Creative AI Briefing October 24, 2026 morning edition 12 min read

The new tools plan or pause before they finish, so creators can give notes in the middle of a job instead of rerolling the whole thing.

Canva Craft Hy-Nomas MOSS-Image Alpha LoRA Phonon Krea 3 3D Model Anything image-gen design-tools video-gen voice-clone local-creative-ai comfyui

Creative AI Briefing: Saturday, October 24, 2026

You can open Canva's Dream Lab today, even on a free account, type a poster brief and get back up to four layouts with the headline you asked for, then (as a Premium step) split the winner into editable text and shapes with Magic Layers. Canva's new Craft model gets there by writing a layout plan before it paints a pixel. That is the week's pattern. Tencent's new open world model builds video one segment at a time and takes a fresh instruction between segments, and Krea's edit API now takes up to 24 boxes, each with its own note. The model is starting to work like a junior who checks in, which means your notes land mid-job instead of after a full reroll.

New models

Canva Craft is Canva's own image model, trained on templates instead of photographs. Canva announced it on October 22 (Canva newsroom; developer forum post, dated October 22). What you can make: posters, menus, infographics, slides, social posts and portraits at up to 2K, with on-image text in 55 languages. A planning model first writes a composition plan (elements, placement, scale, hierarchy, text), then a second model renders it. Training drew on "millions of human-made templates" from Canva's library, then fine-tuning guided by Canva's designers; Canva says "When automated evaluations and human judgments differed, we sided with the designer." Canva says it debuted fourth overall on DesignArena, first for People & Portraits and Landscapes & Nature; BetaNews adds that three OpenAI models rank higher and its rating slipped from 1340 to 1330 by October 23. Where it runs: Dream Lab for every Canva user, and the preview Connect API as craft_v1_2k, alongside two partner models Canva may pick if you leave the model unset. Price: it draws on your Canva AI allowance; Canva lists no per-image figure, and the forum post says to "Expect roughly 50 to 60 seconds end to end." The catch: a hands-on review by Gaurav Tiwari (October 24) found it adds copy nobody asked for (invented opening hours, a made-up subtitle), ignores your Brand Kit in Dream Lab, and re-renders the whole image on one-change edits. The API reference lists six aspect ratios against the "more than 30" in the announcement. Canva itself says fine print and technical drawings still need work. Commercial use of output is allowed under Canva's terms, but the same review notes Canva's AI indemnity covers eligible Enterprise customers only (secondary).

Hy-Nomas lets you steer an open video world one segment at a time. Tencent's Hunyuan team posted the weights on October 21 (created per the Hugging Face API; GitHub repo the same day), with a Diffusers version created at 05:45 UTC today. What you can make: a 720×1280, 24 fps scene that keeps going, where between segments you write a new instruction (add an object, change the style), describe an action with start and end times, or feed a camera path. The GitHub Diffusers example renders landscape 1280×704 instead. A memory of earlier segments is meant to keep objects and lighting consistent. There is an official web demo. Where it runs: the standard setup uses eight GPUs; a low-memory setup runs on one RTX 4090 with compressed 8-bit weights and CPU offload, and the README warns "CPU offload requires enough system RAM for the models" without giving a number. A community ComfyUI repack (October 23) claims 12 GB cards and up. License, plainly: Tencent labels the release Apache 2.0, but the video decoder is a fine-tune of NVIDIA's PixelUMM, and both the ComfyUI repack's card and OrcaRouter say PixelUMM weights carry a noncommercial research license. Treat client use as unresolved until Tencent clarifies. The honest catch: no speed figures, no benchmarks, technical report "Coming Soon," 73 stars (shields.io).

Image

Transparent PNGs straight out of MOSS-Image. paralaif/MOSS-Image-Alpha-LoRA (created October 24) is a small add-on that makes the open poster model output a real alpha channel, so a game item, sticker or logo arrives cut out, glass and wings partly see-through. End the prompt with ", isolated on a transparent background," load the included moss-alpha-lora-t2i.json workflow and the ComfyUI-RGBA-VAE node. It was trained on 39 images in about 40 minutes on a 16 GB RTX 5060 Ti. The catch: the author calls it a proof of concept, fire and smoke are weakest, dark clothing can go translucent, and because 8 of the 39 training images came from Qwen Image 2.1, it is released for non-commercial research and evaluation only; the RGBA decoder it needs carries the FLUX Non-Commercial License on its own card.

Qwen-Image-2.1 in two steps. linoyts/Qwen-Image-2.1-DMD2-2steps (created October 23) cuts the model from 20 passes to 2 for both generation and edits, described as "suitable for real-time and interactive pipelines." Free demo: the Qwen Image 2.1 Realtime Space. The catch: it needs an unreleased Diffusers branch, the card shows comparison images but no measurements, and the Qwen Research License bars commercial use.

