Creative AI Briefing October 18, 2026 morning edition 11 min read

This week's new image and video models take direction by placement, with boxes, numbered references and keyframes, so the sketch or storyboard a creator already makes is turning into the prompt itself.

MOSS-Image OpenMOSS Krea 3 Kling 4.0 SimpleTuner ComfyUI image-gen video-gen open-weights ai-editing prompt-craft lora-finetuning

Creative AI Briefing: Sunday, October 18, 2026

You can now hand an open image model a list of boxes, each holding the exact words to print inside it, and get back a poster with the headline sitting where you drew it, under a licence that lets you sell the result. That model is MOSS-Image, and it arrived the same week Krea shipped an editor that takes up to eight references you point at by number, and Kling officially launched a video model that takes up to ten keyframes. Three companies, one bet: the next prompt is a layout. If you already sketch thumbnails, mark up comps or board your shots, that habit just became the most valuable input you own.

New models

MOSS-Image is an open poster-and-layout model you can use commercially. OpenMOSS announced it on October 16 (the Hugging Face repos were created October 11 and refreshed on October 16; the demo Space was created October 14). It comes in two versions: MOSS-Image-SFT, the one most people should run, and MOSS-Image-Base, kept rawer as a starting point for training your own add-ons. What you can make: posters, infographics and title cards with legible text, from a description, a tag list, or a JSON layout file that says where each piece of text and each object goes. The licence is the headline: the model card says the weights and bundled code are under Apache 2.0, which means client work with no revenue ceiling (FLUX 3 Image Klein, by contrast, is free only under US$5 million in annual revenue). Where it runs: NVIDIA cards with CUDA. OpenMOSS publishes no memory figure, but ComfyUI's repack gives you the sizes that decide it: the picture model is 10.2 GB at full precision or 5.4 GB shrunk, and the text reader that interprets your prompt (a 9-billion-parameter Qwen model) comes only in shrunken versions there, at 9.8 GB or 7.3 GB. ComfyUI added core support on release day (PR #17095, merged October 16), after the latest release tag, so update before you look for it. Free to try in the hosted demo. The catch: the default is 50 sampling passes, so expect to wait, not iterate. Layout control is labelled experimental in the user guide. English has the most training data; Chinese and other languages are supported but less reliable. OpenMOSS publishes no benchmarks and no limitations section.

Krea 3 edits a picture from up to eight numbered references for three to six cents. Krea opened the beta on October 15 (product page, API docs). You give it a photo and an instruction, point at people, products, places or styles by their upload number, and it returns a 1K or 2K edit. A regional mode redraws only what needs to change, so a product label stays put. Pricing from Krea's API docs: $0.03 per 1K image, $0.06 per 2K, plus $0.02 for each 1K reference, $0.07 for each 2K reference, and $0.02 more for the slower "high" thinking mode. That means a 2K edit with four 2K references costs $0.34, not six cents. The under-8-seconds speed claim is Krea's own, footnoted on its page as illustrative playback, beside comparison clips it calls "pre-generated outputs, not a live benchmark." Available in the Krea app, as a ComfyUI Partner Node paid in Comfy credits, and on Replicate. The catch: some launch coverage (TestingCatalog, labelled sponsored) says you can fine-tune Krea 3 with your own add-ons today. The API docs say otherwise: sending add-ons returns an error reading "LoRAs are coming soon to the API," and style references are also unavailable. The API itself is early access for beta workspaces. Krea also warns that whole-image edits (restyle, relight, reframe) re-render everything, so artifacts build up across repeated edits.

Kling 4.0 is out, and it directs 30-second shots from keyframes. Kuaishou's October 16 press release says the model "officially launched on October 15." What you can make: a single continuous take up to 30 seconds, steered by up to 10 keyframes (stills the shot must pass through) and up to 15 references in total, of which up to 10 can be images, 5 videos and 7 subjects. You can also feed it a "white-model" (untextured 3D) or depth-map video to set the motion and blocking, and reference a voice. Prompts run to 8,000 tokens. It outputs up to 10-bit HDR, and Kling says 4K, 1080p and 21:9 widescreen are coming, with extensions to two minutes "soon." It runs in Kling's app and through Kling's MCP connector and command-line tool; the developer API is "coming soon." The catch: the press release names no prices. The only rates I found are secondary: a kie.ai post (from a company that resells API access) says 15 credits per second at 1080p and 10 at 720p, so a 30-second 1080p take is 450 credits before rerolls, and that HDR was not yet live on October 16. Treat both as unconfirmed. Cloud only.

