Creative AI Briefing October 19, 2026 morning edition 10 min read

Over the weekend the open community did the finishing work on last week's image models, sharpening, speeding up and porting them, so the useful version of a release now arrives three days after the launch.

Kromoss MOSS-Image MFlux FLUX 3 Image Klein Fish Audio Drama 3 AuK image-gen open-weights local-creative-ai comfyui audio-gen lora-finetuning

Creative AI Briefing: Monday, October 19, 2026

You can now download a sharper version of last week's open poster model, run it in eight passes instead of fifty, and sell what you make with it, because a community trainer released Kromoss overnight under Apache 2.0. On a Mac, you can now hand FLUX 3 Image Klein up to ten reference photos and get back one composed edit, using a free command-line tool that added support three days ago. The big labs posted nothing creative this weekend. The community spent it finishing their work: fixing the blur in one model, porting another to Apple Silicon, and shrinking a third so it fits on less memory. The pattern is worth planning around. The release-day version of an open model is now the rough draft, and the one you want tends to show up by day three.

New models

Kromoss is a sharper MOSS-Image you can use commercially. The trainer lodestones published it on Hugging Face on October 19 (repo created 01:34 UTC). It is a full fine-tune of MOSS-Image-Base, the starting-point version of the two OpenMOSS released on October 16, and the card says plainly that it "mainly fixes the Base model's blurriness." What you can make: the same layout-aware posters, title cards and text-heavy graphics MOSS-Image does, where you place each block of words with a box, with crisper detail. It drops in where MOSS-Image-SFT goes (same text reader, same image decoder, same pipeline), so existing MOSS workflows should load it unchanged. It ships with an 8-pass fast add-on alongside the full model, which normally takes 50 passes. Licence: Apache 2.0, so client work and resale are fine. Where it runs: Diffusers or ComfyUI on your own card; single-file weights and a conversion script are included, and no hosted provider serves it yet. The catch: lodestones' own comparison (CLIP scores on 100 PartiPrompts, the author's figures) puts Kromoss with the 8-pass add-on at 27.82, against 27.53 for MOSS-Image-Base with the same add-on and 27.19 for Base at 50 passes, while Kromoss alone scores 26.88. Those rows were measured at an earlier checkpoint (update 2000) than the one shipped (update 4000). CLIP scores measure prompt match, not sharpness, so judge the blur fix with your own eyes. The card also says to apply the fast add-on unmerged, because its changes are small enough that rounding during a merge can erase them.

Image

MOSS-Image's long layout prompts got faster. p1atdev/MOSS-Image-SFT-KV-Cache-LoRA (created October 19) is a small correction add-on that lets ComfyUI reuse the text part of the work instead of recomputing it every pass. On a 2,941-token layout prompt on an RTX 6000 Ada, the author measured 7.87 seconds per image normally and 5.26 seconds with the cache, about 1.5 times faster. The card warns that short prompts can be slower, that the same seed will not reproduce a normal render, and that it was trained on 10 authored prompts and validated on 3, so quality evaluation is limited. Apache 2.0.

A 4-pass MOSS-Image in one file. BassionAI/BSAI-MOSS-Image-SFT-4steps-turbo-merged (created October 19) merges the SFT model with a 4-pass fast version into one 9.47 GB ComfyUI file. The card is mostly in Chinese and suggests starting at 4 passes. It is tagged Apache 2.0 but tells you to check the upstream terms before commercial use.

FLUX 3 Image Klein now edits from ten references on a Mac. MFlux v0.23.0 (October 16) added mflux-generate-flux3 and mflux-generate-flux3-edit, covering text-to-image, edits from one to ten reference images, and 4, 6 and 8-bit shrunken versions, plus loading of add-ons in three common formats. On October 19, mflux-community posted ready-made conversions; the 4-bit file is about 5.3 GB on disk against 18.5 GB at full size. The model notes are the useful part: on an M4 Pro with 64 GB, the 4-bit version peaks at 18.5 GB of memory against 30.7 GB at full precision, but each step takes slightly longer, not less (7.0 against 6.4 seconds). A 1024 image at full precision takes about 105 seconds including load; the 4-bit version runs closer to two minutes. Shrinking buys you room, not speed. Licence: FLUX Kommunity License, which allows commercial output for users under its revenue threshold.

