SeedVR2 checkpoints, precisions and the VAE
SeedVR2 model files: 3B or 7B, FP16 or FP8 or INT8
SeedVR2 ships as a diffusion checkpoint plus a separate VAE, and every download is one of those two files at one precision. This page explains what each token in a SeedVR2 file name means, which file the maintained ComfyUI template actually names, and which folder each one belongs in.
This is a SeedVR2.net explainer, not a Comfy-Org or ByteDance publication. The file names, folder names and precision families below come from the maintained ComfyUI template and the Comfy-Org repack as published at the review date. Re-read the template you actually installed before downloading anything to a production machine.
Skip the download when you only need one finished file
Choosing and placing these files is only worth the effort when you want an owned, repeatable graph that runs on your own hardware. If the goal is one upscaled image or clip this afternoon, upload the source, pick a target resolution, and the hosted workspace runs the same SeedVR2 family without a checkpoint download or a folder layout to verify.
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 tokens decide which file you need
A SeedVR2 file name is a specification, not decoration. Read the family, the precision, the role and the format before you download anything, because the wrong combination is the most common reason a workflow fails to load a model.
- Family
- 3B or 7B
- Precision
- FP16, FP8 or INT8
- Role
- Diffusion model or VAE
- Container
- .safetensors, sometimes .gguf
ByteDance publishes both a 3B and a 7B SeedVR2 checkpoint under Apache-2.0. Bigger is not automatically better for a given card; the 3B is the one the maintained ComfyUI template names.
Precision decides how much VRAM the same family needs. Lower precision fits smaller cards, and the cheapest precision the template names is the INT8 build.
One file is the diffusion checkpoint and one is the VAE — in the maintained template that is seedvr2_3b_int8_convrot.safetensors plus the shared EMA VAE. They load in different nodes from different folders, so a VAE dropped into the diffusion folder reads as a missing model.
The template uses .safetensors. GGUF re-exports of the same family also circulate for low-VRAM machines, but they are community repacks, so check the publisher before you trust one.
3B or 7B: the same task at two sizes
The 3B and the 7B are two published sizes of the same SeedVR2 super-resolution model, both released under Apache-2.0. The difference you feel first is memory, not file size on disk: a larger checkpoint needs more VRAM for the same frame and batch settings.
Treat the 7B as a second pass rather than a default. A conservative run on a 12GB card can pass once and fail on the next clip, which is a worse experience than starting from the smaller file and only stepping up after your own footage passes.
That ordering is also what the maintained template encodes: it names a 3B file, so a working install of the template is already a validated 3B baseline you can compare a 7B run against.
- Start with the 3B file the template names; it is the cheaper one to validate on your card.
- Keep one short clip as a fixture and run both files against it before committing a whole batch.
- Compare one crop from each family instead of two different sources, or you are measuring the source.
FP16, FP8 and INT8: what the precision token costs you
Precision is how many bits each weight is stored in. FP16 keeps the most detail per weight and uses the most memory; FP8 halves that footprint again; INT8 is an integer build aimed at the tightest cards. They are the same model, quantised differently.
The practical rule is to go down in precision until the run fits, not up until it looks impressive. A 3B at a lower precision that finishes your clip beats a 7B at FP16 that runs out of memory halfway through.
Repacked file names tend to keep the family and the precision in the name, which is why one name can be long: the family says what the model is, the precision says what it costs, and the tail is the distribution's own label.
- FP16 is the reference precision and the largest of the three.
- FP8 and INT8 exist so the same family fits smaller cards; expect to trade a little detail for the fit.
- Change one token at a time, then rerun the same fixture, or you cannot tell which change mattered.
The VAE is a second file, not a second model
SeedVR2 needs a VAE alongside the diffusion checkpoint. In the ComfyUI graph these are separate loader nodes: one loads the diffusion model, another loads the VAE, and the main node stitches their output into the finished frame.
