AI Image & Video Model News Roundup – November 21, 2026

No new base model stole the week. Instead the ecosystems around the last two big ones filled in: MOSS-Image got a 4-step LoRA and a mask editor, MiniMax H3 got a talking avatar that holds up past ten minutes, and ComfyUI 0.46.0 pulled PAN 2 into core. Seven items, all of them about making last month’s models faster or easier to steer.

If you’re thinking about purchasing a new GPU, we’d greatly appreciate it if you used our Amazon Associate links. The price you pay will be exactly the same, but Amazon provides us with a small commission for each purchase. It’s a simple way to support our site and helps us keep creating useful content for you. Recommended GPUs: RTX 5090, RTX 5080, and RTX 5070. #ad

Here are the highlights from November 15 to November 21, 2026. The pattern holds across image and video alike: the base models are settling in, and the community is now doing the speed and control work on top of them.

New image models

MOSS-Image 4-step LoRA: 50 steps down to 4, texture intact

On November 19, lvladikov published a rank-64 LoRA that renders MOSS-Image-SFT in 4 steps at CFG 1 instead of 50 at CFG 3. That cuts the model calls from 100 to 4, and on the author’s own timing 22 pictures at 1024×1024 took 26.6 seconds of billed GPU time against 387.3 for the teacher, about 14 times less.

It was trained by distribution matching against the 50-step teacher, with a critic comparing student pictures to the teacher’s and to about 8,000 real photo crops. The headline number is fine texture at 1.16 times the teacher’s on the author’s tests, and text rendering holds up too: 14 of 15 prompts spelled every requested string, against 15 for the teacher.

Two things to get right before you run it. Use the ComfyUI file at strength 1.0, not the plain one, because ComfyUI’s bf16 rounding eats part of the change, and use the LoRA’s own four sigma values rather than the simple scheduler. Early reports also note weaker photoreal faces and the occasional missing limb, so treat the author’s numbers as a starting point. The LoRA is on the model page.

MOSS AnyPaint: mask inpainting and outpainting for MOSS-Image-Base

On November 13, yijunwang2 published MOSS AnyPaint, a rank-32 adapter that gives MOSS-Image-Base the inpainting mode its official card never had. One runtime handles same-canvas inpainting, one-pass outpainting onto a larger canvas, several disconnected edit regions at once, and narrow or curved masks that cross the edge of the source.

ComfyUI nodes shipped the same day from alexw5702-afk, reaching v0.2.0 on November 14, and the recommended setup pairs the adapter with the Base 8-step DMD2 LoRA at 8 steps, CFG 1 and shift 5. The adapter itself is a 142 MB file under Apache 2.0, which makes it one of the few fully permissive image-editing releases this month.

The honest limits are on the model card. This is a functional adapter, not a plain LoRA, so ordinary LoRA loaders and the auto-generated diffusers snippet will not run it. Large generated areas can pull in unrelated subjects, unmasked pixels can shift slightly through VAE decoding, and every showcase source is synthetic, so judge it on your own files. The adapter is on the model page.

Seedream-style LoRA brings a closed model’s look to H3-Image

On November 19, Nebsh published a rank-16 style LoRA that bends MiniMax H3-Image toward the contrast, lighting and grain of Seedream 5.5, the closed model the author calls their current favorite. It trained in two stages, 2,000 steps on 43 images then 4,000 more on an expanded 1,134-image set, and runs with the trigger seedreamstyle. at strength 1.0, LCM, 6 steps and CFG 1.

The evidence is a 20-scene comparison gallery, each board showing H3-Image without the LoRA, the 2K and 6K checkpoints, and a real Seedream 5.5 generation side by side. The author also tried the 6K checkpoint on H3 video as a bonus experiment, with matched A/B clips for text-to-video and reference-to-video. It is a style experiment, not a clone, and the H3-Image community license still applies. The LoRA is on the model page.

New video models

LongCat Video Avatar 1.6: talking avatars that hold up past ten minutes

On November 13, Meituan released LongCat Video Avatar 1.6, an audio-driven avatar model that turns one photo and a voice track into lip-synced video past ten minutes long. The update targets the two things that break long talking videos: mouths that look strained or over-red and flicker, and faces that drift from the original as segments pile up. It also adds a faster 8-step path and an optional INT8 DiT at about 15 GB, against 30 GB for BF16.

