Creative AI Briefing October 22, 2026 morning edition 11 min read

The model is becoming the cheap part of creative AI, and what creators now pay for is where it runs, how fast it answers, and whether the endpoint still exists next month.

FLUX 3 Max CorridorKey 2.0 Granite Speech 5.0 MiniMax H3 PTD LoRAs Krea 3 Veo 3.1 video-gen ai-editing local-creative-ai open-weights licensing-provenance creative-workflows

Creative AI Briefing: Thursday, October 22, 2026

You can now pull a clean key off green-screen footage with a free neural keyer, on a four-gigabyte NVIDIA card or, since yesterday, on an Apple Silicon Mac, and use the result on a paid client job. That keyer, CorridorKey 2.0, gives its model away and sells the plugin. The same split runs through the week. fal took Black Forest Labs' video model, post-trained it, and now sells ten-second clips in about twelve seconds of waiting. Krea opened its image API to everyone. And Google's Veo 3.1 preview endpoints can be retired as early as today. The file is getting cheaper. The speed, the wrapper and the uptime are what you pay for now.

New models

FLUX 3 Max turns AI video into something you iterate on instead of waiting for. fal's model went public in mid-October and VentureBeat covered it on October 19. Per fal's launch page, "Black Forest Labs brought the foundation model; fal Research brought post-training." What you can make: 5 to 20 second clips from text, from a start image with an optional end frame, or from a keyframes endpoint, with native audio you can switch off, in aspect ratios from 21:9 to 9:16 (text-to-video endpoint). The headline is speed. fal's own timing puts a 10-second clip at 12.45 seconds against 277.80 for Seedance 2.5, roughly 22 times faster; those are fal's own averages, and VentureBeat notes settings and conditions change the result. Quality is a tie, not a lead: on Artificial Analysis' image-to-video board VentureBeat found 1,138 Elo, level with Seedance 2.5 and two points behind MiniMax H3, with overlapping confidence intervals. Price, in plain terms: $0.0425 per second at 720p and $0.0725 at 1080p (upscaled) until November 7, then $0.17 and $0.29. A ten-second draft costs about 43 cents now and $1.70 later. API output is cleared for commercial use; no open weights. The catch: the price you will budget for in December is four times the price you test with this month, and a "high" reasoning setting (the default) rewrites your prompt for you. Also on Krea and in ComfyUI partner nodes.

CorridorKey 2.0 is a free green-screen keyer you are allowed to use on paid work. Corridor Digital's model, designed and trained by Zachary Harris "in direct collaboration with Niko and Corridor Digital," has weights on Hugging Face (created October 13, per the API) and a GitHub repo that hit today's Trendshift list with 590 stars (shields.io). What you can make: an alpha matte plus a "predicted foreground color," so hair, motion blur and glass come out without the green fringe, from a video file or EXR, PNG or JPG frames. You can run it blind, give it one black-and-white matte for the first frame, or feed a rough roto video, and tell it how much to trust that hint (0.95 for a good matte, 0.50 or lower for a loose one). Overlapping frame windows stop mattes from chattering, and EXR output keeps values above 1.0 for HDR comps. Hardware: an NVIDIA RTX 30-series or newer, 4 GB of graphics memory up to 1440p and 6 GB for 4K. Licence, in plain terms: a modified CC BY-NC-SA 4.0 that says "Use CorridorKey for whatever you'd like, including processing footage for commercial projects," but you cannot sell the tool, and companies over US$20 million in revenue need a plugin seat per user. The paid plugin for Resolve, Nuke, After Effects, Premiere and Blender is $49.50 for a personal licence during beta, rising to $99 on November 21; Mac support for the plugin is in beta. The catch: no-hint mode may key the wrong person in a crowded frame, the first-frame hint "can drift" on long or occluded shots, and it does not convert log or ACES footage for you. No free hosted demo.

