threestudio/models/guidance/stable_diffusion_unified_guidance.py [621:674]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                elevation,
                azimuth,
                camera_distances,
                camera_condition,
            )

        if self.cfg.weighting_strategy == "dreamfusion":
            w = (1.0 - self.alphas[t]).view(-1, 1, 1, 1)
        elif self.cfg.weighting_strategy == "uniform":
            w = 1.0
        elif self.cfg.weighting_strategy == "fantasia3d":
            w = (self.alphas[t] ** 0.5 * (1 - self.alphas[t])).view(-1, 1, 1, 1)
        else:
            raise ValueError(
                f"Unknown weighting strategy: {self.cfg.weighting_strategy}"
            )

        grad = w * (eps_pretrain - eps_phi)

        if self.grad_clip_val is not None:
            grad = grad.clamp(-self.grad_clip_val, self.grad_clip_val)

        # reparameterization trick:
        # d(loss)/d(latents) = latents - target = latents - (latents - grad) = grad
        target = (latents - grad).detach()
        loss_sd = 0.5 * F.mse_loss(latents, target, reduction="sum") / batch_size

        guidance_out = {
            "loss_sd": loss_sd,
            "grad_norm": grad.norm(),
            "timesteps": t,
            "min_step": self.min_step,
            "max_step": self.max_step,
            "latents": latents,
            "latents_1step_orig": latents_1step_orig,
            "rgb": rgb_BCHW.permute(0, 2, 3, 1),
            "weights": w,
            "lambdas": self.lambdas[t],
        }

        if self.cfg.return_rgb_1step_orig:
            with torch.no_grad():
                rgb_1step_orig = self.vae_decode(
                    self.pipe.vae, latents_1step_orig
                ).permute(0, 2, 3, 1)
            guidance_out.update({"rgb_1step_orig": rgb_1step_orig})

        if self.cfg.return_rgb_multistep_orig:
            with self.set_scheduler(
                self.pipe,
                DPMSolverSinglestepScheduler,
                solver_order=1,
                num_train_timesteps=int(t[0]),
            ) as pipe:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



threestudio/models/guidance/zero123_unified_guidance.py [618:671]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                elevation,
                azimuth,
                camera_distances,
                camera_condition,
            )

        if self.cfg.weighting_strategy == "dreamfusion":
            w = (1.0 - self.alphas[t]).view(-1, 1, 1, 1)
        elif self.cfg.weighting_strategy == "uniform":
            w = 1.0
        elif self.cfg.weighting_strategy == "fantasia3d":
            w = (self.alphas[t] ** 0.5 * (1 - self.alphas[t])).view(-1, 1, 1, 1)
        else:
            raise ValueError(
                f"Unknown weighting strategy: {self.cfg.weighting_strategy}"
            )

        grad = w * (eps_pretrain - eps_phi)

        if self.grad_clip_val is not None:
            grad = grad.clamp(-self.grad_clip_val, self.grad_clip_val)

        # reparameterization trick:
        # d(loss)/d(latents) = latents - target = latents - (latents - grad) = grad
        target = (latents - grad).detach()
        loss_sd = 0.5 * F.mse_loss(latents, target, reduction="sum") / batch_size

        guidance_out = {
            "loss_sd": loss_sd,
            "grad_norm": grad.norm(),
            "timesteps": t,
            "min_step": self.min_step,
            "max_step": self.max_step,
            "latents": latents,
            "latents_1step_orig": latents_1step_orig,
            "rgb": rgb_BCHW.permute(0, 2, 3, 1),
            "weights": w,
            "lambdas": self.lambdas[t],
        }

        if self.cfg.return_rgb_1step_orig:
            with torch.no_grad():
                rgb_1step_orig = self.vae_decode(
                    self.pipe.vae, latents_1step_orig
                ).permute(0, 2, 3, 1)
            guidance_out.update({"rgb_1step_orig": rgb_1step_orig})

        if self.cfg.return_rgb_multistep_orig:
            with self.set_scheduler(
                self.pipe,
                DPMSolverSinglestepScheduler,
                solver_order=1,
                num_train_timesteps=int(t[0]),
            ) as pipe:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -