Settings, not superstition
Which SeedVR2 ComfyUI settings change your result
SeedVR2 has a handful of settings that actually change memory use, speed, and output: target resolution, batch size, VAE tile size and overlap, offload device, and BlockSwap. Everything else is mostly noise until one of those is wrong. This page orders them by how often they cause a failed run, so you tune the right knob first.
This is a SeedVR2.net settings explainer built from the public ComfyUI SeedVR2 template and the community parameter guidance already published on this site. It is not a benchmark of a specific GPU. Validate each setting on your own card with one small source.
A hosted test removes the tuning loop
When settings tuning is blocking a deadline, run one short clip or still in the browser. The hosted job handles model, memory, and queue; your job is to judge whether the output is worth further local work.
First eligible 5-second 720p Standard run: free
Verify your email to use your first video run without credits or a card. A failed run does not use the trial; longer clips and other settings show their exact credit cost before submission.
Browser upload limit: 500 MB. For larger files, paste a publicly accessible HTTPS URL. Each job supports videos up to 10 minutes.
Quick scene presets
Presets pick a model for a common scenario. You can still switch models manually below.
Upscaled preview
Compare the enhanced output before exporting the final file.
Upload a clip, paste a direct video URL, or load the sample. No example video is downloaded until you ask for it.
Tip: keep your original FPS and avoid pre-scaling. The upscaler handles cadence better than most NLE plugins.
Test before processing the full video
Upscaling cannot recreate detail removed by heavy messaging compression. Test the hardest 5–10-second segment first; a higher-priced model is not automatically a better fit.
Four settings do most of the work
These four knobs cover the majority of failed and slow SeedVR2 runs. Change them in this order and record each value so a working run is reproducible.
- Resolution
- Start at 720p–1080p
- Video batch
- 4n+1 (5, 9, 13, 17)
- VAE tiling
- 1024 tile / 128 overlap
- BlockSwap
- Last resort
Lower output first, then step up to 2K or 4K only after a clean pass.
Follow the 4n+1 pattern; tiny batches can reduce motion stability.
Enable only when encode or decode hits OOM; raise overlap before accepting seams.
Use only after a repeatable CUDA OOM proves the model cannot fit.
Lower the resolution before anything else
Most SeedVR2 OOM errors are resolution errors. A 4K frame needs far more memory than a 720p frame, and the difference is not linear. Start a new source at 720p or 1080p, confirm it passes, then step up. If 720p fails while 1080p was your only test, you have been guessing.
Resolution is also a quality lever. Pushing a weak source to 4K makes ringing, plastic skin, and broken text more visible, not more true. A clean 2K crop is often a better deliverable than a noisy 4K canvas.
- Confirm 720p or 1080p first.
- Step up one resolution at a time.
- Reject 4K when the 2K crop already fails.
Batch size is a temporal setting, not just memory
For still images, batch size can stay at one. For video, batch size controls how many frames are processed together, which drives temporal consistency. The SeedVR2 workflow expects normal video batches to follow the 4n+1 pattern such as 5, 9, 13, or 17 frames.
On a tight GPU you will be tempted to drop the batch to the smallest value that runs. Do that only after memory optimizations, because a tiny batch can reduce flicker protection and leave visible frame-to-frame drift.
VAE tiling: start near 1024 tile and 128 overlap
VAE tiling solves encode and decode memory pressure, which is separate from the diffusion model. Start near a 1024 tile size with 128 overlap, then reduce the tile only if the log still shows a VAE-side OOM.
Smaller tiles are not automatically better. Each tile boundary is a place the model must reconcile, so aggressive tiling with too little overlap creates seams. If seams appear, raise the overlap before shrinking the tile further.
Offload and BlockSwap come last
When the model itself will not fit, move the DiT offload device to CPU. When a repeatable CUDA OOM remains, enable BlockSwap, which keeps fewer transformer blocks on the GPU and moves the rest through system memory.
BlockSwap is a memory workaround, not a speed or quality upgrade. Every swapped block slows the run. If the asset has a deadline, the hosted workspace is usually cheaper than another evening of swap tuning.
- Offload the model when it will not fit.
- Reserve BlockSwap for repeatable OOM.
- Stop tuning when local time exceeds the value of the output.
Record every working setting
A working run is only useful if you can reproduce it. Record the ComfyUI revision, node revision, model filename, input dimensions, target resolution, batch size, tile size, overlap, offload device, and whether BlockSwap was on.
That one line is the difference between a repeatable workflow and an unexplained break after an update. Save it with the output file, not in your head.
These settings are guidance, not a guarantee
Memory behavior depends on your card, clip length, and graph. The values here come from published SeedVR2 guidance and are starting points to test, not promises that any specific 8GB, 12GB, or 16GB card will pass.
- Different GPUs fail at different resolutions even with identical settings.
- ComfyUI and node updates can change the meaning of a setting.
- Good settings cannot rescue a source with no recoverable detail.
SeedVR2 ComfyUI settings questions
Which setting should I change first when SeedVR2 runs out of memory?
Lower the target resolution first, then the batch size. Only then enable VAE tiling, move the model offload to CPU, or try BlockSwap. Changing memory workarounds before resolution usually just hides the real problem.
What video batch size should I use?
Follow the 4n+1 pattern such as 5, 9, 13, or 17 frames. For still images, batch size can be one. Do not drop video batches to a tiny value before trying memory optimizations, because that can reduce motion stability.
When should I enable VAE tiling?
Enable it when the log shows a memory failure during encode or decode, which is separate from diffusion inference. Start near a 1024 tile with 128 overlap, and raise the overlap if seams appear.
Is BlockSwap always worth enabling?
No. BlockSwap is a memory workaround for a repeatable CUDA out-of-memory during inference. It slows runs, so use it only after resolution reduction, batch reduction, and offload have not resolved the failure.