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.

Samples

The same 15 prompts at the first checkpoint, the halfway point and the final checkpoint. Each prompt keeps its seed at every checkpoint, so the three sheets show the same scenes as training goes on.

Epoch 2

Epoch 8

Epoch 16 (final)

Every 2-epoch checkpoint has a sheet and full-size images in anima/rank64/samples/. They were rendered with ComfyUI's own loaders on Anima base v1.0: LoRA strength 1.0, 40 steps, CFG 4.5, er_sde with the simple scheduler, seed 100 plus the prompt number, and Anima's recommended negative prompt. manifest.jsonl in that folder records the settings of every image. Prompts 13 to 15 are deliberately short, to show what the LoRA brings from the class tag alone.

The 15 prompts

1. liminal, empty mall, A photograph of a deserted two-level shopping mall at night, taken from the upper walkway. A dry fountain sits in the center of the atrium under a glass skylight, surrounded by potted palms and white floor tiles. Storefronts behind lowered metal grilles line both levels, lit by cold fluorescent strips. The palette is teal, cream and sodium orange. (1296x816)
2. liminal, parking garage, A photograph taken from the middle of an underground parking garage level with no cars in it. Concrete pillars with yellow and black striped bases recede toward a green exit sign. Fluorescent tubes run in long rows across the low ceiling, and the painted floor shows faded parking lines and oil stains. The light is flat and greenish white. (1296x816)
3. liminal, suburban, A photograph of a quiet suburban cul-de-sac at dusk, taken from the middle of the street. Two-story houses with dark windows face a circle of asphalt, a single streetlight has just turned on, and a basketball hoop stands over one driveway. The sky is a deep blue gradient with a thin band of pink above the rooftops. (1296x816)
4. liminal, movie theatre, A photograph of an empty movie theater auditorium, taken from the back row. Rows of red velvet seats slope down toward a blank white screen, and dim sconces along the side walls throw soft amber light onto patterned carpet. The ceiling is dark, and a thin beam of light crosses the room from the projection booth. (1152x912)
5. liminal, vanishing point, A photograph of a long, narrow hotel corridor with identical doors on both sides, taken at eye level from one end. Patterned burgundy carpet and beige wallpaper run straight to a single window at the far end. Brass wall lamps are spaced evenly between the doors, and the light is warm and dim. (816x1296)
6. liminal, mega liminal, A photograph of an indoor swimming pool with no one in it, lined with small white and pale blue tiles. Arched openings along one wall lead to darker rooms, and daylight falls from a skylight above, making bright rippled patterns on the water. The walls, floor and steps are all tiled the same way. (1296x816)
7. liminal, landscape, A photograph of a two-lane highway running straight across a flat grassy plain under an overcast sky. A lone billboard with a faded blank panel stands on the right shoulder, and power lines follow the road to the horizon. The colors are muted greens and grays. (1296x816)
8. liminal, cityscape, A photograph of a downtown street at four in the morning after rain, taken from the center line. Tall office buildings with a few lit windows line both sides, traffic lights glow red over the empty intersection, and the wet asphalt reflects the signals. The palette is dark blue with red and white highlights. (1152x912)
9. liminal, simulacrum, A 3D render of a large office floor with rows of empty cubicles under a low ceiling of fluorescent panels. The carpet is gray-blue, the partition walls are beige fabric, and the computer monitors on the desks are dark. The view runs diagonally across the room toward a wall of frosted windows. (1296x816)
10. liminal, hand sourced, A photograph of an elementary school hallway during summer break, taken from low to the floor. Pale yellow lockers line one side, a row of windows lets bright afternoon light in on the other, and the waxed linoleum floor reflects it. The classroom doors are closed, and a trash can stands beside a water fountain. (1152x912)
11. liminal, landscape, A photograph of an empty playground on a foggy morning. A red slide and a metal swing set stand on wood chips in the middle ground, and the trees behind them fade into white fog. The light is soft and gray, and the grass is wet. (1296x816)
12. liminal, mega liminal, A photograph of the stairwell of a concrete apartment building, looking down through the gap between the railings. Green handrails spiral down several floors under ceiling lights, and each landing has a small window of frosted glass. The walls are painted in two tones, pale green above and dark green below. (912x1152)
13. liminal, mega liminal, An empty hallway. (1296x816)
14. liminal, landscape, A field at dusk. (1296x816)
15. liminal, empty mall, A photograph. (1024x1024)

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/samples/	15 samples per checkpoint (eNN/), one sheet per checkpoint (sheets/), the prompts and settings
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, 5,536 steps, 3 h 51 min
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.