AbstractPhil/mega-liminal-lora

Mega Liminal LoRA

A LoRA for liminal spaces: empty malls, fog-bound roads, parking garages, suburbs at night, vacant theatres and hallways. It was trained on the 1,873 curated and captioned images of the mega-liminal dataset.

NON-COMMERCIAL. These files are derived from Anima and inherit its license: the CircleStone Labs Non-Commercial License, plus the NVIDIA Open Model License of the Cosmos-Predict2 weights underneath it. No commercial use.

Status: training (started 2026-09-26 00:14 UTC). A new file lands every 2 epochs. Sample images come when the run finishes.

Files

The folders are <base model>/<LoRA size>/.

Path	What it is
anima/rank64/mega-liminal-anima-r64-eNN.safetensors	the LoRA after epoch NN (02, 04, ... 16; e16 is the final one), ComfyUI format
anima/rank64/adapter_config.json	the PEFT adapter config
anima/rank64/training/	the training configs and the TensorBoard log

Using it in ComfyUI

1. Get the Anima base files from circlestone-labs/Anima split_files/: anima-base-v1.0.safetensors (diffusion_models), qwen_3_06b_base.safetensors (text_encoders) and qwen_image_vae.safetensors (vae).
2. Put the LoRA file in models/loras and load it with LoraLoaderModelOnly at strength 1.0.
3. Sample around 1 megapixel (1024x1024, or anything from 1:2 to 2:1) with 30-50 steps and CFG 4-5 (er_sde or euler_a), the settings the Anima card recommends for the base model.

Prompting

Every training caption has the same shape:

liminal, <class>, <two to four plain sentences describing the image>

The classes are: mega liminal, landscape, suburban, simulacrum, vanishing point, empty mall, hand sourced, parking garage, cityscape, movie theatre.

The sentences name the kind of image, the place and its layout, the camera position, the architecture and materials, and the lighting and colors. Prompts written the same way work best. Two captions from the training set:

liminal, mega liminal, A photograph of a snowy landscape dominated by dense fog that obscures most of the scene. In the middle ground, a blue road sign stands on two thin metal poles, its text illegible due to the heavy mist. The foreground and background merge into a uniform gray-white expanse with no visible terrain features, trees, or sky. Lighting is flat and diffused, typical of overcast winter weather, with no discernible shadows.

liminal, parking garage, Aerial photograph of a large, empty parking lot with neatly arranged rows of marked spaces. The asphalt surface is dark and smooth, marked with white lines and yellow curbs. Tall streetlights are evenly spaced throughout the lot. In the background, a modern building with a flat roof and large windows is visible.

Training

Setting	Value
Base model	Anima base v1.0 (the version the Anima card names for LoRA training)
Data	1,873 images in 10 classes; small classes repeated up to 8x, so one epoch is 2,789 samples
Captions	written by Qwen3.5-9B from each image, one structured pass per image
Resolution	1024, aspect-ratio buckets from 1:2 to 2:1
LoRA	rank 64
Optimizer	Adam, learning rate 2e-5, betas 0.9 / 0.99, no weight decay, 100 warmup steps, gradient clip 1.0
Batch	8 (4 per GPU on 2x RTX 5090)
Length	16 epochs, about 5,580 steps
Trainer	diffusion-pipe (AbstractEyes fork), DeepSpeed data parallel

Images under 512x512 pixels, game screenshots, low-poly and cartoon renders and pixel art were removed before training: 202 of the 2,075 source images.

Dataset

The images and captions are at AbstractPhil/mega-liminal, laid out for any trainer that reads image and .txt caption pairs. Train your own liminal LoRA on it.