The exact files and folders
Download the SeedVR2 model without breaking the graph
Download the two SeedVR2 files from Comfy-Org/SeedVR2 on Hugging Face: the 501 MB VAE seedvr2_ema_vae_fp16.safetensors into models/vae, and the 3B diffusion model seedvr2_3b_int8_convrot.safetensors into models/diffusion_models. A SeedVR2 model download fails almost never because the file is corrupt and almost always because it landed in the wrong folder or got renamed. Match the filenames exactly, then reopen the template.
This is a SeedVR2.net download explainer based on the public Comfy-Org SeedVR2 repack and the maintained ComfyUI template. Filenames and folders are as published at review time; confirm against the live template before downloading. The hosted tool needs no local download.
No download needed for a hosted test
The browser path avoids multi-gigabyte downloads, folder setup, and checksum verification. Upload a short lawful source, read the credit preflight, and decide later whether the local model is worth the disk space.
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.
The two files to download
You need two safetensors files in two specific folders. Everything else in the template is configuration, not a download.
- Diffusion model
- seedvr2_3b_int8_convrot.safetensors
- Shared VAE
- seedvr2_ema_vae_fp16.safetensors
- Disk space
- 2x model size free
- Checksum
- Verify when provided
Goes in ComfyUI/models/diffusion_models; this is the 3B INT8 diffusion model.
Goes in ComfyUI/models/vae; shared across the SeedVR2 template.
Keep at least twice the model size free before downloading and running.
Check the publisher checksum to rule out a truncated download.
The two safetensors files, by name
The 3B INT8 diffusion model is seedvr2_3b_int8_convrot.safetensors and the shared VAE is seedvr2_ema_vae_fp16.safetensors. These names are referenced by the maintained template. Renaming them to something more readable is the single fastest way to produce a model-not-found error that has nothing to do with the download.
The 7B checkpoint exists as an alternative diffusion model, but start with 3B INT8. It is smaller, faster to validate, and the template is built around it. A heavier model only helps after a lighter pass already shows promise.
The exact folders, in order
Place the diffusion model in ComfyUI/models/diffusion_models and the VAE in ComfyUI/models/vae. These are the folders the native template scans. Put the VAE in the diffusion folder or vice versa and the loader will silently find nothing.
After placing the files, restart ComfyUI and reopen the template so the loaders rescan. Select the exact files inside the loaders rather than assuming the defaults are correct.
- Diffusion model to models/diffusion_models.
- VAE to models/vae.
- Restart ComfyUI and reopen the template to rescan.
Verify the download before blaming the model
A truncated download produces confusing errors that look like the model is broken. When the publisher provides a checksum, verify it before your first run. It is a minute of work that removes an entire class of mystery failures.
Keep at least twice the model size free on disk. SeedVR2 weights are multi-gigabyte, and running out of disk mid-download or mid-run corrupts files and wastes a long queue.
3B, 7B, and the community quantization
ByteDance publishes the official 3B and 7B checkpoints under Apache-2.0. FP8, GGUF, and other quantizations are community or integration choices, not upstream releases. Do not attribute them to ByteDance or turn them into an 8GB guarantee.
Validate the source with the lighter 3B model first. Compare a crop with the 7B only if the lighter output is promising; a larger model can invent more believable artifacts rather than more truthful detail on a damaged source.
After the download, run the sample
Once the files are in place, run the template's bundled sample untouched. A passing sample proves the download and folders are correct. A failing sample means a setup problem, and adding your own source will only add variables.
If the download, folders, and checksum are all correct and the graph still fails, the problem is memory or graph configuration, not the files. The hosted workspace is the right fallback when a finished file matters more than the local download.
Download guide limits
Filenames, folders, and hosting locations change. This guide reflects the Comfy-Org repack and ComfyUI template at review time; always confirm the current template names before downloading to a production machine.
- Template filenames can change between ComfyUI releases.
- Disk requirements scale with model size and batch count.
- Multi-gigabyte downloads still need a stable connection and checksum.
Where the two files come from
Checked November 13, 2026. Both files belong to the Apache-2.0 SeedVR2 model family. The ComfyUI repack hosts the exact VAE file this page names; the upstream model card documents the research release.
- Comfy-Org/SeedVR2 on Hugging Face
The ComfyUI repack that hosts vae/seedvr2_ema_vae_fp16.safetensors (501 MB) plus the diffusion models under diffusion_models/. This is where the two template files are downloaded from.
- ByteDance-Seed SeedVR2-3B model card
The upstream Apache-2.0 research release that SeedVR2 builds on. It documents the official 3B checkpoint; the repack's filenames and folders are the ComfyUI integration layer this guide references.
- Direct download: seedvr2_ema_vae_fp16.safetensors
The 501 MB VAE file itself, at huggingface.co/Comfy-Org/SeedVR2/resolve/main/vae/seedvr2_ema_vae_fp16.safetensors. Use this URL to fetch the exact file this page names, then put it in ComfyUI/models/vae.
SeedVR2 model download questions
Where do I download seedvr2_ema_vae_fp16.safetensors?
From the Comfy-Org/SeedVR2 repository on Hugging Face, under the vae/ folder. The direct link is huggingface.co/Comfy-Org/SeedVR2/resolve/main/vae/seedvr2_ema_vae_fp16.safetensors. The file is 501 MB and Apache-2.0 licensed. Put it in ComfyUI/models/vae, and grab seedvr2_3b_int8_convrot.safetensors from the diffusion_models/ folder at the same time.
Which two files do I need for SeedVR2 in ComfyUI?
The 3B INT8 diffusion model seedvr2_3b_int8_convrot.safetensors and the shared VAE seedvr2_ema_vae_fp16.safetensors. They go into models/diffusion_models and models/vae respectively, with filenames unchanged.
Why does ComfyUI say the model is not found?
Usually the file landed in the wrong folder or was renamed. Confirm the diffusion model is in models/diffusion_models and the VAE is in models/vae, then restart ComfyUI and reopen the template so the loaders rescan.
Do I need to verify the checksum?
When the publisher provides one, yes. A truncated download produces errors that look like a broken model. Verifying the checksum removes an entire class of mystery failures for one minute of work.
Can I use SeedVR2 without downloading any model?
Yes. The hosted SeedVR2.net tool runs in the browser and needs no local download, folders, or checksum verification. Download the local model only when you need a repeatable offline graph.