|
|
from ldm_patched.modules.model_base import BaseModel |
|
|
from ldm_patched.modules import model_management |
|
|
from ldm_patched.modules.sd import CLIPType |
|
|
import ldm_patched.modules.utils |
|
|
|
|
|
class SD3ModelConfig: |
|
|
def __init__(self): |
|
|
self.name = 'SD3' |
|
|
self.unet_config = { |
|
|
"in_channels": 16, |
|
|
"pos_embed_scaling_factor": None, |
|
|
} |
|
|
self.latent_format = "SD3" |
|
|
|
|
|
def get_model(self, state_dict, prefix="model.diffusion_model.", device=None): |
|
|
if device is None: |
|
|
device = model_management.get_torch_device() |
|
|
shift = 3.0 |
|
|
return BaseModel(shift=shift, state_dict=state_dict, prefix=prefix, device=device) |
|
|
|
|
|
def process_clip_state_dict(self, state_dict): |
|
|
clip_l = ldm_patched.modules.utils.state_dict_prefix_replace(state_dict, {"clip_l.": ""}) |
|
|
clip_g = ldm_patched.modules.utils.state_dict_prefix_replace(state_dict, {"clip_g.": ""}) |
|
|
t5xxl = ldm_patched.modules.utils.state_dict_prefix_replace(state_dict, {"t5xxl.": ""}) |
|
|
return {"clip_l": clip_l, "clip_g": clip_g, "t5xxl": t5xxl} |
|
|
|
|
|
def clip_target(self, state_dict=None): |
|
|
from ldm_patched.modules.text_encoders import sd3_clip |
|
|
clip_l = sd3_clip.SD1ClipModel(layer="hidden", layer_idx=-2) |
|
|
clip_g = sd3_clip.SDXLClipG() |
|
|
t5xxl = sd3_clip.T5XXLModel() |
|
|
return (clip_l, clip_g, t5xxl) |
