di.FFUSION.ai-tXe-FXAA
Trained on "121361" images.
Enhance your model's quality and sharpness using your own pre-trained Unet.
The text encoder (without UNET) is wrapped in LyCORIS. Optimizer: torch.optim.adamw.AdamW(weight_decay=0.01, betas=(0.9, 0.99))
Network dimension/rank: 768.0 Alpha: 768.0 Module: lycoris.kohya {'conv_dim': '256', 'conv_alpha': '256', 'algo': 'loha'}
Large size due to Lyco CONV 256
For a1111
Install https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris
Download di.FFUSION.ai-tXe-FXAA to /models/Lycoris
Option1:
Insert <lyco:di.FFUSION.ai-tXe-FXAA:1.0> to prompt
No need to split Unet and Text Enc as its only TX encoder there.
You can go up to 2x weights
Option2: If you need it always ON (ex run a batch from txt file) then you can go to settings / Quicksettings list
add sd_lyco
restart and you should have a drop-down now 🤟 🥃
More info:
"ss_text_encoder_lr": "1e-07",
"ss_keep_tokens": "3",
"ss_network_args": {
"conv_dim": "256",
"conv_alpha": "256",
"algo": "loha"
},
"img_count": 121361
}
"ss_total_batch_size": "100",
"ss_network_dim": "768",
"ss_max_bucket_reso": "1024",
"ss_network_alpha": "768.0",
"ss_steps": "2444",
"sshs_legacy_hash": "539b2745",
"ss_batch_size_per_device": "20",
"ss_max_train_steps": "2444",
"ss_network_module": "lycoris.kohya",
This is a heavy experimental version we used to test even with sloppy captions (quick WD tags and terrible clip), yet the results were satisfying.
Note: This is not the text encoder used in the official FFUSION AI model.
Description
FAQ
Comments (7)
2.5 gb lycoris??
Network dimension/rank: 768.0 Alpha: 768.0 Module: lycoris.kohya {'conv_dim': '256', 'conv_alpha': '256', 'algo': 'loha'}
Large size due to Lyco CONV 256
had to work with 80G cards for the task.
feel free to convert it.
Sharing the full Lyco CONV 256 - Float 🤟 🥃
How long did you train him and in what configuration
Training data is left intact in the meta of the file ;)
but overall under a day with 5xA100-80GB x 20 batch sizes each.
ss_random_crop: "False",
ss_min_snr_gamma: "4.0",
ss_max_token_length: "225",
ss_lr_scheduler: "linear",
ss_lr_warmup_steps: "303",
ss_num_batches_per_epoch: "1222",
ss_optimizer: "torch.optim.adamw.AdamW(weight_decay=0.01,betas=(0.9, 0.99))",
ss_bucket_no_upscale: "True",
"ss_text_encoder_lr": "1e-07",
"ss_keep_tokens": "3",
"ss_network_args": {
"conv_dim": "256",
"conv_alpha": "256",
"algo": "loha"
},
"img_count": 121361
}
"ss_total_batch_size": "100",
"ss_network_dim": "768",
"ss_max_bucket_reso": "1024",
"ss_network_alpha": "768.0",
"ss_steps": "2444",
"ss_batch_size_per_device": "20",
"ss_max_train_steps": "2444",
"ss_network_module": "lycoris.kohya",
Could I get another explanation of what this is supposed to do? Is it replacing or adding to the text encoder of whatever checkpoint you're using? And it's supposed to improve it? Are there cases where this could damage anything? like a trained face?
RuntimeError: The size of tensor a (768) must match the size of tensor b (1024) at non-singleton dimension 1
can this not be used with just any model? do i need a specific model for this?
sorry im an idiot, i didnt read the whole title!! sorry!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.