SeedVR2 Image-Space Multiplier Bridge

This ComfyUI custom node decodes an image latent with its associated image VAE, applies the multiplier in pixel space, then encodes the result with the SeedVR2 VAE.

Inputs

• samples: image LATENT from a compatible sampler
• image_vae: VAE associated with the input image latent
• seedvr2_vae: SeedVR2 VAE
• multiplier: image-space scale factor, default 1.0, range 1.0–4.0, step 0.1
• scaling_method: bilinear (default), bicubic, lanczos, nearest, nearest-exact, or area

What changed

1. Decode the input image latent using image_vae.
2. If multiplier is above 1.0, resize the decoded image pixels internally.
3. Encode that resized image using the SeedVR2 VAE.
4. Return the SeedVR2 latent without any post-encode latent interpolation.

This means a 1440p image at 1.5× is resized to approximately 2160p before SeedVR2 encoding. The target width and height are rounded to the nearest multiple of 8 pixels for VAE compatibility. For example, 2560×1440 becomes 3840×2160.

Workflow notes

• Keep the SeedVR2 conditioning node connected to the bridge's seedvr2_latent output as in the existing working workflow.
• Set multiplier=1.0 to preserve the original no-resize bridge path.
• The interpolation mode now applies to decoded image pixels, not latent feature maps.
• This archive has been syntax-checked and tested with mock VAEs for expected dimensions, but it has not been executed inside the user's ComfyUI installation. Back up the existing node folder before replacing it.

Install

Stop ComfyUI, back up and remove the previous bridge folder so ComfyUI does not load two copies of the node. Extract this archive into ComfyUI/custom_nodes/, creating seedvr2_image_space_multiplier_bridge, then restart ComfyUI.

This is a stopgap solution that still relies on upscaling in pixel space. Proposed solution going forward: Do so entirely in latent space, but this will require training.