SeedVR2 (Down)Load DiT Model:
SeedVR2LoadDiTModel, displayed in ComfyUI as “SeedVR2 (Down)Load DiT Model,” prepares the Diffusion Transformer (DiT) used during SeedVR2 video upscaling. It lets you choose a model and configure how it uses your available hardware, including ways to reduce GPU memory use, retain a model between workflow runs, and select an attention or compilation optimization. Think of it as the upscaler’s model setup step: it passes your choices to the Video Upscaler rather than producing an upscaled video by itself. (github.com)
SeedVR2 (Down)Load DiT Model Input Parameters:
model
Required combo selection. Choose the DiT model used for upscaling. The default is seedvr2_ema_3b_fp8_e4m3fn.safetensors. The registry offers 3B models in Q4_K_M, Q8_0, FP8, and FP16 formats, plus 7B models in Q4_K_M, mixed FP8/FP16, and FP16 formats; the 7B choices also include sharp variants. There is no numeric minimum or maximum. Model size, precision, and variant determine which model the upscaler will use, while the file format determines which supplied model file must be available. Registry models can download on first use, and additional DiT files discovered on your installation may also appear in the menu. The listed options can therefore differ from the supplied schema fixture. (github.com)
device
Required combo selection. Choose the GPU that performs DiT inference, meaning the upscaling computation. The supplied one-CUDA-device schema shows only cuda:0 and uses it as the default; this is an example of a detected machine, not a fixed list for everyone. The source builds choices from available CUDA devices and, where available, Apple Metal’s mps device, with the first detected device as the default. There is no numeric minimum or maximum. Selecting a different available GPU changes where this part of the workflow runs; the device must exist on your machine.
blocks_to_swap
Optional integer controlling how many transformer blocks are swapped between the inference device and an offload device to reduce GPU memory pressure. Its widget default is 0 (disabled), with a minimum of 0, maximum of 36, and step of 1. Use no more than 32 for a 3B model or 36 for a 7B model, as specified by the node tooltip. A value above zero requires an offload_device that is neither none nor the same device used for inference. Moving more blocks can make a memory-constrained workflow possible, but transfers can slow processing. The tooltip says this feature is unavailable on macOS because of its unified memory architecture.
swap_io_components
Optional Boolean that also offloads the DiT model’s input/output embeddings and normalization layers to reduce GPU memory use. The default is false; the available values are false and true, with no numeric range. Set it to true only when offload_device is set to a device different from device. It can help when GPU memory is tight, potentially at the cost of additional transfers. The tooltip says it is unavailable on macOS.
offload_device
Optional combo selection specifying where the DiT model is kept when it is not actively processing and which other device BlockSwap uses. The default is none, which keeps the model on the inference device. On the supplied one-CUDA-device fixture, the options are none, cpu, and cuda:0; actual choices depend on detected hardware. cpu uses system RAM, while another available cuda:X GPU can be used for offloading. There is no numeric range. For blocks_to_swap or swap_io_components, choose an available destination different from device; on an MPS-only system, the source excludes cpu from this menu.
cache_model
Optional Boolean controlling whether the DiT model remains loaded on the offload device between workflow runs, which can avoid loading it again during repeated or batch work. The default is false; the available values are false and true, with no numeric range. If you set it to true, you must set offload_device to something other than none: the node explicitly rejects that combination. Keeping a model loaded also means its destination must have room for it.
attention_mode
Optional combo selection for the method used to calculate the model’s attention operations. The default is sdpa, the PyTorch option recommended by the node tooltip for its broad availability. The other options are flash_attn_2 (Ampere or newer, with flash-attn installed), flash_attn_3 (Hopper or newer, with Flash Attention 3 support), sageattn_2 (with sageattention installed), and sageattn_3 (Blackwell/RTX 50xx, with sageattn3 installed). There is no numeric range. A compatible optimized backend may improve speed, but choosing one without its required package and GPU support can prevent the workflow from running.
torch_compile_args
Optional TORCH_COMPILE_ARGS connection for settings supplied by the SeedVR2 Torch Compile Settings node. It has no widget default, numeric range, or menu of choices; when unconnected, its value is None. The setting requests optional PyTorch compilation optimization, for which the tooltip cites a potential 20–40% speedup with compatible PyTorch 2.0+ and Triton. That figure is a tooltip claim, not a measured result for your setup. Leave this input unconnected for a simpler initial workflow, then assess it on your hardware if you want to optimize repeated runs.
