base F1.d doesn't know this building at all, so this is a training experiment for me to better understand how to train F1D LoRA models for non-people/non-characters.I'm pretty satisfied with v5, I don't think I will be training any more versions unless I find out something revolutionary.<h2 id="my-training-notes-g5doyjtqw">My training notes</h2><h3 id="v5-excellent-i-am-satisfy-2cd1mqe81">v5 - excellent, i am satisfy</h3>v5 works with onsite generator but isn't as consistent as local generation<ul><li>gets general size and proportions right, and finally picked up on the two metal spires atop the building</li><li>interior geometry is consistent, size (of room) is acceptable too</li><li>I don't think I will train any more versions, but if i do, i would address<ul><li>reduce film grain/analogue film effect</li><li>pods are slightly fuzzy (i think because some of the training images had them with demolition netting over)</li><li>interiors always tend to generate hardwood floors</li><li>sky tends to be overexposed (will have to photoshop dataset or find new images)</li></ul></li><li>switching to Ostris' ai-toolkit</li><li>network rank (and alpha) increased to 32</li><li>4000 steps</li><li>LR 1e-4 (default for flux preset)</li><li>explicit triggerword set <code>nctower</code></li></ul><h3 id="v4-interior-better-exterior-worse-bhl4lbfov">v4 - interior better, exterior worse</h3><ul><li>It's kind of bad at replicating the exterior, but it is getting some of the interior geometries right</li><li>network rank seems to be max 16 with my current simpletuner setup, i can't seem to make it any higher</li><li>LR reset to default</li><li>changed to CosineAnnealingWarmRestarts scheduler, 4 restarts</li><li>max steps reverted to 2000</li></ul><h3 id="v3-something-went-horribly-wrong-ohzj7n6bk">v3 something went horribly wrong</h3><ul><li>Garbage output, maybe doubling LR wasn't the best idea?</li><li>Dataset updated (more variety, particularly added more that depicted people, more detailed captions)</li><li>all settings reset to default (same as V1), except:</li><li>max steps reduced to 1800</li><li>learning rate doubled to 10e4</li><li><s>network rank doubled to 32</s> somehow the setting didn't change?</li></ul><h3 id="v2-lora-is-okay-18bcdbb3k">v2 LoRA is okay</h3><ul><li>I increased the batch size 4x, and as a result it was trained for much longer (2000 steps × batch size 4, I cut it off at 1500 as it was taking way too long on my rented hourly rate GPU)</li><li>it gets the exterior appearance consistent,</li><li>it is starting to understand the qualities of the interior rooms.</li><li>it is not overtrained or "burned"; it can render other types of towers just fine in the same prompt</li><li>I think it needs a higher network rank to fully understand both interior and exterior qualities.</li></ul><h3 id="v1-lora-is-undertrained-p4waj28ab">v1 LoRA is undertrained</h3><ul><li>using sinmpletuner</li><li>default learning rate, 2000 steps × batch size 1</li><li>exterior appearance: it understands the individual pods but rarely arranges them in the same type of quasi-random geometry</li><li>interior: it understands the narrow room with single round window, but not much else</li></ul>

Files with SHA256: fbb3d02f02c65239ef3b99d5f36860ab94d68a98ab1330be1647ddf47ef8d25e