SeedVR2 (Down)Load VAE Model:
SeedVR2LoadVAEModel configures the variational autoencoder (VAE) used in a SeedVR2 video-upscaling workflow. The VAE converts video frames into a compact representation for processing and converts that representation back into images afterward. This node lets you prepare those encoding and decoding stages for your available hardware. It supports processing large frames in smaller, overlapping tiles to reduce memory use, moving an idle model away from the inference device, retaining a model between workflow runs, and passing through optional compilation settings. The node itself creates a configuration for the Video Upscaler; it does not encode frames, decode frames, or run the upscaler when executed on its own.
SeedVR2 (Down)Load VAE Model Input Parameters:
model
Selects the VAE model filename for encoding and decoding. The default is ema_vae_fp16.safetensors, and that is the only listed option in the supplied registry and schema; no numeric minimum or maximum applies. The model is identified as an FP16 VAE, distinct from the SeedVR2 DiT upscaling models. The interface says registered models can download automatically on first use, so an initial workflow run may take longer. Although its tooltip mentions adding models to a ComfyUI models folder, the supplied VAE-choice function returns only registered VAE names; a newly added file is not shown to appear in this dropdown automatically. This node records your selection in its configuration rather than loading the model during this step.
device
Selects the device that will run VAE encoding and decoding when the configuration is used by the upscaler. There is no numeric minimum or maximum; the options depend on detected hardware. In the supplied one-CUDA-device schema fixture, the only option and default are cuda:0. On another installation, detected CUDA devices are listed as cuda:0, cuda:1, and so on, and an available Apple Metal device is listed as mps. The default is the first detected device. Choose a device with enough available memory for the frames you intend to process; if your system detects no supported inference device, this node may be unable to offer a valid choice.
encode_tiled
Controls whether the VAE encodes frames in smaller image sections instead of processing each frame as a whole. Its options are false and true, with false as the default; no numeric minimum or maximum applies. Enable it when encoding high-resolution frames exceeds available device memory. Tiling can reduce peak memory use, while processing more sections may take longer. The encoding tile size and overlap settings matter only when this is enabled.
encode_tile_size
Sets the width and height, in pixels, of each section used for tiled encoding. The default is 1024, the minimum is 64, and the widget changes in steps of 32; the supplied schema specifies no maximum. Smaller tiles generally use less memory but can increase processing time. This setting is used only when encode_tiled is true, so changing it while encoding tiling is off will not change how frames are encoded.
encode_tile_overlap
Sets how many pixels neighboring encoding tiles overlap. The default is 128, the minimum is 0, and the widget changes in steps of 32; the supplied schema specifies no maximum. The overlap gives adjacent tiles shared image area for blending, helping reduce visible boundaries between them. A larger overlap means more repeated processing and can slow the workflow. This setting is used only when encode_tiled is true.
decode_tiled
Controls whether the VAE converts its processed representation back into images in smaller sections. Its options are false and true, with false as the default; no numeric minimum or maximum applies. Enable it if decoding full-resolution output frames uses more device memory than you have available. Encoding and decoding tiling are independent, so you can enable this setting without enabling encode_tiled. The decoding tile size and overlap settings matter only when this is enabled.
decode_tile_size
Sets the width and height, in pixels, of each section used for tiled decoding. The default is 1024, the minimum is 64, and the widget changes in steps of 32; the supplied schema specifies no maximum. Lower values can reduce memory use at the cost of additional processing time. This setting is used only when decode_tiled is true.
decode_tile_overlap
Sets how many pixels neighboring decoding tiles overlap. The default is 128, the minimum is 0, and the widget changes in steps of 32; the supplied schema specifies no maximum. Overlap supports blending where decoded sections meet, reducing visible tile boundaries. Increasing it can improve those transitions but requires more processing. This setting is used only when decode_tiled is true.
tile_debug
Selects a tile-boundary visualization mode. The options are false (the default, with no overlay), encode (show encoding tile boundaries), and decode (show decoding tile boundaries); no numeric minimum or maximum applies. Use it to inspect how a tiled stage divides frames when diagnosing boundary artifacts. The selected visualization works only when the corresponding encoding or decoding tiling setting is enabled.
offload_device
Selects where the VAE should reside when it is not actively processing. The default is none, which keeps it on its inference device and avoids movement between devices; no numeric minimum or maximum applies. The available options depend on hardware: the supplied fixture lists none, cpu, and cuda:0. Choosing cpu can free GPU memory by moving the idle model to system memory, while another listed GPU can be an offload destination on a multi-GPU system. On an MPS-only system, the supplied device-list logic excludes cpu from these choices. This setting describes idle placement, not the device selected for active encoding and decoding.
