Real North American passenger railroading for MOSS-Image. This LoRA teaches MOSS-Image what passenger trains actually look like: the locomotive models, the operators' paint schemes, the coaches and trainsets, and the look of a real railfan photograph.
Base MOSS-Image draws "a train". With this LoRA it can draw a Coaster F40PH leading a bilevel train along the Southern California coast, or an Amtrak F40PH in Phase III paint at the head of a string of Amfleet coaches, with the right carbody, the right paint, and a road number on the cab.
It is the passenger companion to North American Diesel Freight Trains, trained separately on passenger equipment only.
What it knows
Diesel locomotives
EMD F40PH in Amtrak (Phase II and III), commuter (Metra, MBTA, Caltrain, Tri-Rail, ACE and others) and Coaster (NCTD) paint, the GMD F40PH-2 in VIA Rail paint, and Amtrak's F40PH-based cabbage cars (non-powered control units)
GE Genesis P42DC, P40DC and P32AC-DM
EMD F59PH, F59PHI, GP40PH-2, E8/E9
MPI MP36PH-3C and MP40PH-3C
Siemens Charger SC-44 and ALC-42
MLW LRC
Electric locomotives
Siemens ACS-64, EMD AEM-7, Bombardier/Alstom HHP-8, PRR GG1
Cars and trainsets
Amfleet, Horizon, Viewliner, Superliner, LRC cars, Comet, Bombardier BiLevel and MultiLevel, Kawasaki bilevels, gallery cars, Acela Express, Avelia Liberty (NextGen Acela), Talgo, Brightline, Budd RDC, Silverliner
Operators and paint schemes
It learned Amtrak (including the Phase II and Phase III schemes), VIA Rail Canada, Coaster (NCTD), Pacific Surfliner, Metra, MBTA, Caltrain, Metro-North, New Jersey Transit, GO Transit, SEPTA, Metrolink, Brightline, Amtrak Cascades and others, plus heritage roads such as the Pennsylvania Railroad.
Photography
Realistic railfan-style photography: front three-quarter shots, broadside views, head-on approaches and station platforms. It also handles night, snow, fog, coast, mountains and city terminals, and renders road numbers on cabs fairly reliably.
How to prompt
Start with the operator and the model, then write the rest as a plain description of a photo, the way a railfan would caption it:
Amtrak EMD F40PH diesel locomotive, Phase III paint. Three-quarter front view of an Amtrak EMD F40PH diesel locomotive numbered 321 leading a train of silver Amfleet coaches along a double-track mainline on a clear afternoon. The silver body carries broad red, white and blue stripes, and the number is on the cab front above the windshield.
Tips:
Lead with the operator, builder and model ("Coaster (NCTD) F40PH diesel locomotive", "VIA Rail Canada GMD F40PH-2 diesel locomotive"). Every training caption starts this way, and a bare one-line name prompt works too.
Name the paint scheme and colors ("Phase III paint", "green and white GO Transit").
Give a road number if you want legible numbers on the cab.
Describe the camera angle and setting (side view, station platform, snowy forest, night).
Skip quality tags like "masterpiece". The training captions don't use them.
Recommended settings
LoRA strength: 1.0
Steps: 30–50
CFG: 3
Sampler / scheduler: euler / sgm_uniform
Shift (ModelSamplingSD3): 5
Resolution: ~1 MP, e.g. 1216×832 (landscape suits trains)
Negative (Base): nsfw, worst quality, bad anatomy, watermark, jpeg artifact
The LoRA was trained on MOSS-Image-Base, and the showcase images are Base + LoRA.
Limitations
The Acela Express and the Avelia Liberty can blend together.
Some classes had very few photos: Coaster's F40PH (7 training photos), F59PHI and SC-44, and the MLW LRC. Expect the paint to come through better than exact carbody details there.
Amtrak's cabbage cars and real F40PHs are hard to tell apart in photos, and the LoRA can mix them.
Long lettering and heralds are sometimes garbled. Short road numbers are much more reliable.
This LoRA covers North American passenger equipment. Freight is covered by the freight LoRA; steam and European or Asian equipment are out of scope.
Training data
2,976 photos from Wikimedia Commons (2,667 used for training), all under free licenses: CC BY, CC BY-SA, CC0, public domain and GFDL.
Organized into 40 locomotive, car and trainset classes.
Every image was checked against its class and given a detailed natural-language caption: model, operator, paint, road number, angle, and setting. Each also got medium and short captions, so short prompts work too.
Thank you to the Wikimedia Commons photographers whose work made this possible.