A week dominated by MiniMax H3: Pika distilled it to six steps, NVIDIA’s Sana team opened up its 2K/4K refinement stage, and ComfyUI core absorbed the fast H3 upscaling path. On the image side, Krea 2 learned to put panels and text in exact boxes, and MOSS-Image editing got a stronger community rebuild. Here are the highlights from October 25 to October 31, 2026. This revised edition links each item to its actual source: the model page, the vendor post, or the code itself.
New image models
Krea 2 BBOX: comics and posters with layout you can specify
A community finetune of Krea 2, finished October 27, teaches the model a compact grounding language: every panel, character, object, and line of text gets a box on a 0 to 1000 grid, written as plain text lines in the prompt. The result is multi-panel comics where speech bubbles land inside their own panel, characters stay consistent across panels, and headline text goes where you asked. It ships in two flavors: Turbo, which samples in 8 steps with CFG off, and Raw, which samples in 50 steps at CFG 5.5, each with two EMA weight variants. LoRAs at rank 64, 128, and 256 exist too, but the author is upfront that they carry only a small share of the full finetune’s effect, so the full 26 GB single files are the version to use for precise layouts. Weights, prompting spec, and samples are on the model page.
Anima Reference Adapter: one image, a new scene, no pose copying
Released October 27 as a pilot, the Anima Reference Adapter keeps a character and its drawing style from a single reference image while the prompt sets pose, framing, and background. The trick is in how the reference is attached: it is fed as a clean, unnoised frame with its position coordinates shifted past the target grid, so the model cannot just continue the reference as the next shot. A rank-32 LoRA (about 92 MB) plus three ComfyUI nodes does the job, with separate text and reference guidance scales. In held-out tests it reproduced hair, eyes, outfits, and accessories well, while painterly artist styles transferred weakly, a known limit the author plans to address with separate character and style references. The pilot lives in the GitHub repo, with weights on Hugging Face. Note: the repo has since added a faster v2 adapter, but the October 27 news was the pilot.
MOSS-Image Edit HQ v5: a rebuilt edit adapter with a bigger reference budget
The community MOSS-Image editing capability got a from-scratch rebuild, published October 25 after five training arms. Version 5 raises the reference budget from 144 tokens to 256, uses pixel-diff masks so unedited regions stay close to the source, and trains on 5,644 rows with full scene captions. It ships in two ComfyUI layouts: a native LoRA that works with the experimental Text Encode MOSS-Image Edit node (needs a current master build), and a plugin-format file for the ComfyUI-Moss-Image-Edit node pack. Stacking it with the 8-step DMD speed LoRA cuts sampling to 8 steps, though the author says the 50-step pass still wins on quality. One honest caveat: it was trained on single-reference pairs, so compositing a person from one photo into another is not what it was trained for. Details and downloads are on the project page.
New video models
Pika H3 Flex: MiniMax H3 distilled to six steps, with synced audio
The headline release of the week. On October 29, Pika published H3 Flex, its post-trained version of MiniMax H3, distilled from 50 steps down to six and refined with reward learning and preference optimization. It generates 5 to 15 second clips at 480p or 768p and 24 fps with synchronized stereo audio, and it accepts up to 12 references (up to 9 images, 3 videos, and 3 audio clips) for character, motion, and edit control. For ComfyUI it arrives as two LoRA adapters, FL2VA for text and first/last frames and Ref2VA for multimodal references, loading through the stock LoRA loader with three official template workflows. Pika claims it ranks first among open-weights models on DesignArena’s text-to-video leaderboard. Two things to know before downloading: the adapters are small (about 2 GB each in ComfyUI format) but they sit on top of the roughly 210 GB H3 base, and the 1080p to 4K upscaling and five separate audio stems stay Pika-hosted features. See Pika’s launch post and the weights on Hugging Face.
Open-H3-Regenerate-2K: the missing H3 refinement stage, opened up
MiniMax H3 shipped in two stages, but only the first had open weights: the second stage, which redraws a finished 768p draft at higher resolution, existed only as a hosted API. NVIDIA’s Sana team rebuilt it as Open-H3-Regenerate-2K, a single refiner LoRA on a frozen H3 transformer, published October 27. It takes a draft at 384p to 768p and returns a native 768p, 1080p, 2K, or 4K clip in just three denoising steps, decoding through a fast 24-layer H3 VAE. The team reports a 5-second clip refined from 768p to 2K in 22.45 seconds on a single B200, against 87.26 seconds for the standard path, and says blinded judges preferred the open refiner in 87 percent of comparisons. There are no official ComfyUI nodes yet, but Kijai published the refiner as its own rank-384 LoRA and community conversions exist, with a warning that conversions must apply the SwiGLU transformation or the weights come out broken. Weights, code links, and benchmarks are on the model page.
ComfyUI workflows and tools
ComfyUI core absorbs H3 latent upscaling and the Sol-H3-VAE decoder
Two core pull requests merged October 28 move H3’s high-resolution path out of custom node packs and into stock ComfyUI. One adds a dedicated H3 latent upscale node that grows the latent directly, skipping the decode, pixel upscale, and re-encode round trip, while letting the audio half pass through unchanged. The other teaches the VAE loader to open Sol-H3-VAE, NVIDIA’s distilled 24-layer decoder, which the model card says cuts a 2K decode from 14.5 seconds to 0.89 seconds on a GB200. Both landed after the v0.43.1 tag, so they are in master and nightly builds rather than the current release. Pair them with the refiner above and the full local H3 pipeline, draft to 2K, is now almost entirely stock nodes. The changes are PR 17412 (latent upscale node) and PR 17411 (Sol-H3-VAE loader support); the weights are the H3 latent upscaler and Sol-H3-VAE.
Pruna’s 6-step Qwen-Image 2.1 LoRAs get an official ComfyUI pack
PrunaAI’s v0.2 update on October 27 adds two 6-step adapters for Qwen-Image 2.1, and, more importantly for ComfyUI users, an official ComfyUI export with BF16 single-file adapters and six example workflows. The September release shipped only as diffusers/PEFT files whose scale ComfyUI’s LoRA loader did not read, so this fixes a real friction point. One of the two new adapters is the more interesting one: it distills Qwen-Image 2.1’s prompt upsampler into the LoRA itself, so a short prompt goes straight into sampling with no separate enhancer call. The plain 6-step adapter is tuned for editing where the input must be preserved. Both run at strength 1.0 with Euler at CFG 1 and their own ManualSigmas schedules, which are not interchangeable with the older v0.1 schedules. See the model page; the official ComfyUI export lives in the same author’s P-Q-Image-2.1-Comfy repo.
Bottom line
This week the H3 ecosystem crossed a threshold: six-step generation, three-step refinement, and core support for fast upscaling and decoding mean a full local H3 pipeline is now practical in ComfyUI, even if the base model itself still demands serious hardware. On the image side, the theme is control: boxes for Krea 2 layouts, a reference adapter that holds character without copying pose, and a stronger MOSS editor. One watch-out: Alibaba’s newer Wan models (2.6 onward, including Wan 3.0) remain hosted-only as of this week, so Wan 2.2 is still the newest Wan with downloadable weights. See you next Saturday.
Leave a Reply