cache_model
Controls whether the VAE remains loaded on its selected offload device between workflow runs. Its options are false and true, with false as the default; no numeric minimum or maximum applies. Caching can avoid repeated loading during batch work, but it requires offload_device to be set to something other than none. Enabling caching while leaving offload_device at none causes this node to raise a validation error instead of returning a configuration.
torch_compile_args
Accepts an optional TORCH_COMPILE_ARGS connection from the SeedVR2 Torch Compile Settings node; it has no widget default, numeric minimum, numeric maximum, or enumerated options in the supplied schema. Leave it unconnected to use no compilation settings, or connect compatible settings to request the available torch.compile optimization when the workflow uses this configuration. The tooltip cites a possible 15–25% speedup with compatible PyTorch 2.0+ and Triton, but that figure is not a guarantee for your hardware or workload. This node passes the settings through; the supplied source does not show it compiling or benchmarking the VAE itself.
SeedVR2 (Down)Load VAE Model Output Parameters:
SEEDVR2_VAE
Returns one SEEDVR2_VAE configuration value for connection to the SeedVR2 Video Upscaler node. It is a collection of the selected model name, inference device, idle offload device, caching choice, encoding and decoding tiling settings, tile visualization mode, optional compilation settings, and an internal node identifier. The value tells the downstream workflow how to handle the VAE; it is not an image, an encoded frame, or a loaded model returned by this node. Connect it to the Video Upscaler’s corresponding VAE input so that the upscaler can use these choices.
SeedVR2 (Down)Load VAE Model Usage Tips:
- Start with the default VAE, the detected inference device, and both tiling switches off. Connect
SEEDVR2_VAEto the Video Upscaler, then enable tiling only for the stage that runs out of memory. - For a memory error during encoding, enable
encode_tiledand start with its1024-pixel tile size and128-pixel overlap. If memory remains insufficient, reduceencode_tile_sizein32-pixel steps. - For a memory error during decoding, adjust
decode_tiledanddecode_tile_sizeinstead. The encoding settings do not control decoding memory use. - If tile boundaries are visible, retain or increase the relevant overlap and use
tile_debugset toencodeordecodeto inspect the affected stage. Turn the debug overlay back tofalsewhen finished inspecting. - Leave
offload_deviceatnonewhen the VAE fits comfortably on its inference device. Choose an available offload destination when you need to free device memory between processing stages, and enablecache_modelonly after selecting a destination other thannone. - Connect
torch_compile_argsonly when you have a compatible settings node and environment. Treat the stated speedup as a possibility to assess in your own workflow, not an expected result.
SeedVR2 (Down)Load VAE Model Common Errors and Solutions:
Model caching (cache_model=True) requires offload_device to be set.
- Explanation: This validation message comes directly from the node. Caching is enabled, but
offload_deviceis stillnone, so the node has no configured destination for retaining the idle VAE. - Solution: Set
offload_deviceto an available option such ascpuor another appropriate device, or setcache_modeltofalse.
No inference device is available in device
- Explanation: This is a possible schema-creation issue, not a confirmed message from the supplied source. The device list is built from detected CUDA and Apple Metal devices, and the schema takes its default from the first item. If neither is detected, there may be no first item to select.
- Solution: Check that your GPU or Apple Metal device is available to PyTorch in this installation. Restore the relevant driver or PyTorch device support, then reopen or refresh ComfyUI so the available-device list can be built again.
A VAE model cannot be found or downloaded
- Explanation: This is a possible downstream model-access failure; the supplied node does not provide a specific error string for it. The interface says the registered VAE downloads on first use, so access to the model or its download location may matter when the upscaler first needs it.
- Solution: Confirm that
ema_vae_fp16.safetensorsis the selected registered model, then check the workflow’s model-download error for the specific cause, such as connectivity, permissions, or insufficient storage. Do not assume that placing an unregistered VAE file in a models folder will make it appear in this dropdown.
Out-of-memory error during encoding or decoding
- Explanation: This is a possible error when the downstream upscaler uses the VAE, not an error message produced by this configuration node. Processing a whole high-resolution frame may exceed the selected device’s available memory.
- Solution: Enable
encode_tiledfor an encoding failure ordecode_tiledfor a decoding failure. Start with the displayed tile defaults and lower the affected tile size if necessary; you can also select an appropriate idle offload destination to free device memory between stages.
Compilation or unsupported-device error with torch_compile_args
- Explanation: This is a possible downstream compatibility problem, not a verified error string from this node. Compilation depends on the connected settings and a compatible environment; the node itself only passes those settings along.
- Solution: Disconnect
torch_compile_argsand try the workflow again to isolate the issue. If the workflow succeeds without it, check the Torch Compile Settings node and the installed PyTorch, Triton, and device support before re-enabling compilation.
