SeedVR2 ComfyUI Install Guide: 3B/7B & Missing Nodes

SeedVR2 ComfyUI install starts with ComfyUI 0.28.0 or newer. Restart it, then search the Template Library for a SeedVR2 3B INT8 image or video workflow. The maintained template uses native core nodes, points to seedvr2_3b_int8_convrot.safetensors and seedvr2_ema_vae_fp16.safetensors, and shows their exact folders. This is now the shortest current install path; an old v2.5 custom-node graph should not be your first test.

Download the diffusion model into ComfyUI/models/diffusion_models and the VAE into ComfyUI/models/vae, keep the filenames unchanged, reopen the template, select both files, and queue its bundled sample before adding your own source. A successful sample separates installation problems from source-file problems.

Check the requirements of your chosen ComfyUI build and integration, including GPU support, drivers, system memory, and disk space. ByteDance publishes H100 reference commands. NumZ now documents low-memory community examples; its documentation does not establish the old 18GB floor as a universal minimum. See the sourced example table below before choosing settings.

Keep ComfyUI and custom nodes in a dedicated environment. Updating every dependency at once makes troubleshooting difficult. Record the working ComfyUI revision, custom-node revision, model filename, output resolution, tile settings, and GPU driver before changing anything. That small record is often the difference between a repeatable workflow and an unexplained break after an update.

Use native ComfyUI when you want the maintained 3B/7B image or video templates and normal ComfyUI model loading. Use the v2.5 community custom node only when a known workflow needs its GGUF, BlockSwap, attention, or batch controls. Those extensions have their own version and dependency risks and should stay clearly labeled as community software.

Use the browser workspace when you need one deliverable and local setup is the obstacle. The online path avoids Python, CUDA, model downloads, and custom-node conflicts, but it sends the file to a hosted provider under the published privacy and retention policy. Choose based on the asset and deadline, not on the assumption that local is always free or online is always private.

The separate v2.5 community integration is not the native-template requirement. It splits the workflow into a DiT model loader, a VAE loader, optional torch compile settings, and the main SeedVR2 upscaler node. That split is useful when you specifically need its memory and video controls, but it also adds another package that can break after updates.

The DiT loader is where you choose model family, precision, offload, attention backend, and BlockSwap. The VAE loader is where encode and decode tiling belong. The main upscaler node receives image frames, DiT, and VAE configuration, then controls resolution, max resolution, batch size, temporal overlap, color correction, and noise settings. Keeping that map in mind makes troubleshooting much faster.

ByteDance publishes the official 3B and 7B checkpoints under Apache-2.0. Comfy-Org repackages compatible files for ComfyUI, and NumZ separately maintains a community custom node with additional precision and memory options. FP8 and GGUF are community or integration choices; do not attribute them to ByteDance or turn them into an 8GB guarantee.

A good local process is to validate the source with a lighter model first, then compare one crop with a heavier model only if the lighter output is promising. If the source is a damaged portrait, unreadable text, or a heavily compressed clip, a larger model may create more believable artifacts rather than more truthful detail. Save side-by-side crops so the final decision is visible.

ByteDance reference hardware and a community low-memory workflow describe different software paths. An example for a smaller GPU does not promise that every source, model, or output size will fit. Use the same short source when comparing settings and record failures as well as completed outputs.

Close GPU-heavy applications, avoid multiple queued images, and test the final aspect ratio at a smaller resolution. If the workflow still fails after conservative settings, the browser workspace is a useful fallback because it lets you finish the asset without spending more time on local dependency and memory tuning.

ByteDance documents three important limits: heavily degraded inputs can fail, large motion can break restoration, and lightly degraded inputs can become over-sharpened. Upscaling cannot guarantee recovery of missing text, faces, or objects. Use 2K for the first comparison, and move to 4K only when the source passes crop or playback review.

Portraits with damaged faces, scanned family photos, and heavy grain should start in the old photo enhancer. Transparent PNG assets need careful edge inspection after processing. For client work, compare hair, eyes, small text, product edges, repeating patterns, and flat gradients at 100 percent zoom before approving the result.

An out-of-memory error means the current graph, resolution, or batch is larger than available VRAM. Reduce the output size or batch first, then try tiling or offload. A missing-node error means the workflow references a custom node that ComfyUI cannot load. A model error means the selected filename or directory is wrong. Seams usually point to tile overlap or blending settings. Washed colors or unexpected transparency require checking the input color profile and output format.

If setup keeps failing, save the terminal error, workflow JSON, source dimensions, model, and GPU details. The browser image and video links above offer a separate hosted path while you debug. Use the tool that matches your file, check its quote before submitting, and review the output before accepting it.

For video, SeedVR2 quality is not just about target resolution. Batch size controls how many frames are processed together, which affects temporal consistency. The ComfyUI workflow expects normal video batches to follow a 4n+1 pattern such as 5, 9, 13, or 17 frames. Larger batches can improve consistency and speed on capable GPUs, but they also increase memory pressure.

If a video upscaling job fails on low VRAM, try memory optimizations before reducing batch size below the useful range. For still images, batch size can be one. For video, tiny batches can reduce flicker protection. When the local setup cannot keep a useful batch size, the online video upscaler is the better fallback because it keeps the source in a video-specific lane.

Review the finished image beside the original, not from memory. Check whether the result preserves identity, proportions, typography, logos, and the direction of fine texture. Inspect both the full composition and a 100 percent crop. A larger file is not a better file when edges ring, skin becomes plastic, or small letters change.

Keep the original file, record the model and settings, and download important online outputs before the 30-day creation cleanup window. For a new source type, run one representative sample before committing an entire batch. This keeps local GPU time and online credits predictable.