Image

MOSS-Image got its own second day inside 48 hours. By this morning, accounts other than OpenMOSS had published a stack of supporting files. lodestones/MOSS-Image-Base-8-steps-dmd (created October 18, 01:27 UTC) is a work-in-progress fast version that cuts the 50 passes to 4 to 8, under Apache 2.0. Its card warns that converting it to the common half-precision format "discards most of the distillation," so it needs full precision or the unmerged add-on files. leejet, the developer behind stable-diffusion.cpp, posted shrunken GGUF files of the SFT version on October 18. Style add-ons have started appearing too, such as a gouache add-on (October 18). None publish quality measurements yet.

Training support landed within a day. SimpleTuner merged MOSS-Image add-on training on October 17 (work began the evening before), then edit and reference training the same evening. Memory tiers run from 24 GB up to 144 GB, with 48 GB as the default example.

Open and local

The local story: the week's most useful open image release is the one whose licence lets you bill a client and whose prompt format matches how designers already think in boxes. The fast versions are a day old.

OpenMOSS/MOSS-Image: the code, Gradio app, ComfyUI plugin and user guide for the open layout-aware image model. Why now: released October 16, 523 stars (repo).
bghira/SimpleTuner: a trainer for making your own image add-ons, now with MOSS-Image style, edit and reference training. Why now: MOSS support merged October 17 (quickstart).
Comfy-Org/ComfyUI: the node-based image and video workbench, now with native MOSS-Image nodes. Why now: support merged October 16, ahead of the next tagged release (PR).
obsidianmd/pele: an MIT library from the Obsidian team that turns Mermaid diagram text into SVG you can restyle with CSS colours and fonts; colour changes apply without re-rendering. Why now: about 425 points on Hacker News on October 17 (site).
KnifeLemon/Filee: a free, offline drag-and-drop converter for what its README now counts as 244 formats, including camera RAW, HEIC, video and audio (video and audio use an optional FFmpeg download). Why now: on today's Trendshift daily list, MIT licence (repo).

Creative workflows

1. Lay out a poster as boxes with MOSS-Image: USERGUIDE.md

The steps. Decide your canvas in multiples of 16 (the guide recommends 1, 1.5 or 2 megapixels total; 1024 by 1024 is a safe start). Write a JSON file with a summary block (describing the whole scene and canvas) and an items list. Each item gets a label (text, character or object), a caption, and a bbox; text items also carry the exact words to render. Set "bbox_format": "yxyx" and "bbox_coord_space": "abs_centralized". Pass the whole JSON as the prompt, in the Space, the Gradio app, or ComfyUI's bundled moss-image.json workflow. Defaults: 50 passes, guidance 3, shift 5.

How it works. Coordinates are pixels measured from the centre of the canvas, top before left: [top, left, bottom, right]. On a 1024 canvas the edges sit at minus 512 and plus 512, so the guide's example box [-416, -384, -288, 384] is a 128-pixel-tall band near the top, 768 pixels wide. If you measured your sketch from the top-left corner, as most design apps do, subtract half the width from x and half the height from y.

Why it is good. You place the headline, price and logo where the comp says they go, instead of rerolling until the model guesses.

Where it breaks. Layout control is experimental by the guide's own label. Automatic line breaks "may vary"; add \n yourself for exact breaks. English is the safest language for rendered text. The ComfyUI plugin notes that the same seed and settings will not reproduce the standalone app's image. 50 passes make each try slow.

2. Edit a product shot with numbered references in Krea 3: Comfy's Krea 3 guide and Krea's API notes

The steps. Load your starting photo and your references (Krea's API notes recommend four images or fewer for full detail). In the instruction, say what each numbered reference is for ("use the green bottle from the second image and the window light from the third"; Krea's docs say to refer to each image "by what it shows"), and name what must stay, not only what changes. Turn on regional editing for local changes. Keep references around 1 megapixel; Krea says not to upscale them first. If a chain drifts, branch again from the strongest earlier result.

How it works. References are read in upload order, so numbering is how you point. Regional mode redraws only the parts the instruction touches and keeps the rest.

Why it is good. Explicit pointing replaces guesswork, at cents per step.

Where it breaks. References cost more than the output at 2K ($0.07 each against a $0.06 base). With five or more references, each is shrunk to 1536 pixels on the long side, so fine label text can soften; Krea recommends four or fewer for full detail. Whole-image edits re-render everything and stack artifacts. Seeds do not guarantee repeat results. No custom add-ons yet.

