SeedVR2 Image-Space Multiplier Bridge

A ComfyUI node that decodes a source image latent with image_vae, resizes the decoded pixels in image space, and re-encodes them with seedvr2_vae for the SeedVR2 refinement path.

Inputs

• samples: source LATENT (4-D image latent or 5-D single-frame latent with T=1)
• image_vae: VAE used to decode the source latent
• seedvr2_vae: SeedVR2 VAE used to encode the resized image
• multiplier: image-space scale multiplier (default 1.0)
• scaling_method: resize method (default Lanczos)
• short_side: optional base short-side resolution in pixels (default 0, which uses the decoded image's current short side)

The original widgets remain in their original order (multiplier, then scaling_method). The new short_side widget is appended after them to preserve saved workflow widget values when updating the node.

Output seedvr2_latent connects to the SeedVR2 sampler/refinement path. Connect output_resolution to a Show Text node if you want to display the calculated output dimensions.

Output dimensions are calculated from the decoded image aspect ratio, scaled by short_side × multiplier (or by the decoded image's own short side when short_side=0), then rounded to the nearest multiple of 16.

Example: source image 3392×1440, short_side=1440, multiplier=1.2 → 4064 × 1728.

Fixed an error when scaling with Krea2. Added reference_image output (for use with Post-Process SeedVR2 Output node)

New: added a manual short side field that is automatically multiplied. 0 sets to default behaviour based upon input latent. Added report output to view the final resolution.

Existing workflow compatibility (v12)

This version includes a small ComfyUI frontend migration that repairs an invalid/missing short_side value to 0 when an older saved workflow is loaded. It also repairs an invalid multiplier value to 1.0. Valid saved values are left unchanged. The original widget order remains multiplier, scaling_method, then short_side.