Empty Latent Image Plus

Empty Latent Image Plus

Version 1.1.1 fixes BOM preset loading, prevents changes after preset read errors, and preserves saved dimensions when Clamp on Load is off.

Version 1.1.0 added Auto and manual latent formats. The default remains the original SDXL / SD1.5 format so existing workflows continue to load. <img width="1110" height="867" alt="ELIP" src="https://github.com/user-attachments/assets/19e1d176-1dae-439e-a61f-55a4a2f4a20c" />

Empty Latent Image Plus is a compact ComfyUI custom node based on the core Empty Latent Image node.

It creates an empty latent image and also outputs the current WIDTH and HEIGHT as INT values. This is useful when another node or workflow needs the same canvas size as separate width and height inputs.

Installation

Clone this repository into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/Empty-Latent-Image-Plus.git

Then restart ComfyUI.

Main Features

• Creates a standard ComfyUI latent image.
• Outputs LATENT, WIDTH, and HEIGHT.
• Keeps width, height, and batch_size simple and visible.
• Includes a swap W/H toggle that swaps width and height.
• Includes foldable User Preset controls.
• Includes foldable Advanced controls.
• Stores user presets in presets.json.
• Does not include fixed model-specific presets.
• Accepts an optional MODEL input for Auto detection and a latent_type dropdown for manual selection.

Typical Use

Use this node as a drop-in replacement for the core Empty Latent Image node when you also want WIDTH and HEIGHT outputs.

For example, it can be used with workflows that need one shared size value for:

• latent image generation
• canvas or framing helper nodes
• prompt or layout helper nodes
• nodes that accept INT width and height inputs

Presets store canvas dimensions only; selecting or loading a preset does not change the model connection or latent format.

Presets retain integer dimensions without rounding to a multiple of eight. Clamp on Load rounds only when enabled, using Size Step; disabling it loads the saved dimensions unchanged. Model-specific alignment during generation is separate and does not rewrite the preset.

UTF-8 presets with or without a BOM are supported. Invalid JSON or a read error blocks preset changes and displays an error instead of overwriting the file.

Intended Compatibility

Branch an upstream loader (or LoRA loader) MODEL output into this node. A connected model always takes priority and switches the dropdown to Auto. Disconnecting restores the last manual choice. Explicitly choosing Auto without a model reports an actionable error rather than guessing from a filename.

If a downstream ModelSampling or layout node uses this node's WIDTH / HEIGHT, connect the MODEL before that downstream node, not its output. Feeding its output back into this node creates a graph cycle. Manual selection remains available when an upstream MODEL connection is inconvenient.

| Manual choice | Channels / spatial divisor | Shape | Other examples | |---|---|---|---| | SDXL / SD1.5 | 4 / 8 | B,C,H,W | Pony, AuraFlow | | SD3 / Flux.1 / Lumina2 | 16 / 8 | B,C,H,W | Chroma, Z-Image, HiDream | | Anima / Krea2 / Qwen Image | 16 / 8 | B,C,1,H,W | Original Qwen Image and Edit variants; JoyImage, MingImage via Auto | | Flux.2 / MOSS-Image | 128 / 16 | B,C,H,W | Flux.2 Klein, MOSS Base/SFT | | Qwen Image 2.1 | 64 / 16 | B,C,H,W | Qwen Image 2.1 / Turbo | | Hunyuan Image 2.1 | 64 / 32 | B,C,H,W | Base image model, not the dedicated refiner | | Pixel-space | 3 / 1 | B,C,H,W | Chroma Radiance, pixel-space image models |

Auto reads the host's latent_format and image-model configuration. Quantization, LoRA patches and checkpoint filenames do not determine the geometry. It does not load model weights onto the GPU, change the model or select its VAE.

This is a still-image node. Video (including Wan), audio, 3D objects, Qwen Layered, dedicated refiners and unknown formats should use their dedicated latent nodes. Auto reports unsupported formats rather than silently falling back to 4 channels.

For Auto and non-default manual formats, non-divisible canvas dimensions round down to the spatial divisor (at least one latent cell). WIDTH and HEIGHT report the resulting dimensions. Inputs and saved presets are not rewritten. The unconnected original default path keeps the old clamping, shapes and INT values.

UI Controls

width / height / batch_size

These behave like the core ComfyUI Empty Latent Image node.

swap W/H

swap W/H is a toggle-style control. The default value is false.

Every toggle click swaps the current width and height values, regardless of whether the toggle changes from false → true or true → false. The WIDTH and HEIGHT outputs follow the swapped values.

Example:

Before: width 768, height 1024
After:  width 1024, height 768

batch_size is not changed.

User Preset

User Preset is foldable.

