Creative AI Briefing October 16, 2026 morning edition 11 min read

Two image models launched on the same day and both sell the same thing, a box you draw that changes one part of a picture and leaves the rest alone, so regional editing is now the default image job and the real choice is between running it yourself or paying per edit.

FLUX 3 Image Klein Krea 3 Black Forest Labs ComfyUI Concat VoiceStudio image-gen ai-editing open-weights local-creative-ai lora-finetuning

Creative AI Briefing: Friday, October 16, 2026

You can now draw a box around the mug in a product shot, type "make it matte black," and get the same photo back with only the mug changed, on your own graphics card and for free. That is FLUX 3 Image Klein, which Black Forest Labs released yesterday as open weights. Krea launched Krea 3 the same day with almost the same pitch: draw a box, describe the change, keep everything else. When two labs ship one idea on one day, the idea has become the job. Image models are being built for editing what you already have, and the decision in front of a creator is no longer which model draws best but whether to run the editor yourself or pay six cents a change.

New models

FLUX 3 Image Klein puts box-guided editing on a home graphics card. Released October 15 by Black Forest Labs, Klein is a 5.6 billion parameter model that makes images from text, edits existing images, and combines up to 10 reference pictures in one result. You can place elements with bounding boxes written into the prompt, and the API version copies unedited regions back exactly ("pixel restoration"). Two versions ship: a fast one that finishes in 16 steps, and a slower Base model meant for training your own add-ons. The Hugging Face repos were created October 13 (fast, Base), two days before the public post. Where it runs: BFL says the INT8 build "runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM," with its timings measured in ComfyUI, at roughly 4 seconds per image on an RTX 5090 and under 9 seconds on an RTX 4070 (BFL's own figures). The GitHub README names the RTX 3070 as an example 8 GB card for the ComfyUI int8 files. Without those compressed files, the model card reports 31 GB peak on a B200 data-center card. The honest catch: Klein expects a long, detailed scene description, and BFL ships a separate 9 billion parameter "prompt upsampler" to write one from your short request. The README says "Do not use the model without it," and Diffusers does not run it for you. That second model is the part that strains a small card. Licence: FLUX Kommunity License. Free commercial use for individuals and companies under US$5 million in annual revenue; above that you need a self-hosted licence or the API. The code and the upsampler are Apache 2.0. Price: BFL's post says API "Pricing starts at $0.009 per 1K image." Day-one support in ComfyUI (official templates) and Diffusers. The quality rankings in the post (first among open models, ahead of Ideogram V4.5 and FLUX.2 [pro]) are BFL's internal evaluations. Try it free: the Klein Space (created October 14, shared ZeroGPU hardware) or the BFL playground.

Krea 3 is the paid, hosted version of the same idea, priced for volume. Introduced October 15, Krea 3 is an editing model for product and character work: draw a box and describe the change, name an exact hex colour for a product, and keep a character consistent across a sequence of edits. Krea lists $0.06 per edit at 2K and $0.03 at 1K, with most edits under 8 seconds (Krea's own timing). Its post estimates 100,000 edits a month at $6,000, against $13,500 for Nano Banana Pro, which is the clearest sign of who it is for: shops producing catalogue imagery at scale. The product page lists plugins for Photoshop, After Effects, Premiere Pro, DaVinci Resolve, Blender and ChatGPT. Catches: neither page says the weights are open or that you can self-host, so treat it as closed. The two pages disagree on references (six in the blog, eight on the product page), and fal's own page says Krea 3 is "coming to fal in early access" while Krea's post lists fal as a live partner. The product page labels its comparison "Pre-generated outputs, not a live benchmark." Krea says plainly: "Krea 3 isn't real time yet." Try it: krea.ai/image/k3-edit; the free plan gives 100 compute units a day, and its pricing table lists about 33 edits a day on that plan, so roughly 3 units per edit.

