LTX-2 transformer weights fail strict loading with Diffusers-style keys

[andyelka-creator]

Problem

After fixing Hugging Face config metadata parsing, LTX-2 generation reaches transformer loading but fails before any frames are generated.

The latest local Mini Studio job failed with:

ValueError: Received 1680 parameters not in model:
audio_caption_projection.linear_1.bias,
audio_caption_projection.linear_1.weight,
...
transformer_blocks.0.attn1.to_out.0.weight,
transformer_blocks.0.audio_ff.net.0.proj.weight,
...

Reproduction

1. Start the local Web UI:

mlx_video.webui --server-port 7861

2. Submit an LTX-2 image-to-video job using Lightricks/LTX-2, distilled pipeline, 512x512, 97 frames.
3. Inspect the job log under outputs/logs/<job-id>.log.

The Web UI itself stays alive, but the generation subprocess exits with status 1 and no MP4 appears under outputs/videos/.

Root Cause

LTXModel.sanitize() already contains the key conversions needed by the local MLX module names:

• .linear_1. -> .linear1.
• .linear_2. -> .linear2.
• .to_out.0. -> .to_out.
• .ff.net.0.proj. -> .ff.proj_in.
• .ff.net.2. -> .ff.proj_out.
• audio FF equivalents

However, those conversions currently run only when the incoming checkpoint keys have the raw model.diffusion_model. prefix.

The Hugging Face Lightricks/LTX-2 transformer safetensors are already unprefixed, but still use Diffusers-style submodule names. Because sanitize() returns unprefixed weights unchanged, load_weights(strict=True) sees 1680 unexpected parameters.

Expected Behavior

Unprefixed Diffusers-style transformer keys should be normalized before strict MLX loading, just like prefixed raw checkpoint keys.

Local Evidence

• Web UI: http://127.0.0.1:7861/ returns 200 OK.
• Latest failed job: 0ebf34d8d38a.
• Latest log: outputs/logs/0ebf34d8d38a.log.
• No generated MP4 files under outputs/videos/.

[andyelka-creator]

Opened draft PR #31 with a focused fix for this issue.

Validation performed locally:

• Added regression coverage for unprefixed Diffusers transformer key normalization.
• Compared the real downloaded Lightricks/LTX-2 transformer checkpoint after sanitizer against the MLX model parameter tree: 3510 sanitized keys, 3510 model keys, extra 0, missing 0.
• Verified LTXModel.from_pretrained(.../transformer, strict=True) completes with strict transformer load ok.
• Relevant pytest subset: 18 passed, 2 warnings.

PR: #31