Lightricks

/

LTX-2

like 718

Follow

Lightricks 1.75k

This PR adds diffusers-compatible model checkpoints for the LTX 2.0 model.

The corresponding diffusers PR is https://github.com/huggingface/diffusers/pull/12915. With that PR checked out, you can generate T2V samples with

python scripts/ltx2_test_full_pipeline.py \
    --model_id Lightricks/LTX-2 \
    --revision refs/pr/3 \
    --cpu_offload

and I2V samples with

python scripts/ltx2_test_full_pipeline_i2v.py \
    --model_id Lightricks/LTX-2 \
    --revision refs/pr/3 \
    --image_path https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg \
    --cpu_offload

This seems quite critical to the claimed "Diffusers Support" in https://github.com/huggingface/diffusers/pull/12915. How does that PR claim to be tested without this change? Or are they tethered and both are required?

If you have that PR checked out, you can perform the checkpoint conversion yourself with

python scripts/convert_ltx2_to_diffusers.py \
    --full_pipeline \
    --original_state_dict_repo_id Lightricks/LTX-2 \
    --combined_filename ltx-2-19b-dev.safetensors \
    --output_path /path/to/diffusers/model/

You can then test the PR using the converted checkpoint with for example

python scripts/ltx2_test_full_pipeline_i2v.py \
    --model_id /path/to/diffusers/model/ \
    --image_path https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg \
    --cpu_offload

This PR contains the already converted checkpoint from scripts/convert_ltx2_to_diffusers.py.

LGTM now, will verify again when Diffusers PR is merged.