Creative AI Briefing October 13, 2026 morning edition 11 min read

Creative agents are starting to hand back editable working files, CAD solids and drawings and song arrangements, instead of flat renders, so the creator's job moves from rerolling a picture to correcting a file.

SolidCraft CADCraft Blokkily LTX-2.5 Unsloth Midjourney Thinking Mode Doory design-tools 3d-gen creative-workflows video-gen music-gen local-creative-ai

Creative AI Briefing: Tuesday, October 13, 2026

You can now describe a mounting bracket to an AI agent, watch it build the part step by step in a free CAD program, change one dimension yourself, and hand a machinist a STEP file they can open. That is the shape of today: three new open tools (a parametric CAD app, an AutoCAD-style drafting app and a pattern-based music workstation) all ship a control channel an agent can drive, and all three save to the formats professionals already trade (STEP, DXF and DWG, MIDI and WAV) rather than a flat picture. Even Midjourney's new test mode moves the same direction, checking its own image against your prompt before it hands it back. No major lab shipped a creative product in the sources checked since Monday. The useful work came from people building tools an agent can operate.

Image

Midjourney is testing a "Thinking Mode" that checks its own picture before you see it. On the alpha site since October 7, it appears as a "Rerun (Thinking)" option on existing V8.2 jobs and edits. According to the changelog as mirrored by Releasebot, the model looks for missed prompt details (wrong objects, layout, anatomy, text) and regenerates. A community summary of Midjourney's office hours says developers estimate it fixes roughly 60 to 80 percent of prompt-related issues; that number is an unpublished internal estimate, and AlphaSignal notes no benchmark backs it. No extra cost or speed figure has been disclosed. Secondary sourcing only: midjourney.com/updates returned no readable content. The same changelog adds shared folders, where collaborators can generate into a folder you own, which matters more than it sounds for a small studio sharing one mood.

Doory is a free, self-hostable design editor in the Canva mould. shaongitbd/doory (created October 10, 917 stars on shields.io) makes posters, social posts, slides, invitations and thumbnails, with layers, masks, curved text, 27 entry and exit animations, a music-loop library, live co-editing, and export to PNG, SVG, print-ready PDF with bleed, MP4 and GIF. Background removal runs in your browser. The licence (Attribution Assurance) lets you use it commercially as long as the app shows the author's credit line. Setup is Docker Compose on your own server. The README states no version or stability level.

Video

Unsloth published 20 compressed copies of LTX-2.5, the open video-with-sound model, plus two full-size references, with a measured quality score on every one. unsloth/LTX-2.5-GGUF was created October 13 at 09:07 UTC. Shrunk community versions of LTX-2.5 have existed since mid-August. What is new is the table. Files run from 8.46 GB to 42 GB, and each is scored against the full-size original on the same prompt and seed. The useful surprise: the 12 GB "UD-Q3_K_XL" file scores better than the larger 13.58 GB "Q4_K_M" on both measures, for both the fast and full versions, so bigger is not automatically better. The honest catch sits in the same card: compiling the full-size model versus not compiling it already differs by 0.265 on that score, so for files scoring above about 0.27 the measured loss is hard to separate from noise. Audio was not measured. Licence: free for commercial work under $10M annual revenue; this is not an official Lightricks release.

GenCine lets you plan a camera move and an object's path in 3D from a single still. From Johns Hopkins (project page, code), with Apache 2.0 weights at zonszer/Gencine_ckpts created October 13. The image becomes a coloured point cloud, you move selected regions and a camera through it in a tool such as Blender, and a video model renders the plan. It is research, not a tool you can use this week: the released code only reruns the shipped examples (21 frames at 832x480 on an 80 GB A100), and "new examples need precomputed inputs; preprocessing and training are outside this package." The authors say large viewpoint changes tear, and fluids and fracture are out of scope.

Multics generates one shared game world seen by several players at once. TencentARC/Multics (created October 12, Apache 2.0, built on Wan 2.2 and 2.1) is a four-player checkpoint the card says also handles 6, 8 and 10 players without retraining. Research only; no hardware figures. Project page.

Audio and music

Blokkily is an open music workstation an agent can play. andrewfader/blokkily, shared on Hacker News October 12, lets you edit one pattern as a tracker, piano roll or step grid, arrange across tracks, load CLAP and VST3 instruments, mix with sends and sidechains, and export WAV or MIDI files. Its MCP bridge (the plug that lets an assistant such as Claude operate a running app) exposes tools to read the session, write notes, set instruments and parameters and take a screenshot, and every agent edit goes through the same undo history as your own. The README links a video of an agent building a chiptune. It is "experimental": no comping, no MIDI out to hardware, the Mac build is Apple Silicon only and not notarized, and the README states no licence for Blokkily itself while noting that builds including its Rubber Band time-stretching library fall under the GPL. 36 stars.

