and 90°

---

Outputs

LATENT (list)

Depending on mode:

• Portrait Only → [latent_0deg]
• Landscape Only → [latent_90deg]
• Both → [latent_0deg, latent_90deg]

The output is always a list, even when it contains a single latent.
This is intentional and keeps downstream behavior stable.

---

Usage

Typical wiring:

Empty Latent ↓ Latent Rotate (Portrait / Landscape) ↓ KSampler ↓ VAE Decode ↓ SaveImage

yaml Copy code

ComfyUI automatically iterates over the latent list and generates one image per orientation.

---

Notes on determinism

For meaningful orientation comparison:

• Use a fixed seed in the KSampler
• Keep sampler, steps, CFG, and model settings identical

The node prints a reminder in the console to reduce accidental randomness.

---

Why filenames are not modified

This node deliberately does not modify filenames or emit suffixes.

Rationale:

• Orientation is visually obvious
• Filename logic is often already complex (timestamps, experiment IDs, pipelines)
• SaveImage works best as a simple sink

If filename tagging is required, it can be handled upstream using standard STRING nodes.

---

FAQ

Why does the output port look dotted or stacked?

The dotted icon indicates that the node outputs a list of latents rather than a single latent.
This is normal behavior in ComfyUI and allows downstream nodes to process each orientation automatically.

---

Why not support 180° or 270° rotations?

Those rotations do not introduce new compositional information for diffusion models.
They mainly add redundant computation and clutter datasets without practical benefit.

---

Can this node rotate by arbitrary angles like 45°?

No. Arbitrary-angle rotations break grid alignment in latent space and introduce interpolation artifacts.
This node is intentionally limited to rotations that preserve latent structure.

---

Does this double VRAM usage?

No. The model is loaded once.
Latents are processed sequentially, not duplicated in memory.

---

Design philosophy

This node follows a strict “do one thing well” approach:

• Geometry only
• No UI automation
• No implicit side effects
• No redundant options

The goal is clarity, predictability, and composability within larger ComfyUI graphs.

---

Installation

Clone or copy the folder into:

ComfyUI/custom_nodes/latent-rotate-90/

yaml Copy code

Restart ComfyUI.
The node appears under:

latent → geometry

yaml Copy code

---

License

MIT

---

Acknowledgements

This custom node was developed with assistance from ChatGPT for code generation, debugging, and documentation refinement.
Final design decisions, integration, and validation were performed by the author.