Do I have to use the provided VAE file?

Yes—use seedvr2_ema_vae_fp16.safetensors as recommended by the model authors. A mismatched VAE can shift color/contrast or reduce detail fidelity.

Is the result deterministic or do I need to set a seed?

SeedVR2 is a one-step restoration model and is deterministic in this setup. There’s no random seed involved, so you’ll get repeatable results for the same input and settings.

What’s the benefit of the 3B Int8 version?

Int8 quantization lowers VRAM requirements and speeds up inference, making the model more accessible on mid-range GPUs. Expect slightly lower peak fidelity than full-precision weights, but typically with minimal visible difference.

Can I use this for full videos?

Yes. Export your video to frames, process them with LoadImage → SeedVR2 → SaveImage, then reassemble. Although the graph processes frames individually, SeedVR2’s training often yields stable frame-to-frame results.

Image Upscale: SeedVR2 3B Int8

This workflow demonstrates how to upscale and restore images using SeedVR2 3B Int8, a one‑step diffusion-based video restoration model adapted here for still images and frame-by-frame processing. The core of the graph is a SeedVR2 inference node (id: 1970dcb4-982a-4c01-a373-4d239a3041e4) configured for the 3B Int8 weights and paired with the provided VAE (seedvr2_ema_vae_fp16.safetensors). You load your source with LoadImage, route it through the SeedVR2 node for enhancement/upscaling, compare results with ImageCompare, and finalize with SaveImage. A MarkdownNote in the graph provides inline guidance and links.

Technically, SeedVR2 performs a single forward pass to restore detail, reduce compression artifacts, and upscale, avoiding iterative denoising. The Int8 quantized 3B variant reduces VRAM usage and speeds up inference with a small quality trade-off versus full precision. While SeedVR2 is trained for video restoration and improved temporal consistency across sequences, it also produces strong results on individual images. If you process frame sequences consistently (e.g., exported video frames), the model’s training tends to yield visually stable outputs from frame to frame even when used one-at-a-time in ComfyUI.

API

Use this workflow from code

Every Comfy workflow is a JSON graph. The payload below is this workflow, exactly as ComfyUI runs it — fetch it from the URL, keep it in version control, or load it in ComfyUI and run it node by node.

ff35fea1d184.json

Fetching workflow JSON…

Run it from TypeScript or Python with the Comfy SDK. The same code targets Comfy Cloud or a ComfyUI you host yourself — only the base URL changes.

// Install (beta)
npm i @comfyorg/sdk

// Run this workflow (TypeScript)
import { Comfy } from "@comfyorg/sdk";

const client = new Comfy({ apiKey: "comfyui-..." });
const wf = await client.workflows.fromFile("workflow_api.json");
const job = await client.run(wf);
await job.getOutputs("<output-node-id>")[0].toFile("output.png");

The SDK takes a workflow in API format: open this workflow in ComfyUI and use File → Export Workflow (API).

Comfy Cloud API access requires a plan with an API key.

SDK docs Get an API key

FAQ

Frequently Asked Questions