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.<h2 id="what-it-knows">What it knows</h2>Locomotives (33 models)<ul><li>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</li><li>GE: U23B, C30-7, B40-8, C40-8, C40-8W "Dash 8", C44-9W "Dash 9", AC4400CW, AC6000CW, ES44AC, ES44DC, ES44C4, ET44AC</li><li>ALCO / MLW: RS-3, RS-18</li></ul>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 schemesIt 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.PhotographyRealistic 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.<h2 id="how-to-prompt">How to prompt</h2>Write the prompt as a plain description of a photo, the way a railfan would caption it:<blockquote>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.</blockquote>Tips:<ul><li>Name the builder and model ("EMD SD40-2", "GE ES44AC"). This drives the carbody shape.</li><li>Name the railroad and colors ("blue and yellow Conrail"). The model knows the schemes, but stating the colors helps.</li><li>Give a road number or reporting marks ("UTLX 643218") if you want legible lettering.</li><li>Describe the camera angle and setting (side view, front three-quarter, night yard, snowy mainline).</li><li>Skip quality tags like "masterpiece". The training captions don't use them.</li></ul><h2 id="recommended-settings">Recommended settings</h2><ul><li>LoRA strength: 1.0 (lower it to 0.7–0.8 if images look soft)</li><li>Steps: 50</li><li>CFG: 3</li><li>Sampler / scheduler: euler / sgm_uniform</li><li>Shift (ModelSamplingSD3): 5</li><li>Resolution: ~1 MP, e.g. 1216×832 (landscape suits trains)</li><li>Negative (Base): nsfw, worst quality, bad anatomy, watermark, jpeg artifact</li></ul>Base vs SFT: the LoRA was trained on MOSS-Image-Base, but it also works well on MOSS-Image-SFT.<ul><li>Base + LoRA gives the most authentic look, close to real railfan photos, slightly soft and film-like.</li><li>SFT + LoRA is sharper and gets lettering and heralds right more often, but looks glossier and more "CG". The showcase images include both.</li></ul><h2 id="limitations">Limitations</h2><ul><li>Railroad heralds and long lettering are sometimes garbled (Conrail's logo, "Canadian Pacific"). Short road numbers and reporting marks are much more reliable.</li><li>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.</li><li>Some specialized cars (coil cars, bulkhead flats) are hit-or-miss on Base. SFT does better.</li><li>Images can be a little soft at strength 1.0.</li><li>This LoRA covers North American diesel-era freight. Steam, passenger trains, and European or Asian equipment are out of scope.</li></ul><h2 id="training-details">Training details</h2><ul><li>MOSS-Image-Base, kohya sd-scripts</li><li>LoRA rank 64, alpha 16</li><li>AdamW8bit, learning rate 1e-4, constant schedule</li><li>About 5,000 steps (1 epoch), batch size 1, aspect-ratio buckets around 1024 px</li><li>INT8 base weights, bf16, trained on an RTX 4070 Super</li></ul>

Files with SHA256: 978a7a84bcc57a05ee8c8f0f3584ba207df02e2457ed3a2a8abda55602eb9fa6