Image

The more interesting Klein story is what happened in the first 24 hours. By this morning a Hugging Face search for the model name returned 38 community repos, a few of them straight mirrors, all created October 15 or 16 (createdAt, Hugging Face API):

Smaller files for smaller cards. realrebelai/FLUX-3-Image-Klein_GGUFs (created October 15) offers compressed versions from about 3.4 GB (Q4) to about 6 GB (Q8) for the ComfyUI-GGUF loader. It does not bundle the text reader or the upsampler; the card points to BFL's ComfyUI repo for those and calls the 9B upsampler "demanding on low-VRAM systems."
The prompt writer, shrunk. prithivMLmods/flux-3-image-klein-prompt-upsampler-GGUF (created October 16) packages the upsampler for llama.cpp, LM Studio or Ollama, from 3.83 GB to 17.9 GB, with Q4_K_M (5.63 GB) as its suggested default. Apache 2.0. This is what lets you run the writer on a different machine, or on the CPU, and hand its caption to Klein.
Macs. Several MLX conversions appeared overnight, including VisualInference/FLUX-3-Image-Klein-MLX-8bit and matching upsampler builds. All unofficial.
The first add-ons. ausboss/FLUX-3-Image-Klein-Outpaint-LoRA (created October 15) extends a picture into grey padding without zooming or shifting the original, which the author says the plain model tends to do.

Research to watch: LensErase, posted by Microsoft, is a foundation model for erasing objects from images; it drew 30 upvotes on Hugging Face's daily papers. A code repo is linked; no weights or demo yet.

Open and local

The local story is that the gap between "lab releases it" and "it runs on a laptop" closed overnight again, and this time the community went straight for the part that does not fit: the prompt upsampler. Shrinking the 9B writer separately from the 5.6B painter means a creator with an 8 GB card can run the writer on the CPU or a second machine and keep the card for pictures. What did not change is the licence ceiling: every Klein derivative inherits the FLUX Kommunity terms and the US$5 million line.

black-forest-labs/flux-3-image-klein: the official code, CLI and local demo for Klein, including the box file format. Why now: released October 15; 57 stars on shields.io (repo).
jub0t/concat: a free, offline CapCut replacement with local auto-captions, voice cloning and background removal, on desktop and phone. Why now: on Trendshift's daily list after v0.2.5 added HDR and scopes and v0.2.6 added hardware AV1 export (October 6); AGPL; 7.1k stars on shields.io (repo).
debpalash/VoiceStudio: a desktop app for voice cloning, dubbing, dictation and audiobooks in 646 languages, built on OmniVoice. Why now: on Trendshift's daily list; AGPL app, with model weights under their own terms; 70k stars on shields.io (repo).
wheresryan22/anatomy: an agent skill that draws working isometric machines as explainer figures. Why now: still on Trendshift's daily list a week after launch (repo).
Leonxlnx/taste-skill: an agent skill that pushes AI-built web design away from the generic look. Why now: on Trendshift's daily list again (repo).

Creative workflows

1. Edit one region with a box in FLUX 3 Image Klein: README, bounding-box section (scripts/cli.py, examples/inputs/balloons_boxes.json)

The steps. Install the repo, download the weights, and run an edit with your photo as the first reference: python scripts/cli.py "put a red bandana around the dog's neck" --image examples/inputs/3.png --out out.png --seed 22. To place or move things, add --boxes with a JSON file of boxes, each with an id, a plain description and coordinates measured from the top-left corner in the order top, left, bottom, right. Keep the seed fixed while you adjust the wording so you can compare runs. In ComfyUI, start from the official image_flux3_klein_distilled_image_edit.json template.

How it works. The box tells the model where a thing should go and how big it should be; the prompt says what it is.

Why it is good. You stop rerolling a whole picture to fix one object.

Where it breaks. The model card is blunt: "Boxes guide placement and scale; they do not constrain pixels to a hard mask," and edits can change content outside the box. The README's box file uses coordinates from 0 to 1, while the model card's in-prompt format uses a 0 to 1000 grid, so check which one your tool expects. The CLI download for the fast model plus upsampler is about 52 GB.

2. Extend a photo to a new shape with the grey-padding add-on: ausboss/FLUX-3-Image-Klein-Outpaint-LoRA (workflow included in the repo)

The steps. Pad your picture with flat grey #808080 to the shape you want, use the padded canvas as the reference, and load the add-on at strength 1.5 in ComfyUI 0.41 or later with AusBoss nodes 2.13.0 or later. Start the prompt with the author's instruction ("Outpaint the image: replace the solid gray areas…") and optionally add Scene: and a description. Settings: 16 steps, CFG 1, euler, simple. Paste the original back over the result with Stitch Inpaint.

