What makes SeedVR2 7B Int8 different from traditional upscalers?

SeedVR2 is a one-step diffusion upscaler trained with adversarial objectives and adaptive window attention. It enhances detail in a single forward pass (no iterative sampling), offering fast, consistent results compared to multi-step diffusion or classic GAN upscalers.

Does Int8 quantization reduce quality compared to FP16?

Int8 substantially lowers VRAM and can improve speed. In most cases, quality differences from FP16 are minor, but extremely fine textures may show subtle changes. If you have ample VRAM and need maximum fidelity, try an FP16 variant; otherwise, Int8 is a strong default.

How large can I upscale in one pass?

That depends on your GPU memory and the node’s settings. Moderate upscales (e.g., 2×–4× area increases) are typically reliable. Very large jumps may soften details or introduce artifacts; if needed, upscale in stages by running the output back through the node.

Do I need to set steps, samplers, or CFG?

No. SeedVR2 is a one-step model, so there are no sampler steps or schedulers to configure. Focus on selecting a sensible output size and evaluating results with ImageCompare.

Image Upscale: SeedVR2 7B Int8

This ComfyUI tutorial workflow shows how to upscale a single image using SeedVR2 7B Int8, a one-step diffusion model trained with adversarial objectives and adaptive window attention. Instead of iterative sampling, SeedVR2 performs a single forward pass to enhance resolution and restore detail, making it fast and consistent. The Int8 quantized weights reduce memory usage while preserving most of the model’s fidelity, which is helpful on mid-range GPUs.

Technically, the graph is straightforward: LoadImage feeds your source into the SeedVR2 7B Int8 node (ID: 1970dcb4-982a-4c01-a373-4d239a3041e4), which wraps the SeedVR2 checkpoint and its paired VAE (seedvr2_ema_vae_fp16.safetensors). The model outputs an upscaled image in a single step, which you can preview with ImageCompare (original vs upscaled) and then export via SaveImage. A MarkdownNote in the graph provides quick tips and references. Because it’s a one-step diffusion upscaler, you don’t need to tune sampler steps or schedulers—just set your desired output size (or keep defaults), run, and review.

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.

5f0ab474a355.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