mirrors/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-09-10 11:35:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

Basic Hunyuan Video model support.

Browse Source
This commit is contained in:
comfyanonymous
2024-12-16 19:35:40 -05:00
parent 19ee5d9d8b
commit bda1482a27
18 changed files with 413646 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
comfy/supported_models.py
View File

@@ -12,6 +12,7 @@ import comfy.text_encoders.hydit
import comfy.text_encoders.flux
import comfy.text_encoders.genmo
import comfy.text_encoders.lt
import comfy.text_encoders.hunyuan_video
from . import supported_models_base
from . import latent_formats
@@ -738,6 +739,54 @@ class LTXV(supported_models_base.BASE):
t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
return supported_models_base.ClipTarget(comfy.text_encoders.lt.LTXVT5Tokenizer, comfy.text_encoders.lt.ltxv_te(**t5_detect))
models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV]
class HunyuanVideo(supported_models_base.BASE):
unet_config = {
"image_model": "hunyuan_video",
}
sampling_settings = {
"shift": 7.0,
}
unet_extra_config = {}
latent_format = latent_formats.HunyuanVideo
memory_usage_factor = 2.0 #TODO
supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
vae_key_prefix = ["vae."]
text_encoder_key_prefix = ["text_encoders."]
def get_model(self, state_dict, prefix="", device=None):
out = model_base.HunyuanVideo(self, device=device)
return out
def process_unet_state_dict(self, state_dict):
out_sd = {}
for k in list(state_dict.keys()):
key_out = k
key_out = key_out.replace("txt_in.t_embedder.mlp.0.", "txt_in.t_embedder.in_layer.").replace("txt_in.t_embedder.mlp.2.", "txt_in.t_embedder.out_layer.")
key_out = key_out.replace("txt_in.c_embedder.linear_1.", "txt_in.c_embedder.in_layer.").replace("txt_in.c_embedder.linear_2.", "txt_in.c_embedder.out_layer.")
key_out = key_out.replace("_mod.linear.", "_mod.lin.").replace("_attn_qkv.", "_attn.qkv.")
key_out = key_out.replace("mlp.fc1.", "mlp.0.").replace("mlp.fc2.", "mlp.2.")
key_out = key_out.replace("_attn_q_norm.weight", "_attn.norm.query_norm.scale").replace("_attn_k_norm.weight", "_attn.norm.key_norm.scale")
key_out = key_out.replace(".q_norm.weight", ".norm.query_norm.scale").replace(".k_norm.weight", ".norm.key_norm.scale")
key_out = key_out.replace("_attn_proj.", "_attn.proj.")
key_out = key_out.replace(".modulation.linear.", ".modulation.lin.")
key_out = key_out.replace("_in.mlp.2.", "_in.out_layer.").replace("_in.mlp.0.", "_in.in_layer.")
out_sd[key_out] = state_dict[k]
return out_sd
def process_unet_state_dict_for_saving(self, state_dict):
replace_prefix = {"": "model.model."}
return utils.state_dict_prefix_replace(state_dict, replace_prefix)
def clip_target(self, state_dict={}):
# pref = self.text_encoder_key_prefix[0]
# t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideoTokenizer, comfy.text_encoders.hunyuan_video.hunyuan_video_clip()) #TODO
models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo]
models += [SVD_img2vid]