Best practices for non-square aspect ratios

[PCanavelli]

Hi @bghira

Quick usage question: I need to train Flux adapters on non-square images (e.g., 1200 x 1600, 3:4 aspect ratio).

What changes should be made to a default training config to take this into account?

So far, I couldn't find the way to specify non-square resolutions through the resolution argument in the trainer config, as both (1200, 1600) and "1200,1600" fail.

Also: are there other parameters that should be set to fully propagate this change?

Cheers

[bghira]

please be considerate when pinging project maintainers.. see the dataloader documentation.

[PCanavelli]

This has to be the first time ever I see a maintainer complain about being tagged in a question directly addressed to them.

Also: if this is what you mean by "go read the docs", you would still have to explain to me how this paragraph could possibly answer my question, since all it does is listing all valid options for each parameter, without giving any indication whatsoever regarding which combinations are valid, nor any clue as to which ones should be favoured.

Sorry mate. I'm all for the "figure it out yourself" approach as a first answer to questions on open source projects, but this approach assumes the person hasn't already been digging through your docs (and code), and that those are informative.

I did, and they're really not.

[
  {
    "id": "something-special-to-remember-by",
    "type": "local",
    "instance_data_dir": "/path/to/data/tree",
    "crop": false,
    "crop_style": "random|center|corner|face",
    "crop_aspect": "square|preserve|closest|random",
    "crop_aspect_buckets": [0.33, 0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
    "resolution": 1.0,
    "resolution_type": "area|pixel",
    "minimum_image_size": 1.0,
    "hash_filenames": true,
    "prepend_instance_prompt": false,
    "instance_prompt": "something to label every image",
    "only_instance_prompt": false,
    "caption_strategy": "filename|instanceprompt|parquet|textfile",
    "cache_dir_vae": "/path/to/vaecache",
    "vae_cache_clear_each_epoch": true,
    "probability": 1.0,
    "repeats": 0,
    "text_embeds": "alt-embed-cache",
    "image_embeds": "vae-embeds-example"
  },
  {
    "id": "another-special-name-for-another-backend",
    "type": "aws",
    "aws_bucket_name": "something-yummy",
    "aws_region_name": null,
    "aws_endpoint_url": "https://foo.bar/",
    "aws_access_key_id": "wpz-764e9734523434",
    "aws_secret_access_key": "xyz-sdajkhfhakhfjd",
    "aws_data_prefix": "",
    "cache_dir_vae": "s3prefix/for/vaecache",
    "vae_cache_clear_each_epoch": true,
    "repeats": 0
  },
  {
      "id": "vae-embeds-example",
      "type": "local",
      "dataset_type": "image_embeds",
      "disabled": false,
  },
  {
    "id": "an example backend for text embeds.",
    "dataset_type": "text_embeds",
    "default": true,
    "type": "aws",
    "aws_bucket_name": "textembeds-something-yummy",
    "aws_region_name": null,
    "aws_endpoint_url": "https://foo.bar/",
    "aws_access_key_id": "wpz-764e9734523434",
    "aws_secret_access_key": "xyz-sdajkhfhakhfjd",
    "aws_data_prefix": "",
    "cache_dir": ""
  },
  {
    "id": "alt-embed-cache",
    "dataset_type": "text_embeds",
    "default": false,
    "type": "local",
    "cache_dir": "/path/to/textembed_cache"
  }
]

[bghira]

theres a sequence of examples in that doc that you ignored.