⭐ Star Tiled SeedVR Upscaler

SeedVR2 is, by a wide community margin, the best detail upscaler you can run locally - a one-step diffusion restorer that adds pores, eyelashes and fabric weave instead of just pixels, Apache 2.0 and fast enough to sit in a normal workflow. The catch, for most people, is VRAM: raw SeedVR2 on a big image is a one-way ticket to OOM city. Star Tiled SeedVR Upscaler is the fix - the entire SeedVR2 pipeline (preprocessing, VAE encode, conditioning, single-step sampling, decode, color correction) running tile by tile with feathered blending, so a 3x upscale that would crush an 8GB card becomes routine.

How it works

The pipeline inside this one node is worth spelling out because it explains all the knobs:

1. Your image is upscaled by scale with Lanczos to the target resolution.
2. That big image is split into rows × cols tiles with tile_overlap of shared border.
3. Each tile runs through SeedVR2: padding, tiled VAE encode, conditioning built from the tile latent, single-step Euler sampling (CFG 1.0, denoise 1.0), tiled VAE decode, then a LAB color correction against the source tile so tiles don't shift hue.
4. The processed tiles are blended back with linear feathering over the overlaps - that's why you don't see seams.

It also includes the color-preservation feature the SeedVR2 scene cares about: color_luminance_weight (default 0.8) blends between matching the source's colors and keeping the model's own brightness. Lower it toward 0 for more reference color matching, keep it high to preserve the original's exposure.

Inputs that matter

• model_name - the SeedVR2 diffusion model from models/diffusion_models (e.g. seedvr2_7b-int8_convrot.safetensors). Download it yourself - the repo ships no weights.
• vae_name - the SeedVR2 VAE from models/vae (e.g. ema_vae_fp16.safetensors).
• scale (default 2), rows/cols (default 3×3), tile_overlap (default 0.25) - the VRAM/detail dials. More tiles = less VRAM per step, more total passes.
• model_override (optional) - a MODEL socket that, when connected, replaces the dropdown-selected model. Handy for patched models or custom loaders without editing the dropdown.

Output: the upscaled IMAGE.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

Restart, search star. Manager users can search "Starnodes". Then drop the SeedVR2 diffusion model and EMA VAE into your models folders - the official SeedVR2 ComfyUI repo has download links and the exact file names this node expects.

Where people get burned

Model-placement mistakes are the #1 failure - empty dropdowns mean the files aren't in the folders the node scans. On quality: SeedVR2 adds a slight contrast lift that can make output read as "more AI" than the input; most people compensate by keeping color_luminance_weight near 0.8 and doing final contrast tweaks downstream. And remember the upscaling-ladder rule: if your source is already sharp and you just want pixels, this is the wrong tool - that's ESRGAN's job. SeedVR2 is for adding detail that wasn't there, and it will rewrite faces a little, so give recognizable faces their own pass.