Creative AI Briefing October 26, 2026 morning edition 11 min read

Files you keep on your own drive got faster and better over the weekend while hosted models were switched off on someone else's schedule, so the looks a creator depends on are safest in files they own.

GPU Unlock MOSS-Image-Edit-HQ MOSS Manga Alpha Image LoRA Krea Audio gpt-image-1 Seedance 2.5 4K local-creative-ai creative-hardware image-gen ai-editing comfyui audio-gen

Creative AI Briefing: Monday, October 26, 2026

If you own an older RTX 3090 Ti, one owner's test says you can render the same MOSS-Image poster about 58 percent faster tonight, pixel for pixel identical, by loading a free Linux kernel module that went public eight days ago. Over the same weekend a community editor for MOSS-Image went from version 4 to version 6, and a new add-on started drawing manga characters with real transparent backgrounds. Meanwhile OpenAI switched off gpt-image-1 on October 23 and Krea pulled Veo 3 and Seedance 1.5 Pro from its menus. The pattern for a working creator: the tools you keep as files keep improving, and the tools you rent can vanish on a date you did not pick.

New models

MOSS-Image-Edit-HQ reached version 6, with its failures on the record. akhaliq/MOSS-Image-Edit-HQ (created October 22 per the Hugging Face API, refreshed October 26 at 00:39 UTC) is an unofficial add-on that turns the free MOSS-Image poster model into a photo editor: change an outfit, remove an object, relight a room, swap text, from a one-line instruction. What makes it worth a look is the paper trail. The model card lists six training rounds and says version 1 "lost the eval gate" to the earlier chinoll add-on on keeping faces the same; later rounds added masks that hold unedited areas close to the source and, in v6, captions that describe the finished frame. Use armV6_workflow_single.json for one reference photo or armV6_workflow_multi.json for up to four. License, plainly: Apache 2.0, the same as MOSS-Image, so client work is allowed. Where it runs: ComfyUI through the Chenkin-x node pack (ComfyUI 0.42.3 or newer) or natively on a nightly build; the card asks for about 20 GB of graphics memory and says it will not run on 6 GB. Free demo: moss-image-edit-hq-demo runs both a quality mode and an 8-pass fast mode. The honest catch: every comparison grid is the author's own, combining people or objects from two photos is a weak spot because training used single-reference pairs, and the card says the fast mode weakens added objects and can tint images red.

Image

Manga with transparent backgrounds, no masking. paralaif posted MOSS-Manga-Alpha-Image-LoRA on October 26 (04:48 UTC per the API), a follow-up to Saturday's proof of concept. End a prompt with "isolated on a transparent background" and it paints a character, creature, prop, ink splash, sound effect or speech balloon as a PNG with a real alpha channel, ready to drop onto a comic page. It ships with the four-channel decoder and a small custom node the stock loader needs, plus a ComfyUI workflow and 30 example PNGs. License: Apache 2.0 across the add-on, node and workflow. It was trained on one RTX 5060 Ti 16 GB card from 503 synthetic images. The catch, per the card: smoke and ink sprays can pick up a faint purple tint, dense black subjects and full scenes are harder, and vague prompts like "office worker" can flip gender. A Krea 2 version, Krea2-Manga-Alpha-Image-LoRA, went up the same minute.

Retirement notice for anyone with old image pipelines. OpenAI's deprecations page lists October 23 as the shutdown date for gpt-image-1, announced April 22, with gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest following on December 1. Krea's changelog confirms the knock-on effect: its "ChatGPT Image" option is retired and the old API route now returns "410 Gone." If a client template, a batch script or a no-code automation still calls the old name, it stopped working on Friday, and the replacements OpenAI names (gpt-image-2.5-sunburst or gpt-image-2.5-flare; Krea routes to gpt-image-2, without its presetStyles option) are different models, so expect saved prompts to need retesting.

Video

Seedance 2.5 4K for film work, enterprise trials only. ByteDance's cloud arm Volcano Engine released Seedance 2.5 4K on October 23, according to TechNode (secondary source; no primary announcement found). The useful idea for creators is the draft mode: test shots as cheap 480p previews, then render the keeper in 4K with 10-bit colour, which holds up better in a grade. The catch: selected enterprise customers only, no price, no public access date.

Model menus are shrinking as well as growing. Krea's October 24 entry retires Seedance 1.5 Pro (old links now open Seedance 2.0 Mini) and removes Veo 3 and Veo 3 Fast "because Google no longer serves them." Recreating an old job swaps in a default model with a notice. If a series depends on one of those looks, your existing renders are now the only copies of it.