LAION posted an interim 1.1 of its open voice model. laion/Humaneness-Voice-Large-1.1-Alpha (created October 13, CC BY 4.0) is a checkpoint at step 2,377 of 3,055, still training. It raises the voice-reference clip limit to 15 seconds (from about 3) and adds experimental French, Dutch and Spanish to English and German. The card itself says "no quality improvement is asserted before evaluation completes," and testing covers GH200 data-center chips only. Wait for the finished run.

Inflect v3 previews a voice model small enough for a laptop processor. owensong/Inflect-Mini-v3-Preview (created October 12, Apache 2.0) has about 27.6 million parameters and runs on CPU, with six fixed English voices and no cloning. The author warns of buzz and pacing defects, and you must spell out numbers.

Open and local

The local story today is software, not models. Two Rust rebuilds of professional design programs landed in the same week, both with an agent control channel. SolidCraft is a parametric 3D modeller: sketches with constraints, extrude, revolve, fillet, shell, threads, sheet metal and assemblies, in a feature timeline so changing a dimension rebuilds the part. It runs as a desktop app or in your browser, exports STEP, IGES, 3MF, STL and OBJ, and its own checks find 75 of 76 public reference parts match their analytic or FreeCAD-computed volume, area and centre of mass. The author calls it "early but usable for simple parts." CADCraft, from the ArtCraft team behind PhotoCraft, rebuilds AutoCAD-style 2D drafting. Its roadmap is unusually blunt: 50 percent of menu items work, about 38 percent "ready for real work," 3D at 2 percent, and a DWG it saves "drops everything it doesn't model and has never been checked in AutoCAD."

bherbruck/solidcraft: free parametric CAD for prop makers, product designers and anyone who 3D prints, with an MCP server. Why now: Show HN October 13; created October 7; 1.3k stars on shields.io (repo).
storytold/cadcraft: open AutoCAD-style drafting with an agent command line. Why now: created October 7; 3.7k stars on shields.io; on Trendshift's daily list (repo).
andrewfader/blokkily: a pattern-first music workstation an agent can drive. Why now: Show HN October 12; 36 stars (repo).
shaongitbd/doory: self-hosted Canva-style editor with animation and print export. Why now: created October 10; 917 stars (repo).
unsloth/LTX-2.5-GGUF: compressed LTX-2.5 files from 8.46 GB, each with a published quality score. Why now: created October 13 (repo).
vanillaer/GenCine: research code for camera-and-object plans from one image. Why now: weights posted October 13 (repo).

Creative workflows

1. Have an agent draft a parametric part you can still edit: SolidCraft (solidcraft-cli mcp)

The steps. Install SolidCraft (Windows, Mac, Linux builds) or open the web app to learn the interface first. Start the desktop app with solidcraft --control PORT, then register the agent bridge: for Claude Code the README gives claude mcp add solidcraft -- solidcraft-cli mcp, and says adding --connect 127.0.0.1:PORT makes it drive the running app so you can watch (combining the two in one line is our inference, untested). Ask for a simple part with named dimensions, such as a wall bracket with a hole spacing and thickness you specify. Have the agent call measure and screenshot to check its work, then export to STEP for a machinist or 3MF for a printer.

How it works. Every action is a command (the README's example is {"command": "solid.extrude", "params": {"distance": 20}}), so the agent builds the same feature timeline you would, in order. Change the thickness yourself afterwards and the part rebuilds.

Why it is good. The result is a solid model with history, not a mesh you cannot edit. You watch it happen in the window and can undo any step.

Where it breaks. "Simple parts" is the author's own ceiling. Imported STL, OBJ and 3MF files come in as meshes that solid features cannot touch. Downloads are unsigned, so expect operating-system warnings. One reference part is a tracked bug. And no example agent prompt is published, so expect to iterate on how you describe parts.

2. Pick the right compressed LTX-2.5 file for your card: unsloth/LTX-2.5-GGUF (distilled/ltx-2.5-22b-distilled-UD-Q3_K_XL.gguf)

The steps. In ComfyUI, install the ComfyUI-GGUF custom node. Download one file from the repo plus the separate text encoder, decoders and upscaler from Lightricks/LTX-2.5 (the card gives the commands). Open the built-in Text to Video (LTX-2.5) template and replace the Load Diffusion Model node with Unet Loader (GGUF). Unsloth recommends UD-Q4_K_XL (14 GB) for most cards and UD-Q2_K_XL (9.5 GB) for small ones. Unsloth Studio users can pick the file on its Video page instead.

How it works. Each file stores the same model at lower precision. Unsloth rendered the same prompt and seed with each and scored the frames against the full-size model.

