Downloads · 30 days
226
26% of all-time downloads
anonymous1436778242134213535654
anonymous1436778242134213535654/Wan2.1-T2V-1.3B-Distill-Comfy
Wan2.1-T2V-1.3B-Distill-Comfy is a text-to-video model from anonymous1436778242134213535654. Use it when you need video from a text prompt.
This model is a straight conversion of the original
Downloads · 30 days
226
26% of all-time downloads
All-time downloads
858
Public
Repo size
10.7 GB
Likes
1
Public
Click a slice to open those files.
.gguf7.9 GB · 73%
From the Hugging Face model README
This model is a straight conversion of the original
<video src="Wan2.1-T2V-1.3B-Distill-Comfy/resolve/main/ComfyUI_00229_.mp4" height="480" controls></video>
Here is the code I used to convert it. Tried it in the latest comfyui and it works. See the included workflow file for an example. On my 3090, and using the lightvaew2_1 autoencoder it takes 25 seconds to run the workflow.
import torch
from safetensors.torch import save_file
SRC = '/home/user/convert model/model_iter6000.pt'
DST = '/home/user/convert model/wan2.1-t2v-1.3b-4step-distill.safetensors'
sd = torch.load(SRC, map_location='cpu')
def conv(k):
if k.startswith('blocks.'):
idx, rest = k[len('blocks.'):].split('.', 1)
if rest.startswith('attn1.'):
s = rest[len('attn1.'):].replace('to_out.0', 'o').replace('to_q', 'q').replace('to_k', 'k').replace('to_v', 'v')
return f'blocks.{idx}.self_attn.{s}'
if rest.startswith('attn2.'):
s = rest[len('attn2.'):].replace('to_out.0', 'o').replace('to_q', 'q').replace('to_k', 'k').replace('to_v', 'v')
return f'blocks.{idx}.cross_attn.{s}'
if rest.startswith('ffn.net.0.proj.'):
return f'blocks.{idx}.ffn.0.' + rest[len('ffn.net.0.proj.'):]
if rest.startswith('ffn.net.2.'):
return f'blocks.{idx}.ffn.2.' + rest[len('ffn.net.2.'):]
if rest.startswith('norm2.'):
return f'blocks.{idx}.norm3.' + rest[len('norm2.'):]
if rest == 'scale_shift_table':
return f'blocks.{idx}.modulation'
return None
if k in ('patch_embedding.weight', 'patch_embedding.bias'):
return k
if k == 'proj_out.weight':
return 'head.head.weight'
if k == 'proj_out.bias':
return 'head.head.bias'
if k == 'scale_shift_table':
return 'head.modulation'
if k.startswith('condition_embedder.time_embedder.linear_1.'):
return 'time_embedding.0.' + k.split('.')[-1]
if k.startswith('condition_embedder.time_embedder.linear_2.'):
return 'time_embedding.2.' + k.split('.')[-1]
if k.startswith('condition_embedder.time_proj.'):
return 'time_projection.1.' + k.split('.')[-1]
if k.startswith('condition_embedder.text_embedder.linear_1.'):
return 'text_embedding.0.' + k.split('.')[-1]
if k.startswith('condition_embedder.text_embedder.linear_2.'):
return 'text_embedding.2.' + k.split('.')[-1]
return None
out = {}
for k, v in sd.items():
nk = conv(k)
if nk is None:
raise SystemExit(f'UNMAPPED: {k}')
out[nk] = v.contiguous()
save_file(out, DST, metadata={'format': 'pt', 'framework': 'pt'})
print(f'Saved {len(out)} tensors -> {DST}')
GGUFs made with city96's stuff
Q3_K_S is for the desperate
Models from Q3_K_M upward start looking 'fine'.
The ggufs work great with comfyui. They are also compatible with leejet's stable-diffusion.cpp, but theyre buggy, video outputs seem messed up somewhat, and ram usage spikes on VAE decode to like 60 gb+ on a 5 second video using the TAE, if you use VAE and tile it its fine and vram usage is low. Might look into it later.