Audio and music

Krea opened its Audio tool to every signed-in user (October 24). The changelog adds an Audio option inside Train that clones a voice "in seconds," shows a cost estimate before each generation, and lays voice-design choices side by side. Krea's audio docs list ElevenLabs TTS, ElevenLabs Music, ElevenLabs Video To Music, Qwen TTS, Qwen Voice Clone, ThinkSound and MMAudio inside its node canvas, so a video, its score and its voiceover can come out of one workspace. The catch: neither page states prices, commercial terms, length limits or a consent rule for cloning someone's voice. Read your plan's terms before you clone anyone, including yourself for client work.

Open and local

The local story this week is hardware you already own. A Linux kernel module called GPU Unlock lets GeForce RTX 30-series cards run their BF16 and FP16 maths (the number formats most image and video models now use) at the rate NVIDIA reserves for its professional cards. On a 3090 Ti, the project's own benchmark moves BF16 from 87.5 to 157.4 trillion operations per second. The creator-relevant number comes from one 3090 Ti owner's ComfyUI test, reported by ComfyUI Wiki: MOSS-Image SFT in BF16 rendered 58.6 percent faster and MiniMax H3 video in BF16 60.2 percent, a FLUX.1 dev FP8 file gained 54.6 percent, and 150 matched still images came out pixel-identical. The 8-bit INT8 files were mixed: MOSS-Image INT8 barely moved (0.3 percent) while one H3 INT8 file gained 31.2 percent. Gains depend on the file type, so test your own. A fork for the RTX 4090 followed on October 23.

Zagitta/gpu-unlock: makes BF16-heavy image and video renders run faster on RTX 3060 to 3090 Ti cards under Linux, outputs unchanged. Public October 18, still near the top of Trendshift's daily list; about 8k stars (shields.io), GPL-2.0 module (repo).
JustAResearcher/gpu-unlock-ada: the same idea for the RTX 4090 (experimental for the 4070 Ti Super). Created October 23; the README reports matrix-maths gains of roughly 12 to 34 percent (FP16) and 16 to 20 percent (BF16) in one owner's tests, flat at one size, says these are not whole-app speedups, and notes higher power draw (346 W vs 313 W); 3.3k stars (shields.io) (repo).
akhaliq/MOSS-Image-Edit-HQ: instruction photo edits on MOSS-Image, v6 posted overnight, Apache 2.0 (card).
paralaif/MOSS-Manga-Alpha-Image-LoRA: transparent manga characters, props and balloons as PNGs, Apache 2.0, created October 26 (card).
elayadesign/ai-design-skills: two MIT skills for coding agents, one that builds landing pages and design-to-motion, which turns a static design into a looping animated showcase exported as MP4 and GIF. On today's Trendshift list; 6.6k stars (shields.io) (repo).
Nutlope/hallmark: an anti-generic-design skill for Claude Code, Cursor and Codex, back on Trendshift's daily list; MIT (repo).

Creative workflows

1. Draft fast, finish slow: photo edits with MOSS-Image-Edit-HQ v6. Source: model card, workflows workflows/armV6_workflow_single.json and armV6_workflow_multi.json.

The steps. Install the Chenkin-x node pack on ComfyUI 0.42.3 or newer, load MOSS-Image SFT, and put armV6-moss-edit-lora.safetensors in models/model_patches/ (or use the native path on a nightly build with edit_format set to ai-toolkit). Import the single-reference workflow. For drafts, add lodestones' 8-pass speed add-on at CFG 1 and try three or four wordings of your instruction. When one reads right, remove the speed add-on and render at 50 passes, CFG 3, Euler, shift 5.0. Raise CFG to 3.5 or 4 only if the edit comes out too timid.

How it works. The add-on reads your photo as the thing to keep and your sentence as the change. v6 was trained with region-weighted masks, so pixels outside the edit stay close to the source.

Why it is good. The fast pass is roughly six times quicker per the card, which makes it cheap to find the instruction wording that works before you spend a full render.

Where it breaks. The card says fast mode weakens added objects and can push a red cast, so never approve colour from a draft. Combining two photos is weak. Base and Kromoss compatibility is unverified. You need about 20 GB of graphics memory.

2. Transparent manga assets straight from the prompt. Source: MOSS-Manga-Alpha-Image-LoRA card, workflow comfyui/moss-manga-alpha-image-t2i.json.

The steps. Copy the bundled ComfyUI-RGBA-VAE folder into custom_nodes/, the RGBA decoder vae/flux2_klein_alpha_rgba_vae.safetensors into models/vae/, and the add-on into models/loras/. Download the MOSS-Image SFT int8 model and text encoder from Comfy-Org/MOSS-Image and lodestones' 8-pass add-on. Load the workflow and keep the card's settings: strength 1.0, 8 passes, er_sde sampler, CFG 3.5, 832×1216 for portrait figures. Describe one subject and end with "isolated on a transparent background."