When expanded, it shows:

Preset
Name
Save
Load
Delete

Presets are saved to:

ComfyUI-Empty-Latent-Image-Plus/presets.json

Saved preset data:

{
  "Preset Name": {
    "width": 768,
    "height": 1024,
    "batch_size": 1
  }
}

Advanced

Advanced is foldable.

It contains:

• Size Step
• Clamp on Load
• Show Info
• Info

Size Step controls the rounding step used when loading presets with Clamp on Load enabled.

Clamp on Load keeps loaded preset values inside safe ranges and rounds width and height to the selected step.

Show Info controls the size information line.

The info line shows:

width x height | aspect ratio | megapixels

Installation

Copy this folder:

ComfyUI-Empty-Latent-Image-Plus

into:

ComfyUI/custom_nodes/

Then restart ComfyUI.

Expected node:

Empty Latent Image Plus

Expected category:

latent/plus

Technical Specification

Inputs

width       INT  default 512  min 16  max 16384  step 8
height      INT  default 512  min 16  max 16384  step 8
batch_size  INT  default 1    min 1   max 4096   step 1
latent_type optional combo   default SDXL / SD1.5 (4ch)
model       optional MODEL  connected model always uses Auto

Outputs

LATENT
WIDTH   INT
HEIGHT  INT

Latent Shape

The default backend path creates the latent tensor with:

[batch_size, 4, height // 8, width // 8]

batch_size is clamped to a minimum of 1.

Other shapes follow the compatibility table. Spatial/temporal downscale metadata is included to prevent the sampler from resizing a correctly constructed latent. The new modes use ComfyUI's intermediate device/dtype; the original default keeps its previous CPU/default-dtype behavior.

Existing workflows and presets

The node ID and three outputs are unchanged. Old API prompts that supply only width/height/batch_size remain valid. The frontend migrates old positional widget values in memory, including clipboard and nested subgraph loads, so the added dropdown does not shift the old swap/preset/advanced values. Existing links remain at their original indices. Workflow files are not rewritten by installation.

presets.json remains in its old three-field format and is not replaced during the local update. The preset buttons and width/height swap retain their behavior.

Preset saves and deletes now write a temporary file in the same directory, flush it, and replace presets.json only after the write completes. A write or replace failure keeps the previous file. This protects interrupted writes, alongside the existing BOM support and refusal to overwrite unreadable preset data.

The UI applies the preset list returned by save/delete directly, ignores older pending reads, and disables Save/Delete while a change is in progress on that node. Preset fields, saved selections and workflow migration remain unchanged.

Backend Files

empty_latent_image_plus.py
__init__.py
presets.json

The backend also exposes local preset routes:

GET  /empty_latent_image_plus/presets
POST /empty_latent_image_plus/presets/save
POST /empty_latent_image_plus/presets/delete

Frontend File

web/js/empty_latent_image_plus.js

The frontend adds:

• swap W/H
• foldable User Preset controls
• foldable Advanced controls
• hidden widget layout handling
• size information display

Non-goals

This node intentionally does not include:

• SDXL fixed presets
• SD1.5 fixed presets
• Flux fixed presets
• video latent replacement behavior
• ratio lock enabled by default

The node should remain a small, general-purpose latent helper.

Verification Checklist

• ComfyUI starts without Python import errors.
• The node appears as Empty Latent Image Plus.
• The category is latent/plus.
• Inputs are width, height, and batch_size.
• Outputs are LATENT, WIDTH, and HEIGHT.
• LATENT works like the core Empty Latent Image output.
• WIDTH matches the current width.
• HEIGHT matches the current height.
• batch_size cannot execute below 1.
• swap W/H swaps width and height only.
• User Preset can save, load, and delete presets.
• Advanced can show or hide info controls.
• No fixed model-specific presets are shown by default.

Local CPU verification (2026-10-25)

Tests cover Auto using the host's real latent formats and source model mappings, manual shapes, old API calls, native Flux.2/SD3 tensor equivalence, sampler sizing, unsupported models, preset route round-trips and frontend workflow migration, connections, serialization and preset operations. They do not load real model weights or run image generation. Restart, actual canvas display and GPU generation remain user checks. Tests use existing Python / Node dependencies only.

Run python tests/test_latent_formats.py from an installation inside ComfyUI's custom_nodes, or set COMFYUI_ROOT to a separate ComfyUI checkout. Run frontend tests with node --test tests/test_ui.mjs. Optionally set ELIP_LEGACY_WORKFLOW to an existing workflow JSON for the additional read-only migration check.

The folding controls preserve original widget sizing when reopened. Advanced and User Preset repeated open/close operations and Show Info persistence have regression coverage. Actual canvas verification and GPU image generation remain unverified.