Real North American railroading for MOSS-Image. This LoRA teaches MOSS-Image what diesel-era freight trains actually look like: the locomotive models, the railroad paint schemes, the rolling stock, and the look of a real railfan photograph.
Base MOSS-Image draws "a train". With this LoRA it can draw a Southern Pacific SD45 in gray and scarlet leading a freight around a curve in the Tehachapis, with the right carbody, the right paint, and a road number on the cab.
What it knows
Locomotives (33 models)
EMD: F7, GP7, GP9, GP38, GP38-2, GP40, GP40-2, SD40, SD40-2, SD40T-2 "Tunnel Motor", SD45, SD50, SD60, SD60M, SD70M, SD70M-2, SD70MAC, SD70ACe, SW1500
GE: U23B, C30-7, B40-8, C40-8, C40-8W "Dash 8", C44-9W "Dash 9", AC4400CW, AC6000CW, ES44AC, ES44DC, ES44C4, ET44AC
ALCO / MLW: RS-3, RS-18
Rolling stock (13 types)
Boxcar, reefer, covered hopper, open hopper, gondola, tank car, flatcar, bulkhead flatcar, centerbeam flatcar, coil car, autorack, well car (double-stack containers), and caboose.
Railroads and paint schemes
It learned Class I, regional, and fallen-flag railroads, including Union Pacific, BNSF, Santa Fe (including the warbonnet), Burlington Northern, Norfolk Southern, CSX, Conrail, Southern Pacific, Canadian National, Canadian Pacific, Chessie System, and many more.
Photography
Realistic railfan-style photography: front three-quarter "wedge" shots, broadside views, head-on approaches, and yard scenes. It also handles night, snow, rain, fog, desert, mountains, and prairie. It renders reporting marks and road numbers on cars and cabs fairly reliably.
How to prompt
Write the prompt as a plain description of a photo, the way a railfan would caption it:
Front three-quarter view of a yellow and gray Union Pacific EMD SD70ACe diesel locomotive leading a long train of gray covered hoppers across open prairie under a clear blue sky. The wide safety cab carries the number 8752 in black on the nose, and UNION PACIFIC is lettered in red along the yellow long hood.
Tips:
Name the builder and model ("EMD SD40-2", "GE ES44AC"). This drives the carbody shape.
Name the railroad and colors ("blue and yellow Conrail"). The model knows the schemes, but stating the colors helps.
Give a road number or reporting marks ("UTLX 643218") if you want legible lettering.
Describe the camera angle and setting (side view, front three-quarter, night yard, snowy mainline).
Skip quality tags like "masterpiece". The training captions don't use them.
Recommended settings
LoRA strength: 1.0 (lower it to 0.7–0.8 if images look soft)
Steps: 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
Base vs SFT: the LoRA was trained on MOSS-Image-Base, but it also works well on MOSS-Image-SFT.
Base + LoRA gives the most authentic look, close to real railfan photos, slightly soft and film-like.
SFT + LoRA is sharper and gets lettering and heralds right more often, but looks glossier and more "CG". The showcase images include both.
Limitations
Railroad heralds and long lettering are sometimes garbled (Conrail's logo, "Canadian Pacific"). Short road numbers and reporting marks are much more reliable.
Rare or similar-looking models can blend together. For example, an SW1500 may come out as a GP-style road switcher, and an RS-18 as a generic hood unit.
Some specialized cars (coil cars, bulkhead flats) are hit-or-miss on Base. SFT does better.
Images can be a little soft at strength 1.0.
This LoRA covers North American diesel-era freight. Steam, passenger trains, and European or Asian equipment are out of scope.
Training details
MOSS-Image-Base, kohya sd-scripts
LoRA rank 64, alpha 16
AdamW8bit, learning rate 1e-4, constant schedule
About 5,000 steps (1 epoch), batch size 1, aspect-ratio buckets around 1024 px
INT8 base weights, bf16, trained on an RTX 4070 Super