Comfy-Org/ComfyUI · error · ValueError

SeedVR2 VAE convolution requires an explicit MemoryState.

Error message

SeedVR2 VAE convolution requires an explicit MemoryState.

What it means

SeedVR2 VAE convolutions are streaming-aware: their forward requires an explicit MemoryState enum (ACTIVE for streaming with cache, or another state for non-streaming) because cache lifecycle management differs per state. The default parameter is MemoryState.UNSET, and any call that did not deliberately choose a state raises immediately — this is an API-contract guard, not a runtime data problem.

Source

Thrown at comfy/ldm/seedvr/vae.py:581

                x[idx],
                split_dim=split_dim + 1,
                padding=padding,
                prev_cache=cache
            )

            cache = next_cache

        output = torch.cat(x, dim=split_dim)
        return output

    def forward(
        self,
        input,
        memory_state: MemoryState = MemoryState.UNSET,
        memory_cache = None,
    ) -> Tensor:
        if memory_state == MemoryState.UNSET:
            raise ValueError("SeedVR2 VAE convolution requires an explicit MemoryState.")
        if memory_cache is None:
            memory_cache = {}
        if memory_state != MemoryState.ACTIVE:
            memory_cache.pop(self, None)
        if (
            math.isinf(self.memory_limit)
            and torch.is_tensor(input)
        ):
            return self.basic_forward(input, memory_state, memory_cache)
        return self.slicing_forward(input, memory_state, memory_cache)

    def basic_forward(self, input: Tensor, memory_state: MemoryState = MemoryState.UNSET, memory_cache = None):
        mem_size = self.stride[0] - self.kernel_size[0]
        memory = memory_cache.get(self) if memory_cache is not None else None
        if (memory is not None) and (memory_state == MemoryState.ACTIVE):
            input = extend_head(input, memory=memory, times=-1)
        else:
            input = extend_head(input, times=self.temporal_padding * 2)

View on GitHub (pinned to 1c6d8d45b3)

Solutions

Pass an explicit memory_state, e.g. memory_state=MemoryState.BASIC (non-cached) for a one-shot call, or ACTIVE within the VAE's own streaming loop.
Prefer calling the VAE's public encode/decode rather than individual conv modules.
If porting the module, thread memory_state through your call chain the same way SeedVR2 VAE code does.
Check the MemoryState enum values in comfy/ldm/seedvr/vae.py to pick the right state.

Example fix

# before
y = conv_block(x)
# after
from comfy.ldm.seedvr.vae import MemoryState
y = conv_block(x, memory_state=MemoryState.BASIC)

Defensive patterns

Strategy: validation

Validate before calling

from comfy.ldm.seedvr.vae import MemoryState

def call_conv(conv, x, streaming=False):
    state = MemoryState.ACTIVE if streaming else MemoryState.BASIC
    return conv(x, memory_state=state, memory_cache={})

Type guard

def is_valid_memory_state(s) -> bool:
    from comfy.ldm.seedvr.vae import MemoryState
    return s in MemoryState._value2member_map_

Prevention

Always pass an explicit memory_state when calling SeedVR2 VAE conv modules directly.
Call the VAE's public encode/decode instead of individual conv blocks.
Thread memory_state through custom wrappers — never rely on the default.

When it happens

Trigger: Calling SeedVR2VaeConv forward directly (custom code, ported modules) without memory_state; wrapping the VAE in a tool that re-invokes submodules with default args; calling basic_forward-adjacent paths that forget the parameter.

Common situations: Custom node or research code reusing SeedVR2 VAE blocks; copying the module into another codebase and calling conv(x) idiomatically; migration from a diffusers-style API that had no memory state concept.

Related errors

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/9c86b8be915575ad. Report an issue: GitHub.