ComfyUI had one of those weeks where the changelog reads like three unrelated projects. Core merged a video deflicker node and rewrote the optical flow backend behind it. A music node pack grew a setup wizard, an API and an MCP server. And one of MiniMax's open weight image models landed natively with 4096 pixel output. Here is what shipped, and which parts are worth updating for today.
One Node Kills the Video Shimmer
If you have ever upscaled a clip frame by frame and watched the result crawl like heat haze, this is the fix you have been assembling by hand. ComfyUI core merged a Video Deflicker node on November 15, 2026 at 00:20 UTC in pull request 17754, written by kijai. It removes temporal flicker from per frame upscales, restyles and edits. The same PR replaces the optical flow backend with a native RAFT large implementation, which the author measures at roughly 2.3x faster than torchvision at 1024 by 576.
Under the hood it runs the classic blind video temporal consistency method from Bonneel et al., with a fidelity term that keeps every frame close to its input. Per frame, the node computes RAFT flow from a guide clip, warps the previous output frame along that motion, then solves a screened Poisson system on the GPU. Stable regions stop crawling while real motion survives.
The controls are fewer than the math suggests:
- guide: optional stable source frames at any resolution. Feed the pristine low res clip here and your upscaled frames into the images socket, so motion comes from the clean source instead of the shimmer you are removing.
- strength: default 1.0, range 0 to 100. Higher values remove more flicker and soften more fine detail.
- direction: bidirectional by default, which averages a forward and backward pass and avoids trailing behind motion at twice the compute; forward is cheaper.
- threshold: default 0.1. Changes larger than this count as real motion and get smoothed less.
The backend swap matters as much as the node. The native RAFT loads the same state dict layout strictly, so your existing raft_large.pth keeps working, and it runs the update loop as a CUDA graph under dynamic VRAM. The VOID warped noise nodes share the same loader, so they inherit the speedup with no change on your side. Install is one file: drop torchvision's raft_large weights into models/optical_flow, then wire Load Optical Flow Model into Video Deflicker. The merge landed after v0.45.0 on November 10, so it reaches the next tagged release and runs today on master or nightly.
Plenio 2.0 Turns the Rig Into a Song Studio
Plenio Music Production System 2.0 shipped on November 14, three weeks after 1.5, and it is less about better songs than about surviving first contact with a new machine. The pipeline is unchanged: a local language model writes lyrics, YuE2 or MiniMax Music 3 sings them, and you get a mastered track with cover art, timed lyrics and sheet music. What 2.0 adds is installation logic and remote control.
- A Setup tab reads your graphics card, plans the model files a song needs, downloads only what you confirm, verifies every file against its SHA-256, then renders a one minute test song before unlocking Home.
- Every template now writes auto into its model fields, so a workflow built on one computer picks files that suit another.
- Song API v2 covers all five templates, keeps jobs alive across a ComfyUI restart, stops at review points, streams events and issues tokens for other machines.
- An MCP server, plenio_mcp.py, lets Claude Desktop or Claude Code request a song and have it play in the Plenio Player, which has its own HTTP controls for stream decks and home automation.
- People and access adds Host, Co-host, Maker, Listener and Party guest roles, QR code invites, guest pages in seven languages, and a Director that changes visuals on the first beat of each section.
It needs ComfyUI 0.41.0 or newer, up from 0.37.0. Install with comfy node install comfyui-plenio-music, or through ComfyUI Manager. Version 1.6.0 stays installable on older builds, and workflows saved with 1.x open and run as they are.
MiniMax H3 Image Arrives With Open Weights
On November 10, MiniMax H3 Image became native in ComfyUI, and it also runs through Partner Nodes. One checkpoint handles both text to image and reference editing: connect images and it is an edit graph, leave them disconnected and it is text to image, same weights either way. It is CFG distilled and samples in exactly six steps at guidance 1, so there is no negative prompt to write.
Output is generated at size, not upscaled into it: 4096 by 4096 at 1:1, 4736 by 3552 at 4:3, 5024 by 3360 at 3:2, 5472 by 3072 at 16:9 and 6272 by 2688 at 21:9, with the middle three also offered in portrait. It shares its Qwen3-VL text encoder with the H3 video model, so anyone already running H3 video owns that file. Multi reference editing names inputs as Picture 1, Picture 2 in prompt order, prompts accept up to 4096 tokens, and width and height left at zero default to 2048 by 2048 or follow the first reference image. Connect the VAE while editing to anchor references as latents, not just text encoder context.
Three releases, one pattern: ComfyUI is absorbing the glue. Deflicker removes the need for an external consistency pass, Plenio removes the manual model hunt, and H3 Image removes the API bill for local users. The deflicker node is the one to pull today, because it is free, small and fixes a problem every per frame workflow has. The other two are worth an update when you have the disk space and a free evening.
Comments