Remove LoRA support, add auto-resize, fix GPU memory cleanup

LoRA loading via PEFT crashes on ROCm; removed since user doesn't need
it. Added automatic downscaling of inputs >512px to avoid O(N^2)
attention explosion in the Qwen transformer (262K tokens at 1024^2
exceeds GPU capability). Added explicit GPU memory cleanup after each
command to prevent OOM when chaining generate + edit.

vnassets/cli.py

@@ -53,15 +53,14 @@ def generate_cmd(checkpoint, prompt, negative_prompt, width, height, steps, cfg,
 @click.option("--model", required=True, help="Path to Qwen Image Edit diffusion model (.safetensors)")
 @click.option("--input", "input_path", required=True, help="Input image to edit")
 @click.option("--prompt", required=True, help="Edit instruction")
-@click.option("--steps", default=20, type=int, help="Inference steps (4 for turbo)")
-@click.option("--cfg", default=4.0, type=float, help="CFG scale (1.0 for turbo)")
+@click.option("--steps", default=20, type=int, help="Inference steps")
+@click.option("--cfg", default=4.0, type=float, help="CFG scale")
 @click.option(
     "--seed", default="random", callback=_parse_seed,
     help="RNG seed (integer or 'random')",
 )
-@click.option("--lora", "lora_path", default=None, help="Path to LoRA weights (.safetensors)")
 @click.option("--output", default="output.png", help="Output image path")
-def edit_cmd(model, input_path, prompt, steps, cfg, seed, lora_path, output):
+def edit_cmd(model, input_path, prompt, steps, cfg, seed, output):
     """Edit an image using Qwen Image Edit."""
     edit(
         model_path=model,
@@ -70,6 +69,5 @@ def edit_cmd(model, input_path, prompt, steps, cfg, seed, lora_path, output):
         steps=steps,
         cfg=cfg,
         seed=seed,
-        lora_path=lora_path,
         output_path=output,
     )

vnassets/edit.py

@@ -58,7 +58,6 @@ def edit(
     steps: int = 20,
     cfg: float = 4.0,
     seed: int | None = None,
-    lora_path: str | None = None,
     output_path: str = "output.png",
 ) -> None:
     device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -91,13 +90,20 @@ def edit(
     )
     pipe.to(device)
 
-    if lora_path:
-        pipe.load_lora_weights(lora_path)
-        pipe.fuse_lora()
-
     t_load = time.perf_counter() - t0
 
+    # Resize large inputs to avoid GPU crashes. The Qwen transformer does
+    # O(N^2) attention on image patches. 512px is safe on this GPU.
     input_image = Image.open(input_path).convert("RGB")
+    w, h = input_image.size
+    max_dim = 512
+    if w > max_dim or h > max_dim:
+        scale = max_dim / max(w, h)
+        new_w, new_h = int(w * scale), int(h * scale)
+        new_w = (new_w // 16) * 16
+        new_h = (new_h // 16) * 16
+        input_image = input_image.resize((new_w, new_h), Image.LANCZOS)
+        print(f"Resized input {w}x{h} -> {new_w}x{new_h}")
     generator = torch.Generator(device=device).manual_seed(seed)
 
     t1 = time.perf_counter()
@@ -113,6 +119,10 @@ def edit(
     image.save(output)
     print(f"Saved {output}")
 
+    del pipe, transformer, vae, text_encoder
+    if device == "cuda":
+        torch.cuda.empty_cache()
+
     meta_path = output.with_suffix(".json")
     meta = {
         "model": str(Path(model_path).resolve()),
@@ -123,7 +133,6 @@ def edit(
         "steps": steps,
         "cfg": cfg,
         "seed": seed,
-        "lora": str(Path(lora_path).resolve()) if lora_path else None,
         "load_time_s": round(t_load, 2),
         "inference_time_s": round(t_infer, 2),
     }

vnassets/generate.py

@@ -57,6 +57,11 @@ def generate(
     image.save(output)
     print(f"Saved {output}")
 
+    # Free GPU memory before returning
+    del pipe
+    if device == "cuda":
+        torch.cuda.empty_cache()
+
     meta_path = output.with_suffix(".json")
     meta = {
         "checkpoint": str(Path(checkpoint_path).resolve()),