Krea 3 boxes now work in its top mode. Per Krea's changelog, as of today the API accepts regions with thinking_mode: "ultra", so you no longer have to drop to high to point at part of a reference image with a box.

OpenAI switched off gpt-image-1 yesterday. OpenAI's deprecations page lists the October 23 shutdown, announced in April, with gpt-image-2.5-sunburst or gpt-image-2.5-flare as replacements. Krea's changelog confirms its old route now returns "410 Gone." If a plugin, Zapier step or script of yours stopped making images overnight, this is why. The gpt-image-1.5 and gpt-image-1-mini models follow on December 1.

Video

Beyond Hy-Nomas, Luma added Kling 4.0 to its boards on October 23, so you can judge Kling's fast-action shots beside the rest of a project.

Audio and music

Phonon puts a small, commercially usable English voice in a browser tab. Gradium's phonon-31m-en (created October 19) is a 64.5 MB text-to-speech file the card says is "small enough to run in a browser tab, on a phone, or on a laptop CPU," with about 100 stock voices, Apache 2.0 for weights and voices. It shows up today because llama.cpp and audio.cpp conversions landed on October 23, the second-day work that pulls it into more local tools. Free demo: the Phonon Space runs in your browser on CPU or WebGPU. The catch: English only (the French, German, Spanish and Portuguese siblings are CC BY-NC, non-commercial), no voice cloning in the open model, no speed figures, and the card's "first on all three" ranking is Gradium's own and its own table shows Kokoro ahead on one error measure.

LAION's expressive voice model reached 1.5. Humaneness Voice Large 1.5 Alpha (created October 24) folds in yesterday's 1.4 tuning plus 2,000 more updates. On LAION's own automatic scores, word errors fell from 2.743% to 1.942% and "genuineness" rose from 4.299 to 4.370 against 1.4; LAION calls these "learned proxy scores, not human MOS." Same five languages, same CC BY 4.0, and still a CUDA GPU of unstated size. Do not stack it on 1.4.

Krea opened its Audio tool to everyone. Per the changelog, every signed-in user can use Audio as of today, Train gained an Audio option for cloning a voice, and the tool shows a cost estimate before you generate.

Open and local

The local story today is breadth rather than one big model: an open world model with a single-card mode, a cut-out add-on for a week-old poster model, a TTS small enough for a phone, and a room-capture kit that turns an iPhone scan into an editable 3D scene through an AI coding agent.

Tencent-Hunyuan/Hy-Nomas: steer a generated video world with text and camera paths between segments. Repo created October 21, 73 stars (shields.io); single-4090 mode, licence unclear for commercial work (repo).
wesbos/3d-model-anything: scan a room with a LiDAR iPhone, then let an agent skill (scan-to-threejs) rebuild it as an editable Three.js scene you can walk through. Repo created October 21, 1.3k stars (shields.io); the iPhone app has no App Store release, so you build it in Xcode (repo).
paralaif/MOSS-Image-Alpha-LoRA: transparent PNGs from MOSS-Image with a ComfyUI workflow included. Created October 24, non-commercial (card).
gradium-ai/xn-ptts: the runtime behind Phonon for CLI, Python, browser, iOS and Android. 279 stars (shields.io); model created October 19 (repo).
basketikun/infinite-canvas: an MIT infinite canvas for image and video generation against any OpenAI-compatible API, with a Claude Code or Codex agent wired in over MCP. On today's Trendshift list; 8.9k stars (shields.io) (repo).

Creative workflows

1. A Craft poster with type you control. Source: The Rundown's Craft tutorial (updated October 23) plus Canva's newsroom post.

The steps. Open Dream Lab from Canva's home page (Craft does not live inside the editor) and confirm the model reads Canva Craft. Pick 3:4. Write the subject, style and a limited palette, give the headline once, and ask for no other text or logos. Reserve a clear band at the bottom for details. Read the headline letter by letter at full and small size. If it is wrong, simplify once, or regenerate text-free. Then place the art in a 1080×1350 design and set date, venue and call to action as real text boxes.

How it works. Craft plans the layout first, so naming the text and the empty band tells the planner where things go.

Why it is good. One prompt returns up to four different layouts, a usable first concept round, and the type you care about stays editable.

Where it breaks. Craft invents extra copy, Magic Layers is a Premium step that may not separate a detailed illustration cleanly and spends the same AI allowance, and generated words do not always become selectable text.

2. Cut-out game items and stickers from MOSS-Image. Source: MOSS-Image-Alpha-LoRA card, workflow comfyui/moss-alpha-lora-t2i.json.

The steps. Install the ComfyUI-RGBA-VAE custom node. Download moss_alpha_v1.safetensors into models/loras/, the MOSS-Image int8 model and Qwen3.5 9B text encoder from Comfy-Org/MOSS-Image, and the RGBA VAE from trmz/flux2-klein-alpha renamed to flux2_klein_alpha_rgba_vae.safetensors. Load the workflow, keep strength at 1.0, end every prompt with ", isolated on a transparent background," and describe the subject, never a setting. With the optional 8-pass add-on, use 8 steps at CFG 1; without it, 28 steps at CFG 4. Save with the "RGBA PNG" node.

