ERROR] Error in Phase 1 (Encoding): Tensor type unknown to einops <class 'tuple'>

[jason-mightynice]

Think I have had this error before, but forget how to fix it. Any hints?

███████╗███████╗███████╗██████╗ ██╗ ██╗██████╗ ██████╗ ███████╗
██╔════╝██╔════╝██╔════╝██╔══██╗██║ ██║██╔══██╗ ╚════██╗ ██╔════╝
███████╗█████╗ █████╗ ██║ ██║██║ ██║██████╔╝ █████╔╝ ███████╗
╚════██║██╔══╝ ██╔══╝ ██║ ██║╚██╗ ██╔╝██╔══██╗ ██╔═══╝ ╚════██║
███████║███████╗███████╗██████╔╝ ╚████╔╝ ██║ ██║ ███████╗ ██╗ ███████║
╚══════╝╚══════╝╚══════╝╚═════╝ ╚═══╝ ╚═╝ ╚═╝ ╚══════╝ ╚═╝ ╚══════╝
v2.5.24 © ByteDance Seed · NumZ · AInVFX
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[11:16:45.505] ℹ️ OS: Windows (10.0.26200) | GPU: NVIDIA GeForce RTX 5090 (32GB)
[11:16:45.505] ℹ️ Python: 3.13.6 | PyTorch: 2.10.0+cu130 | FlashAttn: v2 ✓ | SageAttn: v2 ✓ | Triton: ✓
[11:16:45.505] ℹ️ CUDA: 13.0 | cuDNN: 91200 | ComfyUI: 0.19.3
[11:16:45.505]
[11:16:45.505] ━━━━━━━━━ Model Preparation ━━━━━━━━━
[11:16:45.505] 📊 Before model preparation:
[11:16:45.506] 📊 [VRAM] 0.01GB allocated / 0.03GB reserved / Peak: 0.01GB / 30.15GB free / 31.84GB total
[11:16:45.506] 📊 [RAM] 2.93GB process / 11.22GB others / 81.60GB free / 95.75GB total
[11:16:45.506] 📊 Resetting VRAM peak memory statistics
[11:16:45.506] 📥 Checking and downloading models if needed...
[11:16:45.506] 🔧 DiT model found: D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\models\SEEDVR2\seedvr2_ema_7b-Q4_K_M.gguf
[11:16:45.506] 🔧 DiT model already validated (cache): D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\models\SEEDVR2\seedvr2_ema_7b-Q4_K_M.gguf
[11:16:45.507] 🔧 VAE model found: D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\models\SEEDVR2\ema_vae_fp16.safetensors
[11:16:45.508] 🔧 VAE model already validated (cache): D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\models\SEEDVR2\ema_vae_fp16.safetensors
[11:16:45.508] 🔧 Generation context initialized: DiT=cuda:0, VAE=cuda:0, Offload=[DiT offload=cpu, VAE offload=cpu, Tensor offload=cpu], LOCAL_RANK=0
[11:16:45.508] 🎯 Unified compute dtype: torch.bfloat16 across entire pipeline for maximum performance
[11:16:45.509] 🏃 Configuring inference runner...
[11:16:45.509] 🏃 Creating new runner: DiT=seedvr2_ema_7b-Q4_K_M.gguf, VAE=ema_vae_fp16.safetensors
[11:16:45.523] 🚀 Creating DiT model structure on meta device
[11:16:45.635] 🎨 Creating VAE model structure on meta device
[11:16:45.672] 🎨 VAE downsample factors configured (spatial: 8x, temporal: 4x)
[11:16:45.675] 🔄 Moving text_pos_embeds from CPU to CUDA:0 (DiT inference)
[11:16:45.676] 🔄 Moving text_neg_embeds from CPU to CUDA:0 (DiT inference)
[11:16:45.676] 🚀 Loaded text embeddings for DiT
[11:16:45.677] 📊 After model preparation:
[11:16:45.677] 📊 [VRAM] 0.01GB allocated / 0.03GB reserved / Peak: 0.01GB / 30.15GB free / 31.84GB total
[11:16:45.677] 📊 [RAM] 2.93GB process / 11.22GB others / 81.60GB free / 95.75GB total
[11:16:45.677] 📊 Resetting VRAM peak memory statistics
[11:16:45.677] ⚡ Model preparation: 0.17s
[11:16:45.677] ⚡ └─ Model structures prepared: 0.15s
[11:16:45.677] ⚡ └─ DiT structure created: 0.10s
[11:16:45.677] ⚡ └─ VAE structure created: 0.04s
[11:16:45.677] ⚡ └─ Config loading: 0.01s
[11:16:45.678] 🔧 Initializing video transformation pipeline for 1920px (shortest edge)
[11:16:45.793] 🔧 Target dimensions: 1920x1920 (no padding needed)
[11:16:45.798]
[11:16:45.798] 🎬 Starting upscaling generation...
