LTX-2.5

The LTX-2.5 backend supports Stage0.5 native rectified-flow LoRA training, inference, and preencoding for 153-frame 512×768 video.

Availability and data

Route	Data	Status
Stage0.5 preencoded	ltx-153f-h512-w768 latent	Code and weights available; latent upload coming soon
Stage0.5 online	raw-WDS	Available
Inference	ltx-153f-h512-w768 latent	Code and weights available; latent upload coming soon
Preencoding	raw-WDS input	Available

Preencoded training is the recommended route because it avoids running the VAE and text encoder during every training step. Track the payload in the latent-WDS release list, or use Download and access for raw-WDS.

Setup

Activate the LTX environment from Runtime environments, then set:

export SOLAR_REPO=/path/to/SolarWM
export SOLAR_MODEL_ROOT=/path/to/SolarWM-models
export SOLAR_DATA_ROOT=/path/to/SolarWM-Data/releases-v1
export SOLAR_OUTPUT_ROOT=/path/to/outputs
cd "$SOLAR_REPO"

Download the LTX base model and released adapter:

python -m pip install --upgrade huggingface_hub
hf download junchaoh-cs/SolarWM \
  --include "SolarWM-ltx-22B-*/**" \
  --local-dir "$SOLAR_MODEL_ROOT"

Set the model paths used below:

export LTX_BASE="$SOLAR_MODEL_ROOT/SolarWM-ltx-22B-base"
export LTX_TRANSFORMER="$LTX_BASE/diffusion_models/ltx-2.5-22b-dev-transformer-bf16.safetensors"
export LTX_VAE="$LTX_BASE/vae/ltx-2.5-video-vae-bf16.safetensors"
export LTX_GEMMA="$LTX_BASE/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors"
export LTX_NEGATIVE="$LTX_BASE/conditioning/ltx25-negative-caption.safetensors"

Preencoded training

The released training profile uses two eight-GPU nodes. Set NODE_RANK=0 on the first node and NODE_RANK=1 on the second.

export NODE_RANK=0
export MASTER_ADDR=hostname-or-ip-of-node-0
export MASTER_PORT=29500

torchrun --nnodes=2 --node-rank="$NODE_RANK" --nproc-per-node=8 \
  --rdzv-backend=c10d --rdzv-endpoint="$MASTER_ADDR:$MASTER_PORT" \
  -m solarwm train \
  --config configs/examples/ltx25/stage0p5-train-153f-lora384-sp2.yaml \
  --set distributed.world_size=16 \
  --set distributed.local_world_size=8 \
  --set train.global_batch_size=16 \
  --set validation.sample_count=16 \
  --set model.checkpoint_path="$LTX_TRANSFORMER" \
  --set model.codec.video_vae_path="$LTX_VAE" \
  --set data.index_root="$SOLAR_DATA_ROOT" \
  --set data.transport.root="$SOLAR_DATA_ROOT" \
  --set validation.inference.negative_caption_cache="$LTX_NEGATIVE" \
  --set runtime.output_dir="$SOLAR_OUTPUT_ROOT/ltx-stage0p5-153f"

Run solarwm config resolve with the same config and --set arguments before training if you want to inspect the resolved configuration.

New LTX checkpoints restore all reader and RNG state. Older checkpoints remain loadable, but lack the Python and NumPy RNG state needed for an exact resume.

Inference

Inference uses two GPUs and the released adapter:

torchrun --standalone --nproc-per-node=2 -m solarwm infer \
  --config configs/examples/ltx25/stage0p5-infer-153f.yaml \
  --set model.checkpoint_path="$LTX_TRANSFORMER" \
  --set model.codec.video_vae_path="$LTX_VAE" \
  --set model.adapter_checkpoint_path="$SOLAR_MODEL_ROOT/SolarWM-ltx-22B-bid-stage0p5-153f" \
  --set data.index_root="$SOLAR_DATA_ROOT" \
  --set data.transport.root="$SOLAR_DATA_ROOT" \
  --set inference.negative_caption_cache="$LTX_NEGATIVE" \
  --set runtime.output_dir="$SOLAR_OUTPUT_ROOT/ltx-stage0p5-153f-infer"

Optional raw-WDS workflows

Online training uses configs/examples/ltx25/stage0p5-train-online-153f-unpaired.yaml with the same training command. Add --set model.codec.gemma4_path="$LTX_GEMMA" and choose a new output directory.

To create a new LTX latent generation from raw-WDS:

torchrun --standalone --nproc-per-node=8 -m solarwm preencode \
  --config configs/examples/ltx25/preencode-153f.yaml \
  --set model.checkpoint_path="$LTX_TRANSFORMER" \
  --set model.codec.video_vae_path="$LTX_VAE" \
  --set model.codec.gemma4_path="$LTX_GEMMA" \
  --set data.index_root="$SOLAR_DATA_ROOT" \
  --set data.transport.root="$SOLAR_DATA_ROOT" \
  --set preencode.output_root="$SOLAR_OUTPUT_ROOT/preencoded/ltx-153f-h512-w768" \
  --set runtime.output_dir="$SOLAR_OUTPUT_ROOT/ltx-preencode-153f"

Model and camera conventions

Raw records store camera poses as C2W. LTX preencoding converts them to first-frame-relative W2C with normalized intrinsics, which is also the camera format stored in the released LTX latent generation. Released LTX configs use the linear camera-translation transform.

License

The LTX base model and released adapter remain subject to the LTX Community License included with the model packages. Review that license before use or redistribution; the SolarWM code license does not replace it.