What causes "NotImplementedError" in Stability-AI/generative-models?

Pass time_context=None (omit it) and rely on the model's internal spatial/temporal context handling.

Stability-AI/generative-models · error · NotImplementedError

NotImplementedError

Error message

NotImplementedError

What it means

VideoDepthModel/VideoUNet.forward raises NotImplementedError when time_context is passed: the forward path in this variant does not implement separate time-context injection. Passing a non-None time_context therefore aborts immediately as an explicitly unsupported code path.

Source

Thrown at sgm/modules/diffusionmodules/video_model.py:582

            get_alpha,
            merge_strategy,
            self.mix_factor,
            apply_sigmoid=apply_sigmoid_to_merge,
        )

    def forward(
        self,
        x: th.Tensor,
        context: Optional[th.Tensor] = None,
        # cam: Optional[th.Tensor] = None,
        time_context: Optional[th.Tensor] = None,
        timesteps: Optional[int] = None,
        image_only_indicator: Optional[th.Tensor] = None,
        conv_view: Optional[th.Tensor] = None,
        conv_motion: Optional[th.Tensor] = None,
    ):
        if time_context is not None:
            raise NotImplementedError

        _, _, h, w = x.shape
        if exists(context):
            context = rearrange(context, "b t ... -> (b t) ...")
        if self.use_spatial_context:
            time_context = repeat(context[:, 0], "b ... -> (b n) ...", n=h * w)

        x = super().forward(
            x,
        )

        x = rearrange(x, "b c h w -> b (h w) c")
        x_mix = x

        num_frames = th.arange(timesteps, device=x.device)
        num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
        num_frames = rearrange(num_frames, "b t -> (b t)")
        t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False)

View on GitHub (pinned to e8cd657656)

Solutions

Pass time_context=None (omit it) and rely on the model's internal spatial/temporal context handling.
If you need time-context conditioning, use a model variant whose forward implements it, or patch forward to consume time_context.
Restructure so temporal information is folded into `context` instead of `time_context`.

Example fix

// before
out = model(x, t, context=ctx, time_context=t_ctx)
// after
out = model(x, t, context=ctx)  # time_context not supported by this forward

Defensive patterns

Strategy: try-catch

Validate before calling

if time_context is not None:
    raise TypeError("This forward does not support time_context; fold it into context")

Try / catch

try:
    out = model(x, t, context=ctx, time_context=t_ctx)
except NotImplementedError:
    logger.warning("time_context unsupported; retrying without it")
    out = model(x, t, context=ctx)

Prevention

Check the forward signature of the exact model class before passing optional context tensors.
Gate time_context usage behind a feature check/version check of the codebase.
Merge temporal conditioning into `context` rather than passing time_context.

When it happens

Trigger: Calling model(x, timesteps, context=..., time_context=...) with a real tensor for time_context, e.g. reusing a call signature from a different video model version that supported split spatial/temporal context.

Common situations: Migrating code from an older/other SGM fork where time_context was supported, plugging a text/temporal encoder output into time_context, or a wrapper passing time_context unconditionally (even None it's fine; only non-None triggers).

Related errors

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/affb9d5481aa9484. Report an issue: GitHub.