The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Various fixes for broken things from earlier PR. (#9168)
comfyanonymous7991341e89ca
diff --git a/comfy/model_base.py b/comfy/model_base.pyindex 3a9c031ea..f9591f292 100644--- a/comfy/model_base.py+++ b/comfy/model_base.py@@ -401,7 +401,7 @@ class SD21UNCLIP(BaseModel):unclip_conditioning = kwargs.get("unclip_conditioning", None)device = kwargs["device"]if unclip_conditioning is None:- return torch.zeros((1, self.adm_channels))+ return torch.zeros((1, self.adm_channels), device=device)else:return unclip_adm(unclip_conditioning, device, self.noise_augmentor, kwargs.get("unclip_noise_augment_merge", 0.05), kwargs.get("seed", 0) - 10)@@ -409,7 +409,7 @@ def sdxl_pooled(args, noise_augmentor):if "unclip_conditioning" in args:return unclip_adm(args.get("unclip_conditioning", None), args["device"], noise_augmentor, seed=args.get("seed", 0) - 10)[:,:1280]else:- return args["pooled_output"]+ return args["pooled_output"].to(device=args["device"])class SDXLRefiner(BaseModel):def __init__(self, model_config, model_type=ModelType.EPS, device=None):@@ -615,9 +615,11 @@ class IP2P:if image is None:image = torch.zeros_like(noise)+ else:+ image = image.to(device=device)if image.shape[1:] != noise.shape[1:]:- image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")+ image = utils.common_upscale(image, noise.shape[-1], noise.shape[-2], "bilinear", "center")image = utils.resize_to_batch_size(image, noise.shape[0])return self.process_ip2p_image_in(image)@@ -696,7 +698,7 @@ class StableCascade_B(BaseModel):#size of prior doesn't really matter if zeros because it gets resized but I still want it to get batchedprior = kwargs.get("stable_cascade_prior", torch.zeros((1, 16, (noise.shape[2] * 4) // 42, (noise.shape[3] * 4) // 42), dtype=noise.dtype, layout=noise.layout, device=noise.device))- out["effnet"] = comfy.conds.CONDRegular(prior)+ out["effnet"] = comfy.conds.CONDRegular(prior.to(device=noise.device))out["sca"] = comfy.conds.CONDRegular(torch.zeros((1,)))return out@@ -1161,10 +1163,10 @@ class WAN21_Vace(WAN21):vace_frames_out = []for j in range(len(vace_frames)):- vf = vace_frames[j].clone()+ vf = vace_frames[j].to(device=noise.device, dtype=noise.dtype, copy=True)for i in range(0, vf.shape[1], 16):vf[:, i:i + 16] = self.process_latent_in(vf[:, i:i + 16])- vf = torch.cat([vf, mask[j]], dim=1)+ vf = torch.cat([vf, mask[j].to(device=noise.device, dtype=noise.dtype)], dim=1)vace_frames_out.append(vf)vace_frames = torch.stack(vace_frames_out, dim=1)