Image

Krea 3's API is open to every workspace. Per Krea's changelog, the API went public on October 19 with fast and high modes, up to 8 reference images on enterprise zero-retention requests, and pricing from $0.03 per 1K image (from the October 15 early-access note). A preflight option on Krea's MCP tools quotes the cost before a job runs. On October 20 Krea Agent learned to drive Blender, not only Photoshop. Krea 3 Ultra's API thinking mode adds $0.20 per generation and needs a Max plan and at least one reference image. Heads-up: the old instant and medium values stop working tomorrow, October 23.

MOSS-Image now takes a colour map and a depth map. ComfyUI v0.42.2 added "MOSS-Image Color and Depth Control" (changelog): paint one flat colour per object, add a depth map, and the poster model lays the scene out to match. A community adapter doing the same, darask0/MOSS-Image-color-depth-control (created October 19, Apache 2.0), warns that a map traced from a 3D mannequin can show the mannequin's wood or metal surface; the fix is one region prompt per colour and setting color_end and depth_end to 0.5. Evaluated on "a handful of scenes," its style leans photographic, and only the adapter weights are Apache; its training maps came from non-commercial tools.

Video

Veo 3.1 preview endpoints can be retired as early as today. Google's deprecations table lists veo-3.1-generate-preview, veo-3.1-fast-generate-preview, veo-3.1-lite-generate-preview and gemini-omni-flash-preview with a shutdown date of October 22, 2026, replaced by gemini-omni-1.1-flash. Google calls these the earliest possible dates and promises notice of the exact one. If a script, n8n flow or plugin of yours calls those names, move it now rather than finding out mid-job. Luma's Ray 2 endpoints follow on October 24, per ComfyUI's changelog.

MiniMax H3 reference-to-video in four passes. Taobao's content team posted MiniMax-H3-PTD-LoRAs on October 20: 4-step and 8-step add-ons that turn a reference image into video with sound, with one small adapter trained for each step. The card covers Diffusers only (diffusers >= 0.40.0) and mentions no ComfyUI support, 1344×768 with --megapixels 0.98. No quality numbers against the full model, only three demo cases. Licence: the MiniMax H3 Community License, which excludes the EU, UK, South Korea and the US.

Audio and music

Granite Speech 5.0 gives you Apache-licensed transcription with tighter caption timing. IBM's October 21 post covers three models: granite-speech-5.0-2b (transcribes and translates English, French, German, Spanish, Portuguese and Japanese, and accepts a list of up to 2,000 keywords, which is how you stop it misspelling your guest's name), a fast 470M model for on-device use, and an updated 4.1 2B Plus that labels who is speaking and halves average timestamp shift, "yielding tighter timing in captions." All Apache 2.0, with Apple Silicon support through mlx-audio. The 2B weights date from October 7; the post is what is new. Free demo: granite-speech-5-0-2b Space. Catch: no hardware figures, and the post's own WebGPU demo link is a broken placeholder.

Apple bought assets of Samplab in June. MusicRadar spotted the deal on the European Commission's acquisitions list; Samplab made polyphonic note editing and audio-to-MIDI tools; Logic Pro features are the writer's guess, not an Apple announcement.

Open and local

The local story is a keyer going cross-platform in a week. CorridorKey 2.0's official runtime is NVIDIA-only, and on October 21 Peter Teravskis published an Apple Silicon port whose 738 weight tensors match the official checkpoint exactly; only the runtime changes. On his M4 Pro, 32 frames of 720p took about 37 seconds at full precision using 7.05 GiB, or 39 seconds at 8-bit using 3.28 GiB. That is not real time, and the port skips the HDR EXR path, but a Mac editor can now key without renting a PC.