Why it is good. You choose with numbers instead of forum guesses. For the fast ("distilled") version, the 12 GB UD-Q3_K_XL scored 0.310 (lower is closer) against 0.352 for the 13.58 GB Q4_K_M.

Where it breaks. The card's own baseline note says compiling versus not compiling the full model already differs by 0.265, so gaps between files scoring above that are hard to call. Sound quality was not scored. The files score video frames from one prompt only. Text encoder and decoders add their own memory on top of the file size.

3. Let an agent write a first pass in a music session: Blokkily (docs/mcp.md)

The steps. Install the release for your platform and check it against SHA256SUMS. Register the bridge: claude mcp add blokkily -- blokkily --mcp. The first tool call opens Blokkily if it is not running. Ask for a pattern, listen, and undo what you do not like.

How it works. The agent reads the session, writes notes, sets instruments and parameters and starts playback, through the same undoable paths as the on-screen controls.

Why it is good. You keep a real arrangement with real instruments, not a rendered stereo file you cannot pull apart.

Where it breaks. Export always renders the arrangement, so record any scene-launcher performance into it first. No MIDI out to hardware, no comping, and only the user running Blokkily can connect.

Worth testing

SolidCraft web app: free, no install. Tradeoff: built for simple parts; check exported STEP files in the program your manufacturer uses.
Midjourney Thinking rerun on alpha.midjourney.com: try it on a prompt that keeps missing an object or lettering. Tradeoff: alpha only, no disclosed cost or wait time, and the improvement figure is unverified.
Doory on a spare server for a team that wants Canva-style templates without accounts. Tradeoff: you run and update it yourself, and the credit line must stay visible.
LTX-2.5 UD-Q3_K_XL on a 12 to 16 GB card. Tradeoff: 12 GB is file size only; companion parts need memory too, and audio quality is unmeasured.

What actually matters from today's signal

The deliverable is changing. For three years "AI for creators" meant a model that returns finished pixels, and every fix meant rolling the dice again. SolidCraft, CADCraft and Blokkily return the thing a professional actually trades in: a part with a feature history, a drawing on layers, an arrangement with instruments. When the agent gets a dimension wrong, you change one number. That is a cheaper failure than a reroll, and it keeps your judgment in the file instead of in a prompt.

Midjourney's Thinking test and Unsloth's quality table point the same way from the generator side. Both put a check between the machine and you. Neither check is proven yet: Midjourney's figure is an internal estimate passed along secondhand, and Unsloth's own note shows its mid-size scores sit near a noise floor. Treat both as useful habits, not guarantees.

The risk is maturity. CADCraft's roadmap admits its DWG saves are unchecked in AutoCAD, SolidCraft limits itself to simple parts, and Blokkily calls itself experimental. Use them where a wrong file costs you an afternoon, not a client. And note how much of this fortnight's open work comes from one team: the CADCraft README lists PhotoCraft, LightCraft, FilmCraft and more as ArtCraft tools. In my view that is more software than anyone outside the team has had time to test.

Source access notes: Checked with no new creative launches since Monday: OpenAI news (newest Oct 8, none creative), Adobe blog (newest Oct 2), ElevenLabs blog (newest Sep 30), Black Forest Labs (newest Sep 23), Luma (Oct 8 Claude Motion item already covered), Suno (newest Oct 5), fal and Replicate blogs, Runway changelog (newest Oct 2), DeepMind blog (no dated creative item), cdm.link, MusicRadar, Creative Bloq, PetaPixel. Stability AI news showed no dates. midjourney.com/updates returned only metadata; Midjourney item sourced to Releasebot's changelog mirror and AlphaSignal (secondary). ComfyUI GitHub releases blocked by robots; blog.comfy.org returned no posts; api.github.com returned errors, so repo creation dates come from shields.io (relative dates converted: "last wednesday" = Oct 7, "last saturday" = Oct 10) and star totals from shields.io with cache busting. HN listing via Algolia API with a time filter; an unfiltered query returned stale October 2 items and was discarded. All Hugging Face dates are createdAt from the models API. Meta's Muse Wonder (Oct 12) is a text agent model and left to the general briefing. YuE2 mirror (FenomAI/YuE2-3B, created Oct 13) skipped: the official m-a-p release dates to September 9. New models and Writing sections omitted: nothing scored a 5, and nothing creator-relevant in writing. Adversarial fact-check ran (Sonnet subagent) against primary sources. It caught and we fixed: an unverified combined SolidCraft command (now labelled as inference), an overstated "project file" claim for CADCraft (it saves DXF/DWG), a "second-day" mislabel on the Unsloth files, an over-broad noise-floor caveat (it applies only to files scoring above about 0.27), Blokkily's Mac build being ad hoc signed rather than unsigned, an unsupported plugin-editor claim, the wording of the Midjourney estimate, and an unsourced "labs were still" line.