[11:16:45.798] 🎬 Input: 65 frames, 1024x1024px → Output: 1920x1920px (shortest edge: 1920px)
[11:16:45.798] 🎬 Batch size: 65 (uniform), Temporal overlap: 2, Seed: 42, Channels: RGB
[11:16:45.798]
[11:16:45.798] ━━━━━━━━ Phase 1: VAE encoding ━━━━━━━━
[11:16:45.799] ♻️ Reusing pre-initialized video transformation pipeline
[11:16:45.799] 🎨 Materializing VAE weights to CPU (offload device): D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\models\SEEDVR2\ema_vae_fp16.safetensors
[11:16:45.801] 🎯 Converting VAE weights to torch.bfloat16 during loading
[11:16:46.261] 🎨 Materializing VAE: 250 parameters, 478.07MB total
[11:16:46.265] 🎨 VAE materialized directly from meta with loaded weights
[11:16:46.265] 🎨 VAE model set to eval mode (gradients disabled)
[11:16:46.266] 🎨 Configuring VAE causal slicing for temporal processing
[11:16:46.266] 🎨 Configuring VAE memory limits for causal convolutions
[11:16:46.267] 🎯 Model precision: VAE=torch.bfloat16, compute=torch.bfloat16
FETCH ComfyRegistry Data: 85/140
[11:16:46.269] 🎨 Using seed: 1000042 (VAE uses seed+1000000 for deterministic sampling)
[11:16:46.270] 🔄 Moving VAE from CPU to CUDA:0 (inference requirement)
[11:16:46.481] 📊 After VAE loading for encoding:
[11:16:46.481] 📊 [VRAM] 0.48GB allocated / 0.50GB reserved / Peak: 0.48GB / 29.68GB free / 31.84GB total
[11:16:46.481] 📊 [RAM] 2.93GB process / 11.22GB others / 81.60GB free / 95.75GB total
[11:16:46.481] 📊 Memory changes: VRAM +0.47GB
[11:16:46.481] 📊 Resetting VRAM peak memory statistics
[11:16:46.481] 🎨 Encoding batch 1/1
[11:16:46.482] 🔄 Moving video_batch_1 from CPU to CUDA:0, torch.float32 → torch.bfloat16 (VAE encoding)
[11:16:46.775] 📹 Sequence of 65 frames
[11:16:46.897] 🎨 Using VAE tiled encoding (Tile: (1024, 1024), Overlap: (128, 128))
[11:16:46.897] 🎨 Encoding 4 tiles (Tile: (1024, 1024), Overlap: (128, 128))
[11:16:46.906] 🎨 Encoding tiles 1-4 / 4
[11:16:47.526] ❌ [ERROR] Error in Phase 1 (Encoding): Tensor type unknown to einops <class 'tuple'>
[11:16:47.527] 🔄 Moving VAE from CUDA:0 to CPU (VAE offload)
[11:16:47.637] ✅ Cleared 21 VAE memory buffers
[11:16:47.637] 🧹 Starting full cleanup
[11:16:47.637] 🧹 Cleaning up DiT components
[11:16:47.639] ✅ Cleared 36 RoPE LRU caches
[11:16:47.639] 🧹 DiT on meta device - keeping structure for cache
[11:16:47.642] 🧹 DiT model deleted
[11:16:47.643] 🧹 Cleaning up VAE components
[11:16:47.644] 🧹 VAE model deleted
[11:16:47.644] 🧹 Clearing memory caches (deep)...
