license: apache-2.0
base_model:
- Lightricks/LTX-Video
library_name: gguf
tags:
- video
- video-generation
pipeline_tag: image-to-video
Comfyui doesnt natively support these Quants yet but there is a workaround:
First of all you need to load the vae, which you can download in this repo too and then
we need to edit this file Comfyui/comfy/ldm/lightricks/model.py
and change the init function for class LTXVModel(torch.nn.Module):
from this:
def __init__(self,
in_channels=128,
cross_attention_dim=2048,
attention_head_dim=64,
num_attention_heads=32,
caption_channels=4096,
num_layers=28,
positional_embedding_theta=10000.0,
positional_embedding_max_pos=[20, 2048, 2048],
causal_temporal_positioning=False,
vae_scale_factors=(8, 32, 32),
dtype=None, device=None, operations=None, **kwargs):
to this:
def __init__(self,
in_channels=128,
cross_attention_dim=4096,
attention_head_dim=128,
num_attention_heads=32,
caption_channels=4096,
num_layers=48,
positional_embedding_theta=10000.0,
positional_embedding_max_pos=[20, 2048, 2048],
causal_temporal_positioning=False,
vae_scale_factors=(8, 32, 32),
dtype=None, device=None, operations=None, **kwargs):
If you want to use the 2b again just revert it and restart. After a restart the 13b model should work with this and the vae i uploaded.
This is a direct GGUF conversion of Lightricks/ltxv-13b-0.9.7-dev
All quants are created from the FP32 base file, though I only uploaded the Q8_0 and less, if you want the F16 or BF16 one I would upload it per request.
The model files can be used with the ComfyUI-GGUF custom node.
Place model files in ComfyUI/models/unet - see the GitHub readme for further install instructions.
Please refer to this chart for a basic overview of quantization types.
For conversion I used the conversion scripts from city96