The native 3B INT8 graph
Read the SeedVR2 ComfyUI workflow before you queue it
Most SeedVR2 ComfyUI failures are workflow errors, not model errors. This page maps the maintained 3B INT8 template node by node so you can see which loader owns the model, which owns the VAE, where temporal chunking lives, and what the main upscaler actually controls. Bring your own source only after the bundled sample runs.
This is a SeedVR2.net explainer, not a Comfy-Org or ByteDance publication. The node map and folder names come from the public ComfyUI SeedVR2 template and Comfy-Org repack. The hosted tool is a separate credit product; use this page to understand the local graph, then choose your path.
Skip the graph when you only need one file
If your goal is a finished image rather than an owned graph, upload a short lawful source, choose a resolution, and read the credit preflight. You can still build the local workflow afterward from the map below.
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 nodes own the whole workflow
The native template splits SeedVR2 into a diffusion model loader, a VAE loader, an optional temporal chunk, and the main upscaler. Each node owns different settings; fixing the wrong one is the most common mistake.
- DiT loader
- Model + precision
- VAE loader
- Encode + decode
- Temporal chunk
- Video only
- Main node
- Resolution + batch
Chooses model family, FP16/FP8, offload, attention backend, and BlockSwap.
Owns VAE tiling and tile overlap for large frames.
Processes long video latents in pieces when VRAM is tight.
Controls target resolution, batch size, temporal overlap, and color settings.
The two model folders the template expects
The maintained template points to seedvr2_3b_int8_convrot.safetensors in ComfyUI/models/diffusion_models and seedvr2_ema_vae_fp16.safetensors in ComfyUI/models/vae. Keep the filenames unchanged. A renamed file is the fastest way to get a model-not-found error that has nothing to do with the graph.
The 3B INT8 diffusion model and the shared EMA VAE are the Comfy-Org repack of ByteDance's SeedVR2 checkpoints, published under Apache-2.0. They are not the same as the older v2.5 community custom-node layout, which adds its own DiT, VAE, torch compile, and upscaler nodes.
- Diffusion model goes in models/diffusion_models.
- VAE goes in models/vae.
- Keep filenames unchanged, then reopen the template.
Trace the graph in execution order
Read the graph from input to output, not from the prettiest node. The DiT loader hands a diffusion model and precision policy to the main upscaler. The VAE loader hands encode and decode settings. The upscaler receives your image or video frames and applies resolution, batch size, temporal overlap, and color correction before returning the result.
When something breaks, match the error to the owning node. A CUDA out-of-memory during model inference is the DiT side. A crash while encoding or decoding a large frame is the VAE side. Flicker across frames is the batch and temporal-overlap side. This one-to-one map turns an hour of guessing into a five-minute fix.
Run the bundled sample before your own source
The template ships with a sample queue. Run it untouched first. If the sample passes, your installation and model folders are correct and any later failure belongs to your source. If the sample fails, you have a setup problem and adding your own file will only add variables.
After the sample passes, switch to one of your own sources at a conservative resolution. Do not jump to 4K or a long clip. A 2K still or a five-second 1080p excerpt is enough to prove the graph works on your media before you commit a batch.
- Run the bundled sample untouched.
- Then test one of your own sources at a low resolution.
- Record every setting that produced a working run.
Native template versus the v2.5 community node
The native template is the shortest maintained path on ComfyUI 0.28.0 or newer. The separate NumZ v2.5 community custom node exists for workflows that specifically need GGUF, BlockSwap, attention, or batch controls beyond the core nodes. It is useful, and it is also extra software with its own update and dependency risk.
Use the native template unless your workflow names a feature that only the community node provides. Do not install the community node just because an old tutorial screenshot shows it. Every extra package is another thing that can break after a ComfyUI update.
Match every error to its node
Missing node means ComfyUI cannot load a node type the workflow references: update ComfyUI or install the matching community node, then restart. Model not found means the filename or directory is wrong. OOM means the graph, precision, batch, or resolution exceeds VRAM. Seams point to VAE tile overlap.
Solve the first terminal error before changing anything else. The hosted workspace is the right fallback when the asset must ship tonight and the graph is still broken; it does not remove your ability to keep debugging locally later.
What this workflow map cannot promise
This page describes a public template as of the review date, not a benchmark of your specific card. Templates, node versions, and model files change; validate on your own hardware with one small source before trusting a larger batch.
- ComfyUI, node, and model versions move; a screenshot from an old tutorial is not current documentation.
- Nothing here promises 8GB support; upstream documents H100 usage and the integration maintainer reports roughly 18GB in testing.
- A successful launch is not a quality guarantee; the source still has to contain recoverable detail.
SeedVR2 ComfyUI workflow questions
How many nodes does the native SeedVR2 workflow use?
The maintained template centers on four: a DiT loader, a VAE loader, an optional temporal chunk for video, and the main upscaler. The community v2.5 node splits the same work across its own DiT, VAE, torch compile, and upscaler nodes.
Which folders do the model files go into?
The diffusion model goes in ComfyUI/models/diffusion_models and the VAE in ComfyUI/models/vae. The template names the exact files; keep the filenames unchanged and reopen the template so the loaders find them.
Why should I run the bundled sample first?
Because it separates installation problems from source-file problems. If the untouched sample passes, your folders and install are correct. If it fails, you have a setup issue and adding your own media only adds variables.
How do I know which node is failing?
Match the error to the owning node: OOM during inference is the DiT side, encode/decode crashes are the VAE side, and flicker across frames is the batch and temporal-overlap side of the main upscaler.