[11:16:48.262] ✅ Completed full cleanup
[11:16:48.262] 🧹 Cleaned up text embeddings: texts_pos, texts_neg
!!! Exception during processing !!! Tensor type unknown to einops <class 'tuple'>
Traceback (most recent call last):
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\execution.py", line 534, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\execution.py", line 334, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\execution.py", line 308, in async_map_node_over_list
await process_inputs(input_dict, i)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\execution.py", line 296, in process_inputs
result = f(**inputs)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\comfy_api\internal_init.py", line 149, in wrapped_func
return method(locked_class, **inputs)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\comfy_api\latest_io.py", line 1789, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\interfaces\video_upscaler.py", line 580, in execute
raise e
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\interfaces\video_upscaler.py", line 469, in execute
ctx = encode_all_batches(
runner,
...<11 lines>...
color_correction=color_correction
)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\core\generation_phases.py", line 489, in encode_all_batches
cond_latents = runner.vae_encode([transformed_video])
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\utils_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\core\infer.py", line 179, in vae_encode
latent = self.vae.encode(sample, tiled=self.encode_tiled, tile_size=self.encode_tile_size,
~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
tile_overlap=self.encode_tile_overlap).latent
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\models\video_vae_v3\modules\attn_video_vae.py", line 1685, in encode
p = super().encode(x, return_dict=return_dict, tiled=tiled, tile_size=tile_size,
~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
tile_overlap=tile_overlap).latent_dist
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\diffusers\utils\accelerate_utils.py", line 46, in wrapper
return method(self, *args, **kwargs)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\models\video_vae_v3\modules\attn_video_vae.py", line 1186, in encode
h = self.tiled_encode(x, tile_size=tile_size, tile_overlap=tile_overlap)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\models\video_vae_v3\modules\attn_video_vae.py", line 1403, in tiled_encode
encoded_tile = self.slicing_encode(tile_sample)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\models\video_vae_v3\modules\attn_video_vae.py", line 1259, in slicing_encode
self._encode(
~~~~~~~~~~~~^
torch.cat((x[:, :, :1], x_slices[0]), dim=2),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
memory_state=MemoryState.INITIALIZING,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\models\video_vae_v3\modules\attn_video_vae.py", line 1218, in _encode
h = self.encoder(_x, memory_state=memory_state)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\models\video_vae_v3\modules\attn_video_vae.py", line 849, in forward
sample = self.mid_block(sample, memory_state=memory_state)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\models\video_vae_v3\modules\attn_video_vae.py", line 663, in forward
hidden_states = rearrange(
hidden_states, "(b f) c h w -> b c f h w", f=video_length
)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\einops\einops.py", line 616, in rearrange
return reduce(tensor, pattern, reduction="rearrange", **axes_lengths)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\einops\einops.py", line 535, in reduce
backend = get_backend(tensor)
File "D:\ComfyUI_windows_latest\ComfyUI_windows_portable\python_embeded\Lib\site-packages\einops_backends.py", line 62, in get_backend
raise RuntimeError(f"Tensor type unknown to einops {type(tensor)}")
RuntimeError: Tensor type unknown to einops <class 'tuple'>

[naxci1]

I think there are corrupted files in the code. Go into the directory and delete all the seedvr2 files. Download the main repository as a zip file and place it in that directory. Then, restart comfyui.

But first, try closing the VAE tile, and make sure no other applications that strain the GPU are open on your PC while this is running.

[jason-mightynice]

Thanks for reply. I reinstalled seedvr2, and tried setting encode_tiled to false, but got the same error.

I asked Gemini for assistance, and the fix it gave me works. Here is what it told me.

You can likely fix this by forcing the code to use the first element of that tuple.

- Navigate to this file:

D:\ComfyUI_windows_latest\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\src\models\video_vae_v3\modules\attn_video_vae.py

- Go to Line 663 (as indicated in your stack trace).

You will see a block that looks like this:

**Python**
hidden_states = rearrange(
    hidden_states, "(b f) c h w -> b c f h w", f=video_length
)

- Change it to this:

**Python**
# Add this check right before the rearrange call
if isinstance(hidden_states, tuple):
    hidden_states = hidden_states[0]

hidden_states = rearrange(
    hidden_states, "(b f) c h w -> b c f h w", f=video_length
)

- Save the file and restart ComfyUI.

Why is this happening?
In many modern transformer and VAE architectures, "Blocks" return more than just the hidden states—they might return attention maps or "Memory States" (which your logs show are being used: memory_state=MemoryState.INITIALIZING).

The mid_block in this specific version of the SeedVR2 VAE is returning a (tensor, state) tuple, but the line at 663 is trying to "rearrange" the whole package instead of just the tensor.

[naxci1]

excellent. This is a bug, and hopefully it will be fixed in the next version. you keep this issue open.