LoRA strength scaling appears inconsistent: object LoRA (r=16 / alpha=64) requires strength ~3.0 to match trained geometry

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

A LoRA trained with rank (r) = 16 and lora_alpha = 64 (alpha/r = 4.0), focused on a rigid object with clear geometry (vehicle), is expected to produce a recognizable and proportionally correct result at LoRA strength ≈ 1.0, or at least without requiring extreme strength values.

This expectation is consistent with:

- diffusers-based inference
- SimpleTuner training assumptions
- other LoRA pipelines using similar rank/alpha configurations

### Actual Behavior

When the LoRA is loaded in ComfyUI, the effect at strength = 1.0 is extremely subtle.

Observed behavior:

- strength = 1.0 → vehicle shape barely resembles the trained object
- strength = 2.0 → still insufficient similarity
- strength ≈ 3.0 → geometry and proportions finally match the training data

This behavior is consistent across multiple prompts and seeds.

The issue is particularly evident because the LoRA represents a vehicle, where proportions and scale are easy to visually validate.

### Steps to Reproduce

1. Load FLUX-2 klein-9b (tested on both base and distilled variants).
2. Load the LoRA provided below (trained with SimpleTuner, diffusers/PEFT).
3. Use the following prompt (or any similar object-focused prompt):

evospazio, SUV, front three-quarter view, full vehicle in frame,
correct proportions, standard wheelbase,
clean studio lighting, neutral background

4. Keep seed, sampler, steps, CFG, and resolution fixed.

5. Generate images at different LoRA strength values:

- 1.0
- 2.0
- 3.0

6. Compare the generated results against the training reference image.

### Debug Logs

```powershell
- LoRA metadata (from .safetensors):

transformer.r = 16
transformer.lora_alpha = 64

- Training framework: SimpleTuner (diffusers + PEFT)
- LoRA format: .safetensors
- No missing or malformed metadata detected.

If needed, additional logs or metadata dumps can be provided.
```

### Other

- The LoRA is publicly available on Hugging Face for reproducibility:
https://huggingface.co/francipam/evospazio

- Only the last checkpoint is provided

- A training reference image is included to clearly show the intended vehicle geometry and proportions.

- This report is not about subjective style strength, but about objective object geometry and proportional correctness, which should make the issue straightforward to evaluate.