How it works. The add-on was trained to treat flat grey as "fill this" and everything else as "keep this."

Why it is good. A 4:5 photo becomes a 9:16 story or a 16:9 banner without a reshoot.

Where it breaks. The author trained it with the original covering 45 to 93 percent of the canvas and tested only around 1 megapixel. It reports occasional duplicated content, a lighter patch at a seam, and one failure in 20 runs. Only tested on the fast model.

3. Split the prompt writer from the painter on a small card: upsampler GGUF plus the Klein README

The steps. Run the upsampler on its own, either with BFL's python -m flux3.pup or with llama.cpp (./llama-server -hf prithivMLmods/flux-3-image-klein-prompt-upsampler-GGUF:Q4_K_M). Paste its long caption into Klein and turn the built-in writer off with --no-pup, or bypass the upsampler node in ComfyUI.

How it works. The writer runs once per image, before the picture starts, so it never has to share the card with the painter.

Why it is good. It is the README's own advice for when both do not fit, and it lets an 8 GB card spend all its memory on the image.

Where it breaks. The GGUF writer is a community conversion, not BFL's file, and a shrunk writer may write worse captions. Captions written without it are what the README warns against.

Worth testing

Klein Space: edit one of your own product photos with a short instruction, then the same instruction with a full scene description. Tradeoff: shared free hardware means queues, and the Space does not expose box controls.
Krea 3 editor: run the same edit there and compare what changed outside the region you asked for. Tradeoff: closed model, and about 33 free edits a day before you pay.
Klein Q4 GGUF in ComfyUI on an 8 GB card. Tradeoff: you still need the text reader and either the upsampler or your own long captions.
Concat for offline captions and cutouts on a phone. Tradeoff: beta software that, in its README's words, "still has edges."

What actually matters from today's signal

Regional editing just became the default. BFL and Krea, two companies that compete on almost nothing else, shipped the same interface on the same day: a box plus a sentence. For anyone doing product, catalogue or social work, this changes the unit of labour from "generate until lucky" to "fix this one thing." That is closer to how retouchers already think, and it makes AI image work easier to quote and easier to bill.

The money split is now clean. Klein is free on your card if your business earns under US$5 million, and the community had it running on Macs and small cards within a day. Krea 3 is six cents an edit with plugins in the apps you already use. At a few hundred edits a month, Krea's convenience costs less than an afternoon of setup. At tens of thousands, a local Klein pays for a graphics card quickly.

The counter-signal: both companies' quality and speed claims are their own, and both models admit edits can drift. BFL says its boxes are not hard masks. Krea's comparison is pre-generated. Before you promise a client "only the mug changes," run a difference check on your own images.

Source access notes: Checked with no new creative launches in the window: OpenAI news (newest creative-adjacent post Oct 7, GPT-6), ElevenLabs blog (newest Sep 30), Stability AI news (newest Aug 25), Luma blog (Oct 14 posts are listicles), Google AI blog (served an undated, stale list). blog.adobe.com failed to fetch; WebSearch surfaced no Adobe creative news dated this week. runway.com/news returned no dated items. fal.ai/models is blocked by robots rules. GitHub's ComfyUI releases page served a stale single release. comfy.org's Klein page failed to load; Comfy facts come from BFL's ComfyUI repo and the Comfy-Org templates. No Midjourney, Kling, Suno, Udio, Ideogram, Pika or Recraft news dated this week surfaced. Video and Audio sections omitted: nothing new from a primary source in the window (Vidu Q4 preview was Oct 7; Helix World's weights repo dates to Sep 3). Writing section omitted. All Hugging Face dates are createdAt from the models and spaces APIs; star totals from shields.io with cache busting; Trendshift used for presence only, no deltas cited. Hacker News used for discovery (Krea 3 surfaced there). Adversarial fact-check ran (Sonnet subagent) against primary sources. It caught and we fixed: the community repo count (38, not about 30), an altered model-card quote on boxes, an overstated LensErase ranking and a missed code link, fal's exact "early access" wording, the missing ChatGPT plugin, Krea's per-edit unit cost (derivable from its plan table), the 8 GB claim being a ComfyUI claim, and a Concat Windows ARM note contradicted by its releases page. The Krea product page's "pre-generated" label rests on our own fetch; the checker could not see it in the blog.