Klein reached AMD, untested. Harsha001/flux-3-image-klein-ROCmFPX (created October 19) offers 2.5 to 5.5 GB files aimed at AMD cards, but they need a forked build of stable-diffusion.cpp, and the card admits they "have not been tested on AMD hardware yet." Wait for someone to report numbers.

Video

Streaming picture-and-sound clips between two frames, if you own a server rack. Alibaba's TaoLive team released TaoMate-H3-FL2VA (created October 15), an add-on for MiniMax H3 that takes a first and last frame and generates synced video and audio in chunks, aimed at long continuous takes with a steady face and voice. It is research-grade: the validated setup is eight 96 GB data-centre GPUs, under the MiniMax H3 Community License. Score 2: interesting direction, no creator access yet.

Audio and music

Fish Audio Drama 3 on about 30% less memory, with an asterisk. dangvansam/drama-3-pruned-128e-fp8 (created October 19) removes half of the model's internal specialist sub-networks, keeping the 128 used most on 1,700 Vietnamese sentences, with no retraining afterward. Memory under sglang-omni falls from 46.2 to 32.2 GiB (29.9 GiB under vLLM-Omni). The author's small test (120 requests, two voices) shows error rates slightly up and calls the gap "inside the noise," but warns rare voices, accents and emotion tags may degrade, and that only Vietnamese was checked. It is still a 32 GB job. Fish Audio's community licence carries over: commercial use only under US$500,000 annual revenue, and no hosted voice service without a separate licence. Not an official Fish Audio release.

AuK speech editing moves toward the Mac. Tencent's AuK, released September 9, generates and edits speech from one plain-language instruction: change a word, the pitch, the emotion, or turn a line into a whisper. On October 18, mlx-community posted Apple Silicon conversions for the mlx-auk-swift project. Second-day work, dated by repo creation. Read the licence carefully: AuK's own weights are MIT, but the MLX card notes that the required Qwen2.5-Omni-3B encoder is research-only, so a working pipeline is not commercial. The hosted Space shows as running on shared hardware.

Open and local

The local story in plain terms: if you held off on last week's releases because they were blurry, slow or did not run on your machine, each of those objections now has a published answer. Each answer is days old and lightly tested.

lodestones/Kromoss: the sharper, Apache-licensed retrain of MOSS-Image-Base for posters and text-heavy graphics. Why now: posted October 19 with an 8-pass add-on (model).
p1atdev/MOSS-Image-SFT-KV-Cache-LoRA: speeds up long layout prompts in ComfyUI by about 1.5 times on the author's test. Why now: posted October 19 (model).
filipstrand/mflux: the free Mac command-line tool for open image models, now running FLUX 3 Image Klein with ten-reference edits. Why now: v0.23.0 on October 16, 2.5k stars (releases).
OpenMOSS/MOSS-Image: the code, ComfyUI plugin and user guide behind the layout model everything above builds on. Why now: 818 stars three days after release (repo).
KnifeLemon/Filee: a free offline drag-and-drop converter for images, video, audio and documents. Why now: still on today's Trendshift list (repo).

Creative workflows

1. Draft fast, finish slow with Kromoss: Kromoss model card

The steps. Download diffusion_models/kromoss_bf16.safetensors and 8-steps-dmd-lora.safetensors into ComfyUI. Build or reuse a MOSS-Image layout prompt. For drafts, load the add-on unmerged and render at 8 passes with guidance at 1 and no negative prompt. When a layout works, remove the add-on and render the final at 50 passes, guidance 3, shift 5, the card's defaults.

How it works. The add-on teaches the model to jump to a finished image in fewer passes. Keeping it unmerged preserves its small changes, which rounding can erase.

Why it is good. You iterate on layout at a fraction of the wait, then spend the full render time only on the version you keep, all under a licence that allows client work.

Where it breaks. The 8-pass result and the 50-pass result are different images, so a draft is a guide, not a preview. The fine-tune is one day old and the author's numbers measure prompt match, not sharpness.

2. Speed up a long poster layout: p1atdev KV-cache card

The steps. Install the comfy-moss-image-kv-cache custom node the card names. Load lodestones' frozen 8-pass add-on at strength 0.7, then the correction add-on at 1.0. Add the MOSS-Image Text KV Cache node with text_timestep 1000. Sample with Euler, simple schedule, 8 passes, guidance 1.

