|
--- |
|
license: cc |
|
tags: |
|
- text-to-image |
|
- lora |
|
- diffusers |
|
- template:sd-lora |
|
base_model: |
|
- black-forest-labs/FLUX.1-dev |
|
widget: |
|
- text: brazil |
|
output: |
|
url: images/brazil.png |
|
- text: canada, geoguessr |
|
output: |
|
url: images/canada.png |
|
- text: mongolia |
|
output: |
|
url: images/mongolia.png |
|
- text: serbia village |
|
output: |
|
url: images/serbia.png |
|
- text: thailand |
|
output: |
|
url: images/thailand.png |
|
--- |
|
|
|
# fake geoguessr locations lora for flux-dev |
|
|
|
https://x.com/_lyraaaa_/status/1841762752404369745 |
|
|
|
rank 32, trained for 3500 steps on over 200 labeled locations. trigger word ("geoguessr") not always necessary, just name a location |
|
|
|
**run this with diffusers:** |
|
|
|
```py |
|
import torch |
|
from diffusers import FluxPipeline |
|
import time |
|
import random |
|
|
|
# initialize pipeline and lora |
|
pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda") |
|
|
|
lora_weight = 0.8 |
|
pipe.load_lora_weights( |
|
'/workspace/geoguessr_v1_000003500.safetensors', |
|
adapter_name='geoguessr_v1' |
|
) |
|
pipe.set_adapters('geoguessr_v1', adapter_weights=[lora_weight]) |
|
|
|
# set params and generate |
|
seed = -1 |
|
seed = seed if seed != -1 else random.randint(0, 2**32) |
|
print(seed) |
|
|
|
prompt = "sweden, snow" |
|
out = pipe( |
|
prompt=prompt, |
|
guidance_scale=4, |
|
height=624, |
|
width=960, |
|
num_inference_steps=40, |
|
generator=torch.Generator("cuda").manual_seed(seed), |
|
).images[0] |
|
|
|
# save and display output |
|
filename=f"{time.time()}.png" |
|
out.save(filename) |
|
|
|
from IPython.display import Image, display |
|
display(Image(filename=filename)) |
|
|
|
``` |
|
|
|
**known model biases:** |
|
- v1 of this model leans heavily towards rural locations due to dataset bias, will be fixed in v2 as i collect more data |
|
- it managed to generalize to locations not available on geoguessr, like china, although it drifts towards generic locations |
|
- its trained on lowercase country names, and flux is case sensitive. results may vary |
|
- it LOVES orange/red dirt colors. this will be fixed in v2 also |
|
|
|
geoguessr_v2 with a much larger dataset and less location bias will be out eventually. |
|
|
|
since i do not own the data for this model, i can't really claim ownership of the model itself either. have fun! |
|
|
|
<Gallery /> |
|
|
|
trained with https://github.com/ostris/ai-toolkit/blob/main/notebooks/FLUX_1_dev_LoRA_Training.ipynb |
|
|
|
this model is a part of my much larger desterilizer project- a bit more here https://x.com/_lyraaaa_/status/1824003678086590646 |