How it works. The add-on teaches MOSS to produce images that carry transparency information; a four-channel decoder turns that into a PNG with an alpha channel.

Why it is good. No keying, no background-removal pass, and soft partial transparency on glass and wings that a cut-out tool would flatten.

Where it breaks. Halos on unusual subjects, fire and smoke, see-through dark fabric, a bias toward centred full-body subjects, and a non-commercial licence.

3. Scan a room, let an agent rebuild it. Source: 3d-model-anything README.

The steps. Build the iPhone app in Xcode and capture a room on a LiDAR iPhone or iPad. Export the session and unzip it to data/captures-in/<room>/. Ask a coding agent that reads local skills to run scan-to-threejs on it. Run npm ci and npm run dev, then open the viewer at localhost:5188.

How it works. The agent measures the scan, models the room as code, then renders views from the original photo positions and compares them, repeating until they match.

Why it is good. You get an editable scene, walls and furniture as objects, instead of a lumpy scanned mesh.

Where it breaks. The README calls it "an iterative reconstruction workflow, not a one-click scan-to-perfect-model converter," it covers rooms and interiors only, the README mentions no USDZ or Blender export, and the agent time is yours to pay for.

Worth testing

Canva Craft in Dream Lab: free accounts can try it. Tradeoff: it spends a small AI allowance and invents copy you must proofread.
Hy-Nomas web demo: steer a world by text without a GPU. Tradeoff: no speed or quality numbers yet, and the licence is unclear for paid work.
Phonon in the browser: free English narration with no install. Tradeoff: stock voices only, no cloning.
Qwen Image 2.1 Realtime: near-instant edits for exploring ideas. Tradeoff: research licence, so nothing you sell.

What actually matters from today's signal

The reroll is losing its grip on the workflow. Craft writes a plan before it paints. Hy-Nomas stops between segments and asks where to go next. Krea lets you put a box on the part you mean. Each one moves the creator's input from the start of the job to the middle of it, which is how directing works with people. If you have been writing ever longer prompts to get everything right in one shot, the better habit now is shorter briefs and more notes.

Canva is the bigger commercial signal. A company sitting on millions of templates trained a model on them and put it in front of 250 million users inside plans they already pay for, or use free. For freelance designers, the floor for "decent social graphic" just dropped again. Your edge is the part Craft visibly fumbles: exact copy, brand consistency and the call on which of four layouts actually works.

The counter-signal is licensing. Hy-Nomas looks open until you read the decoder's lineage, the MOSS alpha add-on is non-commercial because of its training images, and the two-step Qwen model sits on a research licence. Free weights increasingly come with strings tied to whoever made the parts. Check before a client sees the shot.

Source access notes: Hugging Face dates are API createdAt (fetched through WebFetch; the session's shell could not reach huggingface.co). Hy-Nomas web listing showed only relative times, so dates come from the API and GitHub API. GitHub's HTML releases page stopped at ComfyUI v0.42.1 and the compare API returned stale results, so no new ComfyUI release is reported. PetaPixel's Craft story and Variety failed to load; Craft facts rest on Canva's newsroom, forum post and API reference, with BetaNews for the leaderboard numbers and Gaurav Tiwari's review for hands-on findings and plan limits. OpenAI's news page failed to load; the gpt-image-1 shutdown comes from OpenAI's deprecations page. Bucket A checked with nothing new in the window: ElevenLabs (latest September 30; the UMG deal is from September 10), Suno (October 15), Runway (Nano Banana 2 Lite on its API, October 21), fal blog (October 20, inference topic), Luma (Kling 4.0 integration plus comparison posts), Midjourney and Google (no dated items). PixVerse R2 surfaced but dates to September 28. Utopai's funding round and Block's Dessa voice-agent model were excluded as funding and non-creative. HN search returned nothing creative after October 22. Trendshift's second number per repo was unlabeled; star totals are shields.io. Civitai and blog.comfy.org not fetched. Adversarial fact-check ran (Sonnet subagent) and caught eight issues, all fixed: Magic Layers is a Premium step, not part of the free flow; the six API aspect ratios now cite the API reference (the forum post lists three); an invented quoted Hy-Nomas prompt removed; the RAM quote is from the GitHub README, not the card; a "bottom fifth" detail not in the tutorial removed; the "behind three OpenAI models" detail and the 1340 to 1330 slide attributed to BetaNews; the landscape 1280×704 Diffusers example noted beside the 720×1280 default; and "no USDZ or Blender exports" softened to the README not mentioning them. The 3d-model-anything star count is shields.io (1.3k); the GitHub API returned 493, likely a cached value.