refactor(i2l): remove --rank and --lora-scale from CLI and pipeline

These are unnecessary:
- --lora-scale: already available via --lora-scales at generation time
- --rank: determined naturally by number of images (rank = 4 * N)

CLI is now minimal: --image-path and --output only.

Author: azrahello

src/mflux/models/z_image/cli/z_image_i2l.py

@@ -66,23 +66,6 @@ def main():
         default="lora.safetensors",
         help="Output path for the generated LoRA file. Default: lora.safetensors",
     )
-    parser.add_argument(
-        "--lora-scale",
-        type=float,
-        default=1.0,
-        help="Scale factor applied to all LoRA weights before saving. "
-        "Values > 1.0 amplify the style effect, < 1.0 soften it. Default: 1.0",
-    )
-    parser.add_argument(
-        "--rank",
-        "-r",
-        type=int,
-        default=None,
-        help="Target LoRA rank. Base rank is 4 per image. With N images, "
-        "default rank is 4*N (e.g. 16 for 4 images). Set this to override: "
-        "each image will be repeated ceil(rank / (4*N)) times to reach the "
-        "target. Valid values: multiples of 4. Typical: 4, 16, 32, 64, 128.",
-    )
 
     args = parser.parse_args()
 
@@ -102,7 +85,7 @@ def main():
 
     # Create pipeline and generate LoRA
     pipeline = ZImageI2LPipeline.from_pretrained()
-    pipeline.generate_lora(images=images, output_path=args.output, lora_scale=args.lora_scale, target_rank=args.rank)
+    pipeline.generate_lora(images=images, output_path=args.output)
 
 
 if __name__ == "__main__":

src/mflux/models/z_image/model/z_image_i2l/i2l_pipeline.py

@@ -158,17 +158,15 @@ def generate_lora(
         self,
         images: list[Image.Image],
         output_path: str | Path = "lora.safetensors",
-        lora_scale: float = 1.0,
-        target_rank: int | None = None,
     ) -> Path:
         """Full pipeline: encode images and save LoRA weights.
 
+        Each image contributes rank 4 to the output LoRA. With N images,
+        the output has effective rank 4*N (merged by concatenation).
+
         Args:
             images: List of style reference images.
             output_path: Where to save the generated .safetensors file.
-            lora_scale: Scale factor for LoRA weights.
-            target_rank: Target LoRA rank (must be multiple of 4).
-                If None, uses 4 * num_images (natural merge rank).
 
         Returns:
             Path to the saved LoRA file.
@@ -182,26 +180,7 @@ def generate_lora(
         embeddings = self.encode_images(images)
         encode_time = time.time() - t0
         print(f"  Encoding done in {encode_time:.1f}s")
-
-        # Determine repetitions to reach target rank
-        base_rank = 4
-        num_images = len(images)
-        natural_rank = base_rank * num_images
-
-        if target_rank is not None:
-            if target_rank % base_rank != 0:
-                raise ValueError(f"target_rank must be a multiple of {base_rank}, got {target_rank}")
-            repeats = max(1, (target_rank + natural_rank - 1) // natural_rank)  # ceil division
-            if repeats > 1:
-                print(
-                    f"  Repeating embeddings {repeats}x to reach rank {natural_rank * repeats} (target: {target_rank})"
-                )
-                embeddings = mx.concatenate([embeddings] * repeats, axis=0)
-        else:
-            target_rank = natural_rank
-
-        effective_rank = base_rank * embeddings.shape[0]
-        print(f"  Output LoRA rank: {effective_rank}")
+        print(f"  Output LoRA rank: {4 * len(images)}")
 
         # Decode to LoRA
         print("Generating LoRA weights...")
@@ -211,11 +190,6 @@ def generate_lora(
         print(f"  Decoding done in {decode_time:.1f}s")
         print(f"  Generated {len(lora)} LoRA weight tensors")
 
-        # Apply scale
-        if lora_scale != 1.0:
-            print(f"  Applying LoRA scale: {lora_scale}")
-            lora = {k: v * lora_scale for k, v in lora.items()}
-
         # Save as safetensors (convert MLX bfloat16 -> float32 -> torch bfloat16)
         print(f"Saving to {output_path}...")
         lora_torch = {}