That split is why a folder mistake looks like a missing model. If the VAE lands in the diffusion folder, the VAE loader finds nothing, and the failure points at the VAE node rather than at the file you misplaced.
Precision applies here too. The VAE is small compared with the checkpoint, but an FP16 VAE paired with a quantised checkpoint is the combination the maintained template expects, so matching the template is usually the safer choice than mixing builds.
- Two files, two folders: the checkpoint is a model, the VAE is a VAE.
- Keep the file names exactly as downloaded; templates reference names, not contents.
- Reopen the template after moving files so the node list re-resolves.
Where each file belongs in a ComfyUI install
The template names two folders, and both are relative to the ComfyUI root rather than to the workflow file: the diffusion model belongs in models/diffusion_models and the VAE belongs in models/vae. Extra files in the wrong folder are ignored rather than reported.
If a model does not appear in the loader's dropdown after you move it, reopen the template before you debug anything else, because the node caches its list when the graph loads.
Keep one working folder layout and document it. Most 'the model is broken' reports on a second machine are actually a file that was never moved to the path the template references.
Choosing between the files for your own machine
Decide with one number: the VRAM your card actually gives a single process, not the figure on the box. Then walk down the family and precision list until a short clip completes twice in a row without an out-of-memory error.
Only after that baseline passes should you raise a target resolution or move to a larger family. Raising both at once is how a working setup becomes an unexplained failure.
Write down the combination that worked, because the three tokens are the whole answer: family, precision, and which of the two roles the file plays.
- Smaller family and lower precision first; raise them only after a repeated pass.
- Test at a low target resolution, then raise resolution as a separate change.
- Record the exact file names that produced the pass, next to the settings.
When running the same family online is the shorter path
Downloading and placing files pays off when you want an owned graph you can rerun offline on your own terms. It costs time on every machine you install it on, and it puts the memory ceiling on your card rather than on a service.
The hosted workspace on this site runs the SeedVR2 family for you: upload a supported source, choose a target resolution, and you get a result without choosing a precision or a folder. Local install stays the better answer when the source must not leave your machine or when you need a batch pipeline you fully control.
A reasonable split is to validate the look on a hosted run first, then invest the download time only for the sources you will rerun often.
What this file map cannot promise
File names, folders and repacks move faster than explainers. This page describes the maintained template and the Comfy-Org repack as published at the review date, and says nothing about the card you actually own.
- Repackers rename files; the maintained template references its own names, so always confirm against the template you installed.
- Nothing here promises a specific video resolution on a specific card; run your own fixture before committing a batch.
- Community GGUF re-exports are not published by the model authors, so verify the publisher before trusting one.
- A checkpoint that loads is not a quality guarantee; the source still has to contain recoverable detail.
SeedVR2 model file questions
Should I use the 3B or the 7B SeedVR2 checkpoint?
Start with the 3B file the maintained ComfyUI template names. It needs less memory for the same settings, so it is the easier one to prove on your own card. Move to the 7B only after a short clip passes twice in a row on the smaller file.
What is a GGUF SeedVR2 file, and is it safe to use?
GGUF is a container format used for quantised models. Community re-exports of quantised checkpoints circulate for low-VRAM machines, but they are not published by the model authors, so check who repacked the file before you load it. The maintained ComfyUI template uses .safetensors.
Why does the workflow need a separate VAE file?
SeedVR2 keeps the diffusion checkpoint and the VAE in separate files, and the ComfyUI graph loads them in separate nodes. If the VAE is missing or in the wrong folder, the failure appears at the VAE node even though the checkpoint is fine.
Which folder does each SeedVR2 file go in?
The diffusion checkpoint goes in models/diffusion_models and the VAE goes in models/vae, both relative to the ComfyUI root rather than to the workflow file. Reopen the template afterwards, because the loader caches its file list when the graph loads.
Is FP8 or INT8 worse than FP16?
Lower precision fits the same family into less memory, and the trade is detail per weight rather than a different model. Choose the highest precision that finishes your own clip reliably, and change one token at a time so you can tell which change mattered.