Another busy week in open weights. Black Forest Labs dropped a tiny FLUX 3, a new Chinese lab shipped a model that takes layout as data, ComfyUI pushed two core releases back to back, and the few-step race keeps accelerating. Here are the highlights from October 11 to October 17, 2026.
New image models
FLUX 3 Image Klein: 5.6B open weights that run on 8 GB GPUs
The week’s biggest release came from Black Forest Labs: FLUX 3 Image Klein, a 5.6B-parameter open-weights model released on October 15 for text-to-image, pixel-perfect image editing, and multi-reference generation with up to 10 reference images. It shares the FLUX 3 family’s trick of bounding-box control: draw a box and the model places elements (or confines edits) where you say. Native 4K generation is in there too. The INT8 build runs in ComfyUI on as little as 8 GB of VRAM, with base and step-distilled checkpoints plus FP8 and NVFP4 builds developed with NVIDIA. Day-one tooling includes ComfyUI, Diffusers, Ostris’s AI Toolkit, and NVIDIA NeMo fine-tuning recipes. It ships under the FLUX Kommunity License, so check the terms if you plan to ship commercially. Coverage at Unite.AI and a hardware-focused breakdown at dev.to.
MOSS-Image: layout control as JSON, same-day ComfyUI support
From OpenMOSS (October 16) comes MOSS-Image, a 5B diffusion transformer that treats layout as data. The headline feature is a JSON layout mode: you describe the scene and place text boxes and objects with bounding boxes, and the model obeys, down to exact text at exact coordinates. It renders legible text in English and Chinese, accepts prompts up to 4K tokens, and ships in two checkpoints: Base (built for LoRA fine-tuning) and SFT (the one most people should run). The kicker is the ComfyUI story: core support merged the same day, including a MOSS-Image Layout Prompt node that converts Create Bounding Boxes output into the model’s layout JSON, plus Comfy-Org repacks (SFT in BF16 at 10.2 GB or INT8 at 5.4 GB). Read the full breakdown and note the recommended settings: 50+ steps, CFG 3, shift 5.
New video models
TaoMate-H3-FL2VA: first/last-frame control for streaming MiniMax H3
Alibaba’s TaoLive AIGC team released TaoMate-H3-FL2VA on October 15: a 3-step LoRA that adds first/last-frame control to the streaming TaoMate-H3 runtime for MiniMax H3. The claim is the first real-time streaming first/last-frame-to-audio-video model: you pin down the first and last keyframes and the 3-step denoising blocks generate a 30-second synchronized audio-video clip that has to land exactly on your endpoint. ComfyUI users get it easy: the FL2VA conditioning path already exists in core, and community conversions (Kijai’s rank-averaged 169 MB build, austinit’s full rank-128 BF16) landed the same day. Known weakness of the family, per early testers: fast, large movements still degrade. Details at the ComfyUI Wiki.
SMD: one distillation recipe, 4 steps for everything
SMD (Simplified Matching Distillation, arXiv 2610.18472) is a regression-only distillation recipe published October 13 that ships students for image, video, joint audio-video, and 3D, all at four sampling steps. The headline students for this crowd: MiniMax H3 T2VA and Ref2VA in 4-step and 8-step versions with ready-to-run ComfyUI workflows, plus LoRAs for FLUX.1-dev, Qwen-Image, SD3.5-Medium, FLUX.2 Klein 4B, and Wan2.1. It is unusually honest about limits: the quality claims are the authors’ own, and the ComfyUI packages need the complete bundled workflow (substituting H3 weights into default templates fails with a missing audio_scale error). The Wiki write-up covers all the packages. In the same spirit, LightX2V’s Lightning LoRA collapses Qwen-Image 2.1 from 40 steps to 4, with side-by-side comparisons in the model card.
ComfyUI workflows and tools
ComfyUI v0.40.0 and v0.41.0: Flux3 native, LightVAE, LoRA stacks
Two core releases landed back to back. v0.40.0 (release notes) brought LoRA stack loaders, a Camera Angle node, Qwen 3.5 low-VRAM improvements, and extended comfy_attention support. v0.41.0 (release notes) followed within days with native FLUX.3 Image Klein support (base and step-distilled checkpoints, new layout prompt node, four bundled workflow templates), MiniMax-H3 LightVAE support for lighter local H3 decoding, and SDR-to-HDR conversion. The official changelog tracks the daily node updates. One caution from the community: some users reported slowdowns after the v0.40.0 update, so as always, back up your install before upgrading your production box. Note that v0.41.0 is quite buggy, use v0.41.2 instead.
FluxSR: one-step 4x super-resolution LoRA for Flux.1-dev
An ICML 2025 project finally published its weights: FluxSR is a one-step 4x real-world super-resolution LoRA (717 MB) built on Flux.1-dev, combining flow trajectory distillation with a total-variation LPIPS loss and an attention diversification regularizer to fight the speckle that plagues one-step outputs. The full training loop is open, not just inference: training pairs, prompts, and evaluation scripts all ship. A community ComfyUI node pack appeared the same day (October 13), though it installs manually from a channel zip rather than a repo. The Wiki coverage has the timings: a 446×448 input to 1784×1792 in about 3 seconds on an 84 GB machine.
Open Source Higgsfield: $25,000 to rebuild closed apps in the open
Comfy Org opened the Open Source Higgsfield challenge on October 15: $25,000 in prizes for the best open-source rebuilds of subscription-only creative apps (they name Genjutsu and Cinema Studio) on the Comfy Developer Platform. The grand prize is $10,000 cash, and there is a dedicated RTX 5090 category for builds using open-weight models only. Bonus tracks come from LTX ($2,000 for an LTX-centered VFX workflow), Black Forest Labs (FLUX API credits), and MiniMax. Submissions close October 30 at 9am PT, with winners announced November 6. Full details and entry rules.
Bottom line
This week was about control and speed in equal measure. FLUX 3 Image Klein and MOSS-Image both push the frontier of small, local, controllable image generation, one with bounding boxes and pixel-perfect edits, the other with layout-as-JSON and strong text rendering. On video, the few-step race is now a three-way sprint: SMD, TaoMate’s 3-step LoRA, and LightX2V’s Lightning adapter are all compressing the heavy models down to a handful of steps. ComfyUI keeps pace with native support days after release, and if you have a subscription app you wish were open, Comfy Org will pay you $25k to rebuild it.
Leave a Reply