Text-to-Image
PyTorch
Safetensors
diffusion-transformer
rectified-flow
from-scratch
tinydit-256 / config.json
ivanmikhnenkov's picture
add config.json
def994c verified
Raw History Blame Contribute Delete
703 Bytes
{
"architecture": "TinyDiT",
"latent_ch": 32,
"ctx_dim": 768,
"patch": 2,
"n_registers": 16,
"register_block": 3,
"n_null": 2,
"dim": 896,
"depth": 16,
"heads": 14,
"autoencoder": "black-forest-labs/FLUX.2-dev (vae subfolder)",
"text_encoder": "google/flan-t5-base",
"max_tokens_long": 128,
"max_tokens_short": 48,
"latent_whitening": "stats.json (per-channel mean/std, applied after encode and undone before decode)",
"sampler": {
"steps": 20,
"cfg": 4.0,
"shift": 2.8,
"solver": "euler"
},
"training_shapes": [
[
256,
256
],
[
288,
224
],
[
224,
288
],
[
320,
208
],
[
208,
320
]
],
"step": 400000,
"ema": "0.9999, bf16"
}