wm_ltx

wm-serving Infer Ready

This repo is a slimmed inference-only export for internal visual quality testing.

Included

• packages/ltx-core
• packages/ltx-pipelines
• chunkwise_overlap8_blend_async_decode.py
• scripts/
• copied test_data/

Reused From Existing Repo

• checkpoints -> /mnt/data04/144632/[email protected]/world/wm-serving/checkpoints
• distilled LoRA default: /mnt/data04/144632/[email protected]/LTX-2/LTX-2/ltx-2-19b-distilled-lora-384.safetensors

Default Model Choices

• New DIT: checkpoints/0517-sfpp-4step-distil_generator_weights_step_04000_merged.safetensors
• Fine-tuned decoder: checkpoints/ltx2_vae_decoder_tune_lpips0.01_fdl10.1_lr1e-6_ckpt10800_vae_only.ckpt
• Spatial upsampler: checkpoints/latent_upsampler

Environment

cd /mnt/data04/144632/[email protected]/world/wm-serving-infer-ready
. env.sh

env.sh is copied from the original repo and only handles repo-local dependency setup:

uv sync --frozen
source .venv/bin/activate

This ready repo does not add an extra CUDA / NVIDIA installer script.

Three 1024 Modes

All three scripts default to:

• HEIGHT=1024
• WIDTH=1536
• copied test_data/
• new DIT
• 10800 VAE decoder

1. One-stage 1024

bash scripts/infer_ti2vid_one_stage_1024.sh

2. Two-stage refine 1024

Flow: stage1 DIT -> upsampler -> stage2 refine -> 10800 decoder

bash scripts/infer_ti2vid_two_stage_refine_1024.sh

3. Two-stage direct decoder 1024

Flow: stage1 DIT -> upsampler -> 10800 decoder

bash scripts/infer_ti2vid_two_stage_ft_decoder_1024.sh

Batch Comparison

Run all three variants on the copied test_data/ cases:

bash scripts/batch_compare_1024_variants.sh

Outputs will be written under:

results/batch_compare_1024_<timestamp>/

Notes

• scripts/batch_run_ti2vid_cases.sh now defaults to scripts/infer_ti2vid_one_stage_1024.sh
• Override any default with environment variables, for example:

GPU_ID=1 IMAGE_PATH=test_data/image/zelda_water.png \
OUTPUT_PATH=results/custom/zelda_water.mp4 \
bash scripts/infer_ti2vid_two_stage_ft_decoder_1024.sh