MiniMax H3 kept branching in every direction this week: Pruna turned the video model into a fast image maker, Viggle taught it to reshoot finished clips, and a set of tiny dial files now steers motion and mood. On the image side, Qwen-Image 2.1 dropped to four steps and MOSS-Image got a fix for half-precision runs.
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
NVIDIA opened four world-simulation models for driving and robotics, and ComfyUI 0.44.0 pulled most of this week’s news into stock nodes. Here are the highlights from November 1 to November 7, 2026. The H3 ecosystem is the thread running through most of it.
New image models
P-H3-Image: a video model rebuilt as a fast image maker
The headline release of the week. On November 3, Pruna AI published P-H3-Image, an image model distilled from MiniMax H3, the video model that has dominated this column all autumn. Instead of training from scratch, Pruna shrank each part: the DiT from 33B to 20B parameters, the text encoder from 32B to 4B, and sampling from 50 steps to 8.
It makes 8-megapixel images in 8 steps with no CFG and no negative prompt, and it edits too: one reference image to change, or up to 8 to combine, driven by a short plain request such as making the pickup truck red. A built-in Qwen3.5-9B prompt expander turns a one-line request into the detailed description the model draws best from.
Two things to know before downloading: the full-quality version is about 100 GB and wants a data-center card, while the ComfyUI version is about 20 GB and runs in 21 GB of VRAM. The license is MiniMax’s community license, which excludes the EU, UK, South Korea, and the US unless you hold a separate license. ComfyUI supports it natively from 0.44.0. See the model page and the ComfyUI files.
Lightning v2 RL: Qwen-Image 2.1 in four steps, polished by trial and error
Published November 2, LightX2V’s Lightning v2 RL is a full replacement for the Qwen-Image 2.1 DiT that finishes in 4 steps, down from the original 40. It handles three jobs in those 4 steps: text-to-image, single-image editing, and multi-reference composition.
The RL stands for reinforcement learning: the model was tuned by testing its own outputs and keeping the changes that scored better, rather than only imitating examples. Unlike the earlier Lightning v2 adapter, a 67 MB LoRA, this release ships the whole DiT, so you swap it in place of the original and keep the same text encoder and VAE.
The authors show side-by-side examples where the 4-step version matches the 40-step original. One license note: it falls under the Qwen Research License, which has non-commercial terms, so check it before using the pictures in paid work. Details and downloads are on the model page.
MOSS-Image-heal: a repair that makes fp16 safe
MOSS-Image had a quiet flaw: a few of its activation channels ran so hot that the largest value reached 58,411, while fp16 tops out at 65,504. Push it slightly harder and the numbers overflow to infinity, and the picture comes back black or corrupt. That is why the official ComfyUI repack shipped bf16 and int8 files but no fp16 one.
On November 4, KohakuBlueleaf, the author of LyCORIS, published MOSS-Image-heal, a full retraining of every weight that spreads those hot channels out. The largest value drops to 3,083, about 21 times below the fp16 ceiling, while the pictures stay essentially the same as the base model.
It ships as three single files that load through the normal model loader: bf16 and fp16 at 10.2 GB each, and an int8 version at 5.4 GB. If you run MOSS-Image on a card where fp16 is the fast path, this is the version to use. Weights and reports are on the model page.
New video models
NVIDIA Cosmos 3 Nano: four open models for driving and robot simulation
On November 5, NVIDIA released four open-weight checkpoints in its Cosmos 3 family, all built for interactive world simulation: generating video of a place that responds to actions, instead of just playing a fixed clip. Each is a version of Cosmos 3 Nano, a 16B model tuned for a different job.
Transfer-Auto creates driving footage across up to 11 camera views at once, including fisheye lenses, with matching LiDAR data. Sim-Auto is the streaming version: it generates the next chunk of camera and LiDAR views as the driving controls change. Sim-Transfer turns simulator controls like depth and outlines, plus a text prompt, into realistic-looking video.
Sim-Bimanual predicts video from the actions of a two-armed robot, for training robot policies without a real robot. The license is OpenMDW-1.1, which allows commercial and non-commercial use with attribution. The catch is hardware: NVIDIA puts Nano at 96 GB of VRAM or more, so this is workstation and server territory. The announcement is on the Hugging Face blog.
