LTX-2 loading fails on Hugging Face config metadata keys

[andyelka-creator]

Summary

Running LTX-2 image-to-video from the current main can fail after downloading model weights because LTXModel.from_pretrained() constructs LTXModelConfig directly from config.json.

Recent Hugging Face/Diffusers-style configs may include metadata keys such as _class_name and _diffusers_version. Those keys are not dataclass fields on LTXModelConfig, so loading aborts before generation starts.

Reproduction

Run LTX-2 with the default model repo after weights are available/downloaded:

mlx_video.ltx_2.generate \
  --prompt "Animate the image" \
  --pipeline distilled \
  --height 512 \
  --width 512 \
  --num-frames 97 \
  --image horse1.png \
  --output-path output.mp4

Actual Result

After the model snapshot is fetched and the text encoder loads, model loading fails:

Text encoder loaded successfully
✓ Text encoder loaded
TypeError: LTXModelConfig.__init__() got an unexpected keyword argument '_class_name'

Expected Result

from_pretrained() should ignore metadata keys that are not part of LTXModelConfig, matching the existing BaseModelConfig.from_dict() behavior.

Proposed Fix

Change LTXModel.from_pretrained() to build config with:

config = LTXModelConfig.from_dict(config_dict)

instead of:

config = LTXModelConfig(**config_dict)

This preserves known fields and filters unknown metadata keys.