Upload luminars/config.py

luminars/config.py

@@ -3,27 +3,26 @@ from dataclasses import dataclass
 
 @dataclass
 class LuminaRSConfig:
-    # Core dims
-    latent_dim: int          = 16      # SD-compatible VAE latent dim
+    # Latent dims
+    latent_dim: int          = 16      # VAE latent channels
     latent_h: int            = 32      # 32x32 latent = 1024x1024 px
     latent_w: int            = 32
-    # SSM Core
-    d_model: int             = 256     # main hidden dim
-    d_state: int             = 64      # mamba S4D hidden state
-    d_conv: int              = 4       # local convolution width
-    expand: int              = 2       # expansion factor in SSM
-    n_layers: int            = 8       # SSM-repeat stack
-    drop_path: float         = 0.0     # stochastic depth
-    # Semantic (text) encoder
-    text_embed_dim: int      = 768     # CLIP-T hidden dim
+    # UNet channels at each scale
+    channels: tuple          = (64, 128, 256, 256, 384)
+    # Bottleneck depth
+    n_bottleneck: int        = 4
+    # Iterative refinement depth (TRM-style shared-weight recursion)
+    n_recurse: int           = 2
+    # Time embedding
+    t_embed_dim: int         = 256
+    # Text conditioning
+    text_embed_dim: int      = 768
     max_text_len: int        = 77
-    prompt_embed_dim: int    = 256     # compressed prompt plan
-    # Style/Mood/Philosophy modules
-    style_dim: int           = 128
-    mood_dim: int            = 64
-    philosophy_dim: int      = 64
-    # Flow head
-    n_flow_steps: int        = 10      # ODE solver steps at inference
     # Training
-    vae_pretrained: str      = "madebyollin/sdxl-vae-fp16-fix"  # 31M param VAE
+    drop_path: float         = 0.05
+    # VAE (frozen)
+    vae_pretrained: str      = "madebyollin/sdxl-vae-fp16-fix"
+    # Text encoder (frozen)
     clip_pretrained: str     = "openai/clip-vit-large-patch14"
+    # Flow matching inference steps
+    n_flow_steps: int        = 12