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.

My training notes

v5 - excellent, i am satisfy

v5 works with onsite generator but isn't as consistent as local generation

• gets general size and proportions right, and finally picked up on the two metal spires atop the building

• interior geometry is consistent, size (of room) is acceptable too

• I don't think I will train any more versions, but if i do, i would address

  • reduce film grain/analogue film effect

  • pods are slightly fuzzy (i think because some of the training images had them with demolition netting over)

  • interiors always tend to generate hardwood floors

  • sky tends to be overexposed (will have to photoshop dataset or find new images)

• switching to Ostris' ai-toolkit

• network rank (and alpha) increased to 32

• 4000 steps

• LR 1e-4 (default for flux preset)

• explicit triggerword set nctower

v4 - interior better, exterior worse

• It's kind of bad at replicating the exterior, but it is getting some of the interior geometries right

• network rank seems to be max 16 with my current simpletuner setup, i can't seem to make it any higher

• LR reset to default

• changed to CosineAnnealingWarmRestarts scheduler, 4 restarts

• max steps reverted to 2000

v3 something went horribly wrong

• Garbage output, maybe doubling LR wasn't the best idea?

• Dataset updated (more variety, particularly added more that depicted people, more detailed captions)

• all settings reset to default (same as V1), except:

• max steps reduced to 1800

• learning rate doubled to 10e4

• network rank doubled to 32 somehow the setting didn't change?

v2 LoRA is okay

• 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)

• it gets the exterior appearance consistent,

• it is starting to understand the qualities of the interior rooms.

• it is not overtrained or "burned"; it can render other types of towers just fine in the same prompt

• I think it needs a higher network rank to fully understand both interior and exterior qualities.

v1 LoRA is undertrained

• using sinmpletuner

• default learning rate, 2000 steps × batch size 1

• exterior appearance: it understands the individual pods but rarely arranges them in the same type of quasi-random geometry

• interior: it understands the narrow room with single round window, but not much else