MiniMax owned the week again, this time on the image side: H3-Image puts 16 megapixel pictures in six steps, and a compressed H3 video model now fits a single 24 GB card. Around them, a 29B world model learned to simulate what happens next, and ComfyUI 0.45.0 pulled most of it into stock nodes.
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 8 to November 14, 2026. The thread running through most of it is the H3 family, now spanning image, video, and the tools around both, with a world model and two video upscalers filling out the rest.
New image models
MiniMax H3-Image: 16 MP in six steps, if you have the disk and the license
On November 10, MiniMax published the weights for H3-Image, an image model built from its H3 video system. One 33.1B checkpoint handles text-to-image and editing alike, producing about 16 megapixels natively, from 4096×4096 up to 6272×2688, in six steps with no CFG and no negative prompt. It takes up to 16 reference images in a single edit.
The catch is size and license. The full BF16 set runs about 135 GB: a 68.65 GB DiT plus the 66.7 GB Qwen3-VL text encoder borrowed from the H3 video repo. Comfy-Org’s repack is kinder, with an int8 DiT at 21 GB and a 15.7 GB NVFP4 encoder, and ComfyUI 0.45.0 supports it natively with official T2I and Edit templates.
Two honest notes before you download. The license excludes the US, EU, UK, and South Korea, and commercial use needs written permission above $1 million in annual revenue, so check it before building anything on these weights. There is also no technical report and no independent benchmark yet, so the “strongest instruction following” claim is MiniMax’s own. See the model page and the ComfyUI repack.
Pruna’s Remove-Object LoRA: Qwen-Image 2.1 edits that leave the rest alone
PrunaAI published P-Q-Image-2.1-Remove-Object on November 10, a rank-32 LoRA that deletes the object you mask from a photo and leaves the rest of the frame where it was. The six-step variant runs on Pruna’s distillation LoRA without CFG at a removal strength of 0.7, and the ComfyUI export ships two ready workflows plus a small node that draws the mask outline the model expects.
This one matters because Qwen-Image 2.1 edits tend to repaint the whole frame: colors shift and the output stops lining up with the input, and few-step distillation makes it worse. This LoRA was trained to hold the pixels outside the object, so the result drops back onto the original at its own size. In Pruna’s examples, a removed person takes their long reflection on a wet plaza with them, even though it sat outside the mask.
It is removal only: adding or replacing objects was never trained and runs no better than the base model. A mask that also covers something you want to keep can delete that too, and filled areas can look smoother than their surroundings or show a faint grid. The LoRA and ComfyUI export are on the model page.
MOSS-Image control LoRAs: seven maps, one 290 MB file each
On November 10, thedeoxen published seven control LoRAs for MOSS-Image-SFT: canny, depth, pose, normal, lineart, scribble, and segmentation. Each is a single rank-64 LoRA of about 290 MB with its own ComfyUI workflow and a control-to-result example grid, so you pick the map you already have and run it.
These are edit LoRAs, not a new control model: the map goes in as the reference image and the prompt decides what the shapes become. Sampling is 30 steps at CFG 4 with the LoRA at strength 1.0, dropping toward 0.7 when the control holds too tightly. The preprocessors in the workflows come from the ControlNet Auxiliary pack, not from core ComfyUI. Start from the canny LoRA; the other six live alongside it under the same author.
New video models
HyperFlow H3 Flash: H3 cut to 13.7B, in six steps, on one GPU
Video Rebirth released HyperFlow H3 Flash on November 6, a compressed MiniMax H3 that samples in 4, 6, or 8 steps at CFG 1. It drops three of the base model’s 50 transformer blocks and factors 102 linear layers, landing at 13.7B parameters, and the 9.1 GB Q4_K_M GGUF runs on a 24 GB card with stereo audio intact.
The headline speedup needs a second look. Video Rebirth claims 12.7x over the H3 baseline on a single B200, but Kijai measured roughly 10 percent over the pruned full model in ComfyUI, because the attention cost stays the same at this resolution. By the release’s own numbers, a 5.2-second clip takes 92 seconds on an RTX 5090. LoRAs trained on the full H3 do not load on the compressed transformer either.
