LTX-2.3-distilled-1.1-fp8

This is an FP8 quantization of the LTX 2.3 Distilled 1.1 model located at https://huggingface.co/Lightricks/LTX-2.3

The model was quantized using the script shown below. The intent was to quantize only the diffusion model tensors, leaving other things (like the VAE) in full precision. This may or may not have been the right approach but it made sense at the time and the model produces the expected results. AI assistance was used in creating the script.

This model was quantized and hosted for my own convenience. If it is useful for others, that's great. I am unable to provide any information or assistance beyond what is posted here.

Quantization script:

from safetensors import safe_open
from safetensors.torch import save_file
import torch
import gc

tensors = {}
with safe_open("ltx-2.3-22b-distilled-1.1.safetensors", framework="pt", device="cpu") as f:
    tensor_num = 1
    tensor_total = len(f.keys())
    metadata = f.metadata()
    for key in f.keys():
        print(f"Processing {key} (tensor {tensor_num}/{tensor_total})...")
        tensor = f.get_tensor(key)          # BF16 tensor loaded
        if key.startswith("model.diffusion_model"):
            tensors[key] = tensor.float().to(torch.float8_e4m3fn)
        else:
            tensors[key] = tensor  # keep original BF16
        del tensor
        gc.collect()                         # don't wait for GC to catch up
        tensor_num += 1

save_file(tensors, "ltx-2.3-22b-distilled-1.1-fp8.safetensors", metadata=metadata)