01
What the ComfyUI port is
Saganaki22's ComfyUI-VDN-H3 is a community port, not an official OpenVDN or MiniMax product. It adapts the released Video DeltaNet checkpoints to ComfyUI and substitutes portable PyTorch implementations for parts of the datacenter-oriented reference stack.
02
1. Install the custom node
From the ComfyUI `custom_nodes` directory, clone `https://github.com/Saganaki22/ComfyUI-VDN-H3`, then restart ComfyUI. The current repository documents Windows with NVIDIA as its primary tested target, says Linux NVIDIA should work, and describes the port as single-GPU.
The maintainer currently says this path adds no new Python dependencies beyond the ComfyUI environment it expects. AMD, Intel and CPU-only execution are not documented as tested targets.
03
2. Put the VDN checkpoint in the expected folder
For the released distilled path, download `stage-dmd-step-250/*` from `OpenVDN/vdn-minimax-h3` and place the stage under `ComfyUI/models/vdn/`. Keep the stage metadata, linear branch and adapters together rather than converting the files on disk.
The released 8-step stage is about 5.1 GB and the 50-step stage about 4.3 GB, in addition to the MiniMax H3 base components. The branch itself is therefore not a complete-system VRAM estimate.
04
3. Apply VDN-H3 and choose the memory mode
Add the port's VDN-H3 apply node to a MiniMax H3 workflow. For the DMD stage, use its matching turbo adapter and eight sampling steps as documented by the port; stacking unrelated attention replacements or turbo adapters can create an invalid combination.
The port can keep branch weights resident on the GPU when memory permits or stream branch blocks when memory is tighter. The maintainer describes the full branch at roughly 4.3 GB and a much smaller block-at-a-time streaming working set, trading speed for lower persistent memory.
05
What performance and VRAM claims mean
In one documented RTX 5090 workload at 1280×736 and 145 frames, the maintainer reports about 17 seconds per iteration and roughly 2 minutes 15 seconds of sampling across eight evaluations. That is a community measurement tied to its workflow, quantization and backend choices, not a reproduction of OpenVDN's eight-B200 result.
The repository also describes specific INT8-oriented configurations that can fit 736p work on some 12–16 GB cards. Treat that as configuration-specific community guidance rather than an official VDN-H3 minimum specification.
06
VAE decode can still be the memory spike
Sampling can fit while untiled VAE decoding still runs out of memory at higher resolutions or longer clips. The community documentation recommends tiled VAE decoding when decode memory becomes the bottleneck.
Conditioning variants exposed by the ComfyUI integration should also be described as integration behavior. OpenVDN's documented training and benchmark evidence does not establish that every community conditioning layout was independently trained or benchmarked.
07
Official OpenVDN vs the ComfyUI port
Use the official OpenVDN stack when you need reference-code parity, training code or its H200/B200 distributed paths. Use the ComfyUI port when a visual single-GPU workflow is more useful and you can accept community-maintained kernels, platform boundaries and benchmark evidence.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.