How it works. Long layout prompts make the model reread the text every pass. The node stores that reading once; the add-on corrects the small errors storing it introduces.

Why it is good. Dense posters with many boxes are exactly where waiting hurts most, and that is where the gain shows up.

Where it breaks. Loading the add-on alone does nothing; the node is required. Short prompts can run slower. Seeds will not match normal renders, and other add-on strengths or pass counts sit outside what it was trained for.

3. Compose an edit from up to ten references on a Mac: MFlux FLUX 3 notes

The steps. Install with uv tool install --upgrade mflux. Accept the FLUX Kommunity License on Hugging Face and log in. Run mflux-generate-flux3-edit with your reference images (one to ten), writing the prompt as a full caption of the finished picture with the edit included. Add -q 4 if memory is tight. For placement, append a JSON list of elements with [top, left, bottom, right] boxes on a 0 to 1000 grid.

How it works. The distilled model runs 16 passes and reads every reference at once; the output size defaults to the first reference.

Why it is good. Multi-reference product and character composites, free, offline, on a laptop you already own.

Where it breaks. Shrinking saves memory but not time. Negative prompts are not supported and exit with an error. There is no prompt rewriter, so short prompts give generic results. The undistilled Base model takes about 1,400 seconds for a single-reference edit.

Worth testing

MOSS-Image Space: test a layout JSON free before downloading Kromoss. Tradeoff: it runs the original model, not Kromoss, on shared hardware with queues.
Kromoss in ComfyUI: run the same layout through MOSS-Image-SFT and Kromoss and compare at 100% zoom. Tradeoff: a full model download, and no hosted way to try it first.
MFlux with -q 4: a three-reference edit on any Apple Silicon Mac with room to spare. Tradeoff: about two minutes per image, and shrinking does not make it faster.
AuK Space: try an instruction edit on a short voice clip. Tradeoff: the pipeline's encoder licence rules out commercial use.

What actually matters from today's signal

Stop judging open releases on launch day. MOSS-Image went public on October 16, and its Base version was the one people found soft. Three days later a trainer had a fix for the blur, someone else had cut the wait on long prompts by a third, and a third person had collapsed it to four passes. FLUX 3 Image Klein followed the same arc onto the Mac. If you evaluate on day one, you judge the draft.

For your time, the gain is iteration speed: 8-pass drafts and cached prompts mean more layouts tried per hour. For your money, the Apache licence on Kromoss is the headline. It keeps the commercial freedom of MOSS-Image while improving the version most people found unusable.

The risk is that every one of these fixes is a single person's work, tested lightly and sometimes measured with numbers that do not capture what you care about. The Drama 3 cut was checked only in Vietnamese; the AMD Klein files were not checked on AMD at all. Treat day-three community releases as promising, and keep the original model installed until your own tests agree.

Source access notes: Adversarial fact-check pass ran (Sonnet subagent, all load-bearing claims checked). It caught: the KV-cache add-on's 10 prompts were its training set (plus 3 validation), not its test set; Drama 3's memory saving is about 30%, not half (half is the expert count); the Kromoss comparison omitted the like-for-like Base-plus-add-on row (27.53) and the fact that scores came from an earlier checkpoint; OpenMOSS never called Base raw, so that framing was removed; 'retraining' corrected to 'fine-tune'; the 105-second MFlux figure is full precision, not 4-bit. Unverifiable and softened: yesterday's star count comparison. Bucket A primary sources checked: OpenAI news (no creative posts since October 15), blog.google (no dated creative posts), Adobe blog (nothing creative October 13 to 19; Adobe MAX runs November 10 to 12), BFL blog (latest October 15), Stability news (latest August 25), Luma blog (listicles October 14), ElevenLabs blog (latest September 30), Suno blog (latest October 13). runway.com/news rendered empty (JS). Midjourney, Kling and ByteDance Seed surfaced nothing new via WebSearch. fal blog latest September 17; fal models page lists H3 Max Relight, Recast and Reference Director as "new" but undated, and Relight was already integrated elsewhere by October 8, so they were omitted. Replicate blog latest April 15. blog.comfy.org returned only the Substack shell; the ComfyUI GitHub releases page served stale data (August). GitHub trending served an obviously stale list and was ignored. HN Algolia results stopped at October 12 and were not used. Hugging Face papers listed nothing creative. All Hugging Face dates are from API createdAt. Civitai skipped.