It still fills a real gap: a few-step H3 that fits consumer hardware, with text-to-video, first-frame, first-and-last-frame, and reference-to-video modes, the last taking up to nine images. The license is MiniMax’s H3 community license, which excludes the US, EU, UK, and South Korea. Weights and code are on the model page.
PAN 2: a 29B world model that simulates what happens next
MBZUAI’s Institute of Foundation Models opened PAN 2 on November 10, a 29B world model that simulates up to 20 seconds of 704×1248 video at 24 fps from text or a starting frame. A frozen LLM first rewrites your request as a structured world-state caption, then a 29B MMDiT predicts the video latents, and a stock Wan 2.2 VAE decodes them.
A community ComfyUI node appeared the next day, loading the diffusers-layout weights directly, with a 33.7 GB fp8 repack for smaller setups. The technical report benchmarks six simulation capabilities plus Physics-IQ, PhyGround, and RBench, with a 1,400-pairwise human study behind it. The authors are direct about the weak spot: about two in three simulated multi-step plans contain at least one step the model does not carry out.
Appearance also drifts across long sequences, since each clip simulates from at most one starting frame plus text. Still, a world model with open weights, training configs, and a working ComfyUI node on day two is a strong release. The weights are on the model page.
ComfyUI workflows and tools
ComfyUI 0.45.0: Qwen3.x text generation twice as fast, H3-Image native
Version 0.45.0 landed November 10, six days after 0.44.0, and the headline is speed: Qwen3.x text generation was rewritten, with Qwen3.5 9B on an RTX 5090 going from 106 to 219 tokens per second with adaptive tree MTP and INT8 attention. MiniMax-H3-Image support is now in core, so the model above works with stock nodes and official templates.
RelayVSR joins core too, with two nodes for the Wan 2.2-based video upscaler, and WeVisOne-VL adds detection, masks, and alpha mattes from a single text query. The partner-node list keeps growing: GPT Image V3 and 2.6, Seedream 5.5, Breeze TTS 2, Pika H3 Flex, Kling 4.0 Flash, and 4K Seedance 2.5. The frontend is now 1.58.5. Full changes are on the release page.
ReCaVSR and RelayVSR: two one-step video upscalers reach ComfyUI
ReCaVSR, a one-step video super-resolution model from USTC built on Wan2.2 TI2V-5B, got its ComfyUI port on November 10. It upscales a clip 4x with one denoising step per block, installs from ComfyUI Manager, and ships a 6.3 GB int8 file that lets a 1440p 4x pass run on an 8 GB card.
The numbers are solid for a one-step model: a 960×480 clip to 3840×1920 takes 59 seconds on a 4090 in bf16, 36 seconds with the int8 file, and a 360×360 portrait to 1440×1440 finishes in 15 seconds. RelayVSR, the companion upscaler, landed in ComfyUI 0.45.0 core the same day with x2 and x4 repacks, though an open bug report flags a progress-bar crash in the new code. Weights are on the model page, with nodes at ComfyUI-ReCaVSR.
Bottom line
This week the H3 family finished swallowing the pipeline: the video model became an image model, the compressed video model fit a single GPU, and ComfyUI absorbed both plus two video upscalers. The practical story is that 16 MP images and 4x video upscales are now local jobs, as long as you have the VRAM and the license allows it where you live.
Two things to watch: Qwen4-Max Preview opened November 11 on Token Plan only, with no weights, model card, or license, and it ends November 25, so Qwen3.8 stays the newest Qwen you can download. A November 13 roundup put Wan 3.0 top of the video arena at 1175 Elo, with H3 the first open-weights model in the top three at 1161, but Wan 3.0 is API-only, so H3 is the best open-weights video you can run. 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 Gets Day-One Support in ComfyUI with Open Weights
FastH3 V2: 8-Step MiniMax H3 Video Generation in ComfyUI
How to Use FlashVSR Upscaling in ComfyUI
AI Image & Video Model News Roundup – November 7, 2026
Leave a Reply