SeedVR2 (Down)Load DiT Model Output Parameters:
SEEDVR2_DIT
The sole output is a DiT configuration for the SeedVR2 Video Upscaler. Connect it to the upscaler’s corresponding DiT input; it is not an image, video, or completed upscale. Its value carries the selected model filename, inference and offload devices, caching and BlockSwap choices, attention mode, optional compile settings, and an internal node identifier. The upscaler uses these settings to determine how to obtain and run the model. In the shown execute method, this node constructs and returns that configuration; the provided verification did not run model loading or GPU inference.
SeedVR2 (Down)Load DiT Model Usage Tips:
- Start with the default 3B FP8 model, your detected GPU,
offload_deviceset tonone,blocks_to_swapset to0, andattention_modeset tosdpa. ConnectSEEDVR2_DITto the Video Upscaler, then change one setting at a time so you can tell what helps. - If upscaling runs out of GPU memory on a CUDA system, try a model format or size that uses less memory, or set
offload_devicetocpuor another GPU and increaseblocks_to_swapgradually. Keep within the model-specific block limit. BlockSwap reduces GPU memory pressure but may add transfer time. (huggingface.co) - For repeated workflow runs, set an appropriate
offload_devicebefore enablingcache_model. Caching without an offload destination produces an explicit error. - Use
swap_io_componentsalongside BlockSwap only when you need further GPU memory relief and have a different offload destination. Leave both swapping controls disabled on macOS, as directed by their tooltips. - Keep
sdpaunless you know your GPU and installation support another attention backend. Test optional compilation after the basic workflow works; the stated speed benefit is not guaranteed for every setup.
SeedVR2 (Down)Load DiT Model Common Errors and Solutions:
Model caching (cache_model=True) requires offload_device to be set.
- Explanation: This is an explicit
ValueErrorfrom this node whencache_modelistruebutoffload_deviceisnone. The full message also reports the current offload value and asks you to choose a storage device or disable caching. - Solution: Select an available
offload_device, such ascpuon a supported CUDA setup or another GPU, or setcache_modelback tofalse.
CUDA out of memory
- Explanation: This is a possible downstream GPU-memory error during model loading or upscaling, not an error string raised by the shown loader method. The selected model and workload may exceed available GPU memory.
- Solution: Try a smaller or lower-memory model choice. On a supported CUDA setup, configure a different
offload_deviceand gradually increaseblocks_to_swapwithin the selected model’s limit; reduce the upscaling workload if needed. (huggingface.co)
Model file not found or model download failed
- Explanation: These are possible downstream file or download failures, not exact messages established by the shown node method. Registered models are described as downloading on first use, while extra models depend on files present in your installation.
- Solution: Confirm the selected filename is available or that the first-use download can complete. For an added model, check that its file is in the ComfyUI model location recognized by this installation, then refresh the model list.
Attention backend package missing or unsupported GPU
- Explanation: An optimized
attention_modemay require a separately installed package and a compatible GPU. The provided source identifies those requirements but does not establish one exact failure message. - Solution: Switch to
sdpato check the basic workflow. Useflash_attn_2,flash_attn_3,sageattn_2, orsageattn_3only after confirming the stated package and hardware requirements for your choice.
BlockSwap destination is missing or matches the inference device
- Explanation: The tooltips require a different, configured
offload_devicewhenblocks_to_swapis above0orswap_io_componentsistrue. The shownexecutemethod does not itself raise a specific error for this combination, so any later message depends on the rest of the workflow. - Solution: Set
offload_deviceto an available device other thandevice, or disable BlockSwap by settingblocks_to_swapto0andswap_io_componentstofalse.
torch.compile or Triton is unavailable
- Explanation: Optional compile settings may fail later if the PyTorch or Triton installation is incompatible. The provided loader does not show an exact error message for this case.
- Solution: Disconnect
torch_compile_argsand confirm the ordinary workflow runs. Then check the compilation settings and the tooltip’s PyTorch 2.0+ and Triton prerequisites before enabling it again.
No device choices appear
- Explanation: The source builds inference-device choices from detected CUDA or MPS devices. If neither is detected, the list is empty, and the source’s
devices[0]default cannot be selected. The exact UI or startup error is not established by the supplied fixture. - Solution: Check that ComfyUI’s Python environment detects a supported GPU backend and that the relevant drivers and PyTorch installation work. Restart ComfyUI after correcting the environment so the device choices can be rebuilt.