Viggle Remake: reshoot a finished video, keep the motion and the sound
Published November 4, Viggle Remake is a pair of LoRAs for MiniMax H3 that re-render video you already have. Swap mode replaces the people in a clip with different characters, while Edit mode changes objects, backgrounds, weather, and lighting. In both modes the original motion, camera moves, timing, and soundtrack stay as they were.
You paint one to four key frames to show what should change, and the model carries the change through the whole shot. It ships with ComfyUI nodes and int8 and GGUF versions, so it runs from 12 GB of VRAM. The key files are about 5.5 GB, and there is also an MCP server that lets coding assistants cut a long clip into windows, render each one, and stitch the result back together.
Two honest notes: each render window covers about 12 seconds, and the release includes no quality measurements, only curated demos. Also, this tool puts a person into footage they did not shoot, so only use it on your own material or material you have permission for. The weights and ComfyUI guide are on the model page.
ComfyUI workflows and tools
ComfyUI 0.44.0: H3 editing nodes, P-H3-Image support, and ProRes video
Version 0.44.0 landed November 4 as the biggest ComfyUI release in a while. The headline is native P-H3-Image support, so the new image model above works with stock nodes and official workflow templates. It also adds the MiniMax H3 Max editing nodes: Recast, Extend, Reframe, Insert, Lip Sync, and Camera Control.
The H3 work from last week is now in core too: the latent upscaler, support for the fast Sol-H3-VAE decoder, and context windows that carry information across shots. For video saving, ProRes encoding arrives with chroma subsampling options. The interface is now at version 1.58.1.
There is also a node to set the VAE compute precision, Flux Video Upscale V2 targets at 1080p, 2K, and 4K, and new partner nodes including OpenAI’s GPT-6.1 Terra and Sol. Full changes are on the release page. Worth updating for the H3 nodes alone.
H3 Dials: 14 KB sliders that steer motion and mood, no training needed
Published November 6, H3 Dials is a new way to control MiniMax H3 without training anything. A “dial” is a tiny file, about 14 KB, made from a pair of opposite prompts, like “frozen still” versus “running and jumping.” Load it into a ComfyUI workflow and a slider moves every generation toward one side or the other.
The motion dial runs from near-frozen, about 0.05 times the normal movement, to about 2.7 times. There is also a villainy pack of eight dials: sinister, creepy, cold, unhinged, seductive, corrupted, dread, and impish, each pushing the mood of a scene in its own direction. The dials work on Pika H3 Flex as well as the base model.
The project also includes a “de-rope” workflow that finds parts of a clip moving too fast for the frame rate, re-renders them slowed down, and maps them back to real time. One warning from the author: the sinister dial can crush dialogue audio at higher strengths, so keep it low when characters are talking. The dials and the ComfyUI node pack are on the model page.
Bottom line
This week the H3 ecosystem filled in around the edges: the base video model is now an image model, a video reshooter, and a set of mood sliders, and ComfyUI 0.44.0 makes most of it work with stock nodes. On the image side the theme is speed: Qwen-Image 2.1 in 4 steps, P-H3-Image in 8, and a 4K editor in 4, all without giving up much quality.
Two things to watch: several of this week’s releases inherit MiniMax’s community license, which excludes the US, EU, UK, and South Korea, so check the license before building anything commercial on H3 derivatives. And Alibaba’s Qwen4, expected around now, still has no published weights, license, or model card as of this week, so Qwen3.8 remains the newest Qwen you can actually download. See you next Saturday.
Further Reading
More from kombitz.com, related to this week’s items and popular on the site:
How to Use Qwen-Image 2.1 GGUF in ComfyUI
MiniMax H3 Video VAE Gets Up to 2.2x Faster in ComfyUI
FastH3 V2: 8-Step MiniMax H3 Video Generation in ComfyUI
How to Use Flux.2 Klein 9B GGUF in ComfyUI
MiniMax H3 Gets Day-One Support in ComfyUI with Open Weights
Leave a Reply