Meituan ran its own arena of 33,051 pairwise human votes against four commercial systems, and 1.6 came second with an Elo of 1068.1, behind OmniHuman-2.0 and ahead of HeyGen and Kling Avatar 2.0. Treat that as a vendor claim: the margin over 1.5 is only 16 Elo, and no independent check exists yet. The weights are MIT, so commercial use is allowed. The model is on the model page.

The ComfyUI path arrived five days later. Comfy-Org repackaged the weights on November 18, including the int8 DiT at 14.9 GB, both LoRAs and the Whisper encoder, and a community node pack runs the model today with 8-step distilled and 50-step base nodes. Kijai’s core support is still open for review as PR 17831. The port’s author measured about 162 seconds for a 10.1-second clip at 768×512 on a single 4090. The repack is on the Comfy-Org page.

ComfyUI workflows and tools

ComfyUI 0.46.0: PAN 2 goes native, plus a video deflicker node

Version 0.46.0 landed November 18, eight days after 0.45.0, and the two local additions that matter are both from Kijai. PAN 2 support is now in core, so MBZUAI’s 29B world model loads through stock nodes, and a new Video Deflicker node smooths per-frame processed video with a native RAFT optical flow implementation the release notes call 2.3 times faster than torchvision.

The rest is mostly partner nodes and housekeeping, and most of those nodes call hosted APIs rather than running locally. LTX gets AlphaGen, SDR-to-HDR, Reframe and Upscale nodes, MiniMax gets an H3 Image node, and Kling 4.0 Image, HeyGen Avatar 6 and Seedance 2.5 Flash join the list. The frontend is now 1.59.6, DirectML support is gone, and ComfyUI warns on GPUs that support CUDA 13 but are not running it. See the release page.

Wuli’s Gemma-4 extender writes H3 prompts for you

On November 16, Wuli-Art released a Gemma-4-12B fine-tune that rewrites a plain idea into the audiovisual prompt MiniMax H3 expects, covering shot, action over time, camera movement and soundscape. One weight set serves both input formats, FL2VA and Ref2VA, so you do not need a different extender for first-and-last-frame work and for reference-conditioned generation. It ships in bf16 at 23.9 GB and fp8 or int8 at 13 GB.

Two ComfyUI workflows come with it, one per format, and they use core nodes only, though you need a build with TextGenerateV2 and Gemma-4 support, which 0.46.0 has. The model card runs a 45-clip five-way comparison against no extension, plain Gemma-4, MiniMax’s official expander and lightx2v’s rewriter, and the enhanced prompt comes back in English even from a Chinese request. The model is on the model page.

NKD Event Curve times the action in H3 clips

NKD Basic Tools v1.27.0, out November 15, adds an Event Curve node that decides when things happen in a MiniMax H3 clip. You draw a curve, pick one of six guides, punch, flow, motor, blast, uncanny or words, and the node turns it into an audio guide that H3 follows while it generates picture and sound together. A test clip with eight peaks produced eight impacts.

It can also copy timing instead of drawing it: wire in audio and the node follows its loudness, so a hit lands on a beat, or wire in video frames and it follows how much the picture moves. Start at strength 0.5 and lower it if the result looks forced, and note that one words guide per character is not supported yet. The pack is on the repo page.

Bottom line

This week the base models stood still and the tooling moved. MOSS-Image went from a 50-step model to a 4-step one with a mask editor on top, H3 went from a video model to a talking-avatar and prompt-timed one, and ComfyUI absorbed the world model that was a community node a week ago. If you run these models locally, the week was mostly good news about speed and control.

Two things to watch: Qwen4-Max Preview closes on November 25 with still no weights, model card or license in sight, so Qwen3.8 stays the newest Qwen you can download. And Black Forest Labs’ FLUX 3 Dev, announced back on July 23 with no date, is still unreleased as of November 11, so the open FLUX line remains the Klein models. See you next Saturday.

Further Reading

More from kombitz.com, related to this week’s items and popular on the site:

How to Use Krea 2 Turbo GGUF for Inpainting in ComfyUI

MiniMax H3 Video VAE Gets Up to 2.2x Faster in ComfyUI

How to Generate Video Prompts in ComfyUI with Qwen3-VL

FastH3 V2: 8-Step MiniMax H3 Video Generation in ComfyUI

AI Image & Video Model News Roundup – November 14, 2026

How to Use Flux.2 Klein 9B GGUF in ComfyUI

How to Download Hugging Face Models Using Command Line

Be the first to comment

Leave a Reply

Your email address will not be published.