How it works. A decoder with a fourth channel lets the model paint transparency directly, so the edge of a hair strand is partly see-through instead of cut out.

Why it is good. Speech balloons, sound effects and spot characters land on a page with no masking pass, and everything is Apache 2.0.

Where it breaks. Purple tint in smoke and ink spray, heavy black subjects, full backgrounds, and gender drift on vague nouns. Check every asset over white, black and mid-grey before lettering.

3. A before-and-after speed test on an RTX 30-series card. Source: GPU Unlock README and the ComfyUI Wiki test summary.

The steps. On Linux, render five images in ComfyUI with a fixed seed and a full-size (BF16) model, and note the median time. Install the module with make install (or install-deb, install-rpm, install-arch), add gpu_unlock.targets=all to the kernel command line, and power the machine fully off before booting. Confirm with journalctl -k -b | grep gpu_unlock. Rerun the same five renders and compare times and pixels in a Difference blend.

How it works. The module sets the card's speed selectors before NVIDIA's driver loads, so it survives driver updates.

Why it is good. If your card is supported, it is free speed on hardware you already paid for.

Where it breaks. Linux only; an unsigned kernel module (per ComfyUI Wiki); cold boot required; hibernation not covered; one OEM 3080 Ti stayed at half speed; most 8-bit INT8 files see little gain; ComfyUI Wiki says to "treat it as experimental." Back up first.

Worth testing

MOSS-Image-Edit-HQ demo: edit your own photo in quality or fast mode with no install. Tradeoff: a shared queue, and the Space may lag behind the newest version.
Manga Alpha example gallery: 30 transparent PNGs to inspect before you download anything. Tradeoff: the author chose them.
Krea Audio: music, voiceover and voice cloning beside your images and video. Tradeoff: prices and commercial terms are not on the docs page.
GPU Unlock: if you run Linux on a 3090 or 3060. Tradeoff: low-level system software with a small chance of a card that will not start until power-cycled.

What actually matters from today's signal

The weekend made one thing plain: there are two clocks in creative AI. Open files run on the community's clock, which moves fast and in your favour. MOSS-Image went public ten days ago and now edits, cuts out manga with transparency, and, in full-size BF16 files, runs close to 60 percent faster on a used RTX 3090 Ti, per one owner's test. None of that cost a subscription. Hosted models run on the vendor's clock. gpt-image-1 was OpenAI's first GPT Image model; on Friday it stopped answering.

For a creator's time and money, the lesson is to keep the parts of your style you cannot afford to lose as files on your own drive: the model, the add-on, the workflow and the seed. Use hosted models for what they are good at, which is speed and polish, and save the outputs you love, because the model behind them may not be there next year.

The counter-signal: owning files means owning the maintenance. A kernel module that touches your graphics card, an add-on on version 6 in four days, and a manga model with a purple-tint habit all ask you to test before you trust. And speed on old hardware does not replace memory; the edit add-on still wants 20 GB.

Source access notes: Thin primary-lab window (Saturday to Monday morning). Checked with nothing new dated October 24 to 26: OpenAI news (page failed to load; deprecations page used instead), Google blog (no dates shown, no Veo/Imagen/Lyria/Flow items), Adobe blog (latest October 19), BFL (October 15), Stability (August 25), ElevenLabs (September 30), Suno (October 13), Luma (October 23 Kling 4.0 posts, already covered), fal blog (October 20), Runway news (no dates rendered), Replicate explore (no dates; Qwen-Image-3 listed there dates to July/August per OrcaRouter and AlphaSignal, so omitted), PetaPixel and CDM (no AI items October 24 to 26), Hugging Face papers (one non-creative paper). HN Algolia returned nothing newer than October 12, so it was not used. Seedance 2.5 4K is sourced to TechNode only. GitHub API star counts were stale (gpu-unlock showed 866 against 8k on shields.io); shields.io totals are used, and Trendshift figures are shown only as list placement. Hugging Face dates are API createdAt; the MOSS-Image-Edit-HQ file list required a cache-busted API call, since an uncached call returned a two-day-old revision. Midjourney, Civitai and blog.comfy.org not fetched. Adversarial fact-check ran (Sonnet subagent) and caught four issues, all fixed: the lead named a 3090 instead of the tested 3090 Ti; the FLUX.1 row was an FP8 file and one H3 INT8 file gained 31.2 percent, so the "8-bit barely moved" framing was softened; the 4090 fork's gains were restated as the README's matrix-maths ranges; and the gpt-image-1 replacement sentence now names OpenAI's and Krea's actual replacements. It could not verify Trendshift placements (taken from this run's own Trendshift fetch) or the 4090 fork's creation date (taken from this run's GitHub API call, created October 23).