3. Teach MOSS-Image an edit with paired folders in SimpleTuner: MOSS_IMAGE.md quickstart

The steps. Put your "after" images in datasets/moss-edits and the matching "before" images in datasets/moss-references, with identical filenames including extensions. Put the instruction for each pair in a .txt file next to the "after" image. Start from the moss_image-sft-edit.peft-lora example config. The recipe runs 4,000 training steps and saves a test image every 250.

How it works. The trainer learns the change between each pair under your instruction, at a 512-by-512 equivalent area, keeping the model's text reader frozen.

Why it is good. A house edit you repeat (backdrop swap, packaging mockup) becomes a reusable file on an Apache-licensed base.

Where it breaks. The default examples assume a 48 GB card; the 24 GB tier runs one image per batch. References add memory. The automatic JSON captioning needs an 80 GB card. The support is one day old and the guide gives no training-time figure.

Worth testing

MOSS-Image Space: paste a JSON layout from the user guide and change one box at a time. Tradeoff: shared free hardware with queues, and 50 passes per image.
Krea 3 in the Krea app: try a regional edit on a product photo with two numbered references. Tradeoff: reference fees add up fast at 2K, and repeated whole-image edits degrade.
ComfyUI's MOSS-Image nodes: update ComfyUI to the latest code and load the plugin's moss-image.json. Tradeoff: about 13 GB of model files with both shrunken versions, about 20 GB with the full-precision picture model, and support sits ahead of the tagged release.

What actually matters from today's signal

Description is losing to placement. For three years the skill was writing paragraphs a model would read the way you meant. This week the models read boxes, numbers and keyframes instead. The people who gain most already work visually: thumbnails, boards, grids.

The money side moved too. MOSS-Image under Apache 2.0 stands out against the licences this briefing has covered in recent weeks, which kept landing on research-only terms (Qwen-Image-2.1, InfoVis) or revenue ceilings (FLUX 3 Image Klein). Krea 3's prices look tiny until you count references, which cost more than the picture at 2K. Fewer, smaller references cut the bill sharply.

The counter-signal is honesty about readiness. MOSS-Image calls its layout control experimental and ships no benchmarks. Krea's own docs contradict sponsored coverage on custom add-ons. Kling announced features faster than it priced them. Placement is the right direction, not yet a finished tool. Test on a real brief before you promise a client anything built on it.

Source access notes: Adversarial fact-check pass ran (Sonnet subagent, about 60 claims checked). It caught: leejet's GGUFs cover the SFT version only, not both; Kling's term is "white-model" video; a workflow step wrongly attributed a three-reference cap to Comfy's post (now cited to Krea's four-or-fewer advice); the ComfyUI text-reader files are both shrunken versions, not full precision; the model-file total was miscalculated (now 13 GB shrunken, 20 GB with full-precision picture model); Filee's README now counts 244 formats; Pele's HN score drifted to about 425; and an unsupported licence superlative was replaced with a comparison to licences covered here. Checked with no new creative launches in the window: OpenAI news (newest October 15, education items), BFL blog (newest October 15, Klein, covered), ElevenLabs blog (newest September 30), Suno blog (newest October 13), Stability news (newest August 25), Luma news (October 14 comparison guides), fal blog (newest September 17), Runway changelog (newest October 2), Hugging Face blog (October 15 to 16 items are text models plus the Klein Diffusers post, covered), Higgsfield changelog (October 16 beginner course only). Meta's Muse Spark Open blog post (October 15) failed to load, so its creative relevance is unverified and it is omitted. Bucket D (PetaPixel, CDM) had no creative-AI launches October 16 to 18. Kling pricing and HDR status come only from kie.ai, a reseller, and are flagged inline. Krea's open-beta date rests on Comfy's October 15 post and Replicate's October 15 version timestamp (Krea's own pages give no date); the "default app model since October 16" detail appears only in secondary coverage and is omitted. Blocked or unusable: x.com (robots), GitHub REST API for the MOSS-Image repo and ComfyUI releases (fetch errors; PR status read from the API successfully), shields.io for three repos returned "invalid" (star counts omitted for those), Midjourney (no new items in search), civitai and blog.comfy.org listing (skipped; the Krea post was read directly). Hugging Face dates are API createdAt; the MOSS-Image repos' lastModified of October 16 marks the public release refresh, not creation. No Video, Audio and music, or Writing sections today: Kling 4.0 sits under New models, and nothing verifiable shipped in audio or writing.