ComfyUI v0.45.0 for video makers: RelayVSR video upscaling, H3-Image, Seedream 5.5 and Breeze TTS 2 nodes

RelayVSR, natively. RelayVSR is an Apache-2.0 research upscaler from Xijun Wang and co-authors. Per the ComfyUI pull request, it runs a one-step Wan2.2 5B pass on every 15th frame and a small 'Flash' decoder that rebuilds the frames in between from those keyframes and the low-res video, so the expensive model touches only a fraction of the clip. It ships as two nodes (RelayVSRConditioning and VSRDecode, which has a detail control) with no text encoder, in 2x and 4x variants. CG Pixel's tutorial uses it on MiniMax H3 clips and argues it keeps the original detail better than generative upscalers; that is the creator's comparison, not ours.

New model nodes. Native MiniMax-H3-Image (open-weights text-to-image and editing; see our separate post), plus partner (API) nodes for ByteDance's Seedream 5.5 image model and Breeze TTS 2 voices and text-to-speech. ByteDance and early coverage pitch Seedream 5.5 on edits that stay stable over many rounds, transparent PNG output and better small text; those are vendor claims. Partner nodes bill through Comfy's API credits.

Also in this cycle. Kling 4.0 Flash with an original-audio option, 4K output for Seedance 2.5, Gemini Omni 1.2 Flash in the Omni video node, Pika H3 Flex and soundtrack/SFX nodes, WeVisOne-VL support, texture rebaking with UltraTex for 3D, and a 'Get Video Components' fix that stops decoded frames being held twice in memory.

Not yet: the countdown ComfyUI teased on Nov 8 had not launched when we checked; the new @ComfyDevs account said this morning that its 'OS Higgsfield' winners announcement is coming today.

Our take: if you finish open-model video in ComfyUI, update and try RelayVSR on one short clip next to whatever upscaler you use now (SeedVR2, FlashVSR or a paid node). Signals: release only, so not hot.