I put 64 of my LoRAs on the same eight prompts. Here's what each one actually does.

Update: seven of the 64 LoRAs have since been withdrawn. They were trained on Google Flow pictures, and Google's terms don't allow that. Their posts are gone; they still appear on the comparison sheets.

Most showcase pictures show a LoRA doing the one thing it was trained to do. That tells you very little about what it will do in your prompt.

So I ran the opposite test. I took eight prompts from my own wildcard files and put 64 of my published SDXL LoRAs on every one of them, one LoRA at a time. Same seeds, same settings, nothing picked. Then I took the 18 LoRAs my library audit found no defects in and stacked every possible pair of them, on two of the prompts.

826 pictures. Here's what they showed, and a link to every post so you can check me.

How it was run

Base: Juggernaut XL v9 · 1024×1024 · Euler a, Karras · 24 steps · CFG 4.5 · negative prompt low quality, malformed, duplicate
One LoRA at a time, at 0.75 on the model and the text encoder, with its trigger word first: trigger, prompt.
A control for every prompt: the same prompt and seed with no LoRA and no trigger. That's the first post in the series, and every other picture can sit next to its no-LoRA twin.
The trigger is the word each LoRA actually learned, read from its own metadata. For eleven of them that isn't the word their card originally listed. Any card still showing the old word gets fixed as its post goes up.
Pairs: the 18 LoRAs with no recorded defect, all 153 pairs, each LoRA at 0.5, both triggers first, on prompts G1 and G4.
Hardware: one RTX 4070 SUPER (12 GB) running ComfyUI. 1 hour 41 minutes, about 7 seconds a picture. I planned the run with Claude, and a Claude Code session on that desktop wrote the render script and ran it.

The eight prompts:

G1 (seed 75001): shamanic mechanical-organic hybrid figure, underground ice city, brass pipes, mycelium, glowing fungi, glowing particles, volumetric light rays, floating spores or dust in the air
G2 (seed 75002): ritual figure, masked, symbolic presence, Velvet Room cabaret stage, velvet curtains, smoke, broken mirrors, bioluminescent highlights on skin, fabric flowing in motion
G3 (seed 75003): androgynous figure, ambiguous and striking presence, rooftop overlooking glowing city, wind and fabric movement, cinematic lens flare, anamorphic light streaks, high fashion editorial styling, magazine quality lighting
G4 (seed 75004): twilight settlement, organic architecture, bioluminescent networks, god rays through smoke or mist, mycelium growing through structures, environment feels alive and responsive
G5 (seed 75005): hodgepodge fringe village, mixed-styles cobbled-together dwellings, varied materials, willing-to-live-on-edge residents, first to take the storm, soft ambient turquoise-blue bioluminescent glow along every shoreline
G6 (seed 75006): polished brass astrolabe, desert ruins, ancient stone and brass artifacts, prismatic light scattering, chromatic aberration, intricate linework, mixed media texture
G7 (seed 75007): tarnished silver censer trailing smoke, gothic atmosphere, cathedral scale, dramatic shadows, warm gold, crimson, deep amber tones
G8 (seed 75008): flower of life pattern overlaid on composition, subterranean cavern, crystalline formations, bioluminescent water, torus field of radiant energy, symbolic composition with hidden patterns

What I found

1. Subject LoRAs don't add. They replace.

Put a LoRA trained on a subject onto a prompt about something else, and you get its subject. On the ice-city prompt, 20 of the 64 replaced the figure: a dog, cats, a rodent, a spider with a glass dome, a small robot creature, a crystal tower, a brass diving-helmet machine, an ice rover with a bearded driver.

That isn't a defect. It's what a subject LoRA is for. But it means one at 0.75 won't politely decorate your scene.

G1, the ice-city figure: no LoRA first, then all 64.

2. The smallest training sets behave worst

Three of my LoRAs were trained on three images each, for 200 epochs, so every picture was seen about 2,000 times. I flagged them in my September audit. Here's what that looks like from the outside:

joort_reef put its coral-and-mushroom creature into four of the eight prompts, whatever they asked for.
jcolored_pencil came out faded and grey, and nearly blank in four of the eight.
janimism wrote script around the astrolabe's rim.

3. Letters turn up where nobody asked for them

On the astrolabe prompt, three LoRAs wrote on the brass: janimism (script around the rim), jsolarum_punk_registers (numerals around the dial) and jsolarum_sigils (a clear capital J). On the village prompt, jsolarum_fungal_caves drew a top-down map and signed it in the corner with a made-up signature.

G6, the brass astrolabe.

4. Some LoRAs draw panels instead of a scene

Five LoRAs split the village prompt into a collage of panels: jrecipe_vault, jtridium_antenna, jsolarum_badges, jsolarum_geometry and jsolarum_faces_and_emotions. Why those five, I don't know yet. My first suspect is sheets and grids in their training sets.

G5, the fringe village.

5. Vehicles go everywhere, except with the LoRA named for them

jsolarum_living_vehicles put a vehicle into four of the eight prompts: the ice rover, a car on the rooftop, a station wagon with a jellyfish on its roof rack, and a glowing houseboat at the village shore. jspace_winnebago_v2 did much the same. The original jspace_winnebago never drew a vehicle at all away from its home subject.

6. Stacking held up

None of the 306 pair pictures broke: no noise, no mush, no black frames. What happens instead is that a subject LoRA wins the pair. Stack anything with jsolarum_living_vehicles or jsolarum_spiders and the ice-city figure becomes a vehicle or a spider. Stack those two together and you get both: a brass spider-walker carrying a snow globe.

Every pair on G1. Each LoRA alone, at 0.75, sits on the red-outlined diagonal.

7. Two came out broken

jcolored_pencil (above), and jason_shape_logic_v2, which came out hazy and washed out on four of the eight.

8. Some of them undressed the figure

Nothing in any prompt asked for nudity, and the no-LoRA versions stayed dressed. Still, on the cabaret prompt 12 of the 64 LoRAs came out nude or topless, and on the rooftop prompt 5 more opened the jacket over a bare chest. Those 17 pictures aren't posted; where one is missing, the post says so. Next run, figure prompts get nude, topless in the negative.

What happens next

Each LoRA's model page gets a short field note at the top: the trigger it actually learned, what it did here, and links to its posts.
The posts go up in small batches, vehicles first. Mistakes stay in. That's the point.

Every post

62 of 89 posts are up so far. The list grows as each batch goes up.

All my LoRAs are free here. They're visual research for Solarum, a biopunk novel I'm writing: two hundred years after a collapse, what's left of humanity living around Saturn, and an authority offering to let everyone sleep through the hard part.

Chapters and world notes: signalofsolarum.substack.com

Questions welcome in the comments. If you've run something like this on your own library, I'd like to see it.

I put 64 of my LoRAs on the same eight prompts. Here's what each one actually does.

How it was run

What I found

1. Subject LoRAs don't add. They replace.

2. The smallest training sets behave worst

3. Letters turn up where nobody asked for them

4. Some LoRAs draw panels instead of a scene

5. Vehicles go everywhere, except with the LoRA named for them

6. Stacking held up

7. Two came out broken

8. Some of them undressed the figure

What happens next

Every post

Batch 1 · Vehicles

Batch 2 · Vehicles, stacked

Batch 3 · Cats, dogs, mice and spiders

Batch 4 · Critters, birds, fish and pets

Batch 5 · Spiders and pets, stacked

Batch 6 · Places I

Batch 7 · Places II

Batch 8 · Signs and symbols

Batch 9 · Architecture, sigils and punk registers, stacked

Batch 10 · Things

Batch 11 · People