tiny-random/ltx-2-5

This is a randomly initialized tiny model based on Lightricks/LTX-2.5 for testing and debugging.

Feature	Source	This repo
Checkpoint tensor elements	35.03B BF16, 4.73M F32, 32.19M U8	4.68M BF16, 8.45K F32, 32.18M U8
Safetensors weight files	70.12 GB (4 files)	41.65 MB (4 files)
Transformer layers / video inner width	48 / 4,096	4 / 64
Video heads / head dimension	32 / 128	2 / 32
Audio inner width / heads / head dimension	2,048 / 32 / 64	64 / 2 / 32
Video / audio cross-attention width	4,096 / 2,048	64 / 64
Text connector blocks / registers	8 / 128	2 / 4
Text projection factor (layers + 1)	49	3
Text encoder layers / width / head dimension	48 / 3,840 / 256	2 / 8 / 32
Video VAE latent channels / patch size	128 / 4	Same
Audio VAE latent channels / mel bins	8 / 64	8 / 4
Vocoder hidden width	1,024	32
Embedded tokenizer and processor assets	U8 tensors in the text encoder file	Same layout
Stored precision	BF16 weights, FP32 modulation tables	Same

The source is a ComfyUI split single-file pack: the transformer file also carries the text connectors, and the text encoder file carries the connector projections, two modality projectors and the tokenizer assets as U8 tensors. This repo keeps that layout and adds config.json plus ltx2_loader.py, which the split files do not contain, so the reduced components can be rebuilt. LTX-2.5's tokenizer is embedded in a 26 GB file that cannot be fetched alone, so tokenizer_json and the hf_asset__* entries come from the pinned google/gemma-4-12b-it release, which has the same 262,144-row vocabulary. The embedded 32.17 MB tokenizer is the floor for that file; the remaining non-tokenizer weights are about 9.5 MB, so the four safetensors files total 41.65 MB. The text encoder keeps a 32-wide attention head on an 8-wide residual width to preserve the minimum head dimension.

File path	Size
diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors	2.3MB
text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors	36.5MB
vae/ltx-2.5-audio-vae-bf16.safetensors	0.1MB
vae/ltx-2.5-video-vae-bf16.safetensors	2.7MB

Example usage

Diffusers (✅ verified)

ltx2_loader.py maps the split files back to Diffusers components; this runs a 2-step text-to-video request that also decodes the audio branch.

def run_diffusers(model_id: str):
    import sys
    from pathlib import Path

    import torch

    model_id = Path(model_id).resolve()
    sys.path.insert(0, str(model_id))
    from ltx2_loader import load_pipeline

    pipeline = load_pipeline(model_id)
    with torch.no_grad():
        result = pipeline(
            prompt="A tiny red cube rotating on a white table",
            height=64,
            width=64,
            num_frames=9,
            num_inference_steps=2,
            output_type="pt",
            guidance_scale=1.0,
            stg_scale=1.0,
            modality_scale=1.0,
            audio_guidance_scale=1.0,
            audio_stg_scale=1.0,
            audio_modality_scale=1.0,
            generator=torch.Generator(pipeline.device).manual_seed(0),
        )
    frames, audio = result.frames, result.audio
    assert frames.ndim == 5 and frames.shape[0] == 1
    assert audio.shape[0] == 1 and audio.shape[1] == 2
    assert torch.isfinite(frames).all() and torch.isfinite(audio).all()
    print("frames", tuple(frames.shape), "audio", tuple(audio.shape))


if __name__ == "__main__":
    run_diffusers(model_id="tiny-random/ltx-2-5")

Creating this repo

• Code: .meta/create_model.py
• Model structure: .meta/model.log
• Environment: .meta/env.json

Change log

• 2026-11-01: Initial tiny LTX-2.5 split checkpoint with video and audio branches.

This repo is generated by DeepSeek-V4.1-Flash.