nikopueringer/CorridorKey2.0: a free neural green and blue screen keyer with a commercial-use licence and HDR EXR output. On today's Trendshift list; 590 stars (repo).
cdelta/CorridorKey-v2-MLX: the same weights for Apple Silicon in full, half and 8-bit versions; the 8-bit version can shift alpha by up to 6.9 percentage points on the author's glass test. Created October 21 (card, code).
taobao-content-ai/MiniMax-H3-PTD-LoRAs: four-pass H3 reference-to-video with sound. Created October 20 (card).
teawhite/ActionRoPE: research checkpoints that steer a top-down game clip with camera moves, on four base models including a 1.2 GB H3 add-on; trained on one game. Created October 20 (card).
TencentARC/Looking-Ahead: Apache 2.0 image-to-video research weights (configs point to Wan 2.1 and 2.2) testing a new way to keep long clips coherent; 63.2 GiB of checkpoints, no demo, no hardware notes. Created October 21 (card).
darask0/MOSS-Image-color-depth-control: lay out a MOSS-Image poster with flat colours and a depth map. Created October 19 (card).

Creative workflows

1. Key a hard shot with a first-frame hint in CorridorKey 2.0. Source: the CorridorKey 2.0 README.

The steps. Install CUDA PyTorch (torch==2.9.1, torchvision==0.24.1), pip install -r requirements.txt, then python scripts/download_weights.py (about 5 GB including three helper models). Paint one black-and-white matte of frame one (white is subject). Run with --hint_first_frame pointing at it and --hint_quality 0.95 if the matte is clean, 0.75 if it came from an automatic tool. Test on a slice first with --num_frames and --start_mode middle. For the comp, add --output_format exr; for linear EXR plates add --frame_dir_linear. Ask for --outputs including checker to eyeball spill, and tune --despill_strength (default 0.5). The drag-and-drop launcher picks up <clip>FrameHint.png automatically.

How it works. The hint tells the model which subject you mean; "carry-hint" then feeds each prediction forward as the next hint, and overlapping frame windows smooth the edges over time.

Why it is good. A deliverable key with clean hair and blur on a 4 GB card.

Where it breaks. The hint drifts on long, crowded or occluded shots, so cut long takes into sections. No log or ACES conversion. Video outputs clip to 0 to 1, so HDR needs EXR. The README says its brightness and contrast controls are for "controlled tests, not for ordinary production use."

2. The same key on a Mac. Source: corridorkey-v2-mlx.

The steps. On Python 3.12: clone the repo, create a venv, python -m pip install '.[download]', then hf download cdelta/CorridorKey-v2-MLX-fp32 --local-dir <folder>. Run corridorkey-mlx --compile --image-sequence --model-root <folder> --input plate.png --output-dir outputs/plate --outputs alpha fg raw_fg checker cutout. Add --precision fp16 or q8 only if memory is tight.

How it works. Same weights, rewritten for Apple's GPU.

Why it is good. No NVIDIA card needed, and full precision matches the reference run to within 1/65535.

Where it breaks. Only one Mac tested (M4 Pro, 48 GiB), SDR output only, no audio passthrough, and shapes must match a prebuilt resolution profile.

3. Four-pass H3 reference-to-video for drafts. Source: MiniMax-H3-PTD-LoRAs.

The steps. In Diffusers 0.40 or newer, load the H3 Ref2VA pipeline, call apply_ptd_lora with MiniMax-H3-Ref2VA-PTD-4Step.safetensors (it reads the step count from the file), and set num_inference_steps to the nfe value it returns plus one. Weight 1.0, --megapixels 0.98 for 1344×768. Keep the pipeline's video shift 12.0 and audio shift 3.0.

How it works. Each pass gets its own small trained adapter instead of one adapter stretched across all passes.

Why it is good. Four passes is fast enough to check blocking and voice timing before a full render.

Where it breaks. The card mentions no ComfyUI path, three demo cases and no measured quality, and the territory clause in the licence.

Worth testing

FLUX 3 Max on fal: ten-second drafts at about 43 cents. Tradeoff: paid only, and the price quadruples after November 7.
granite-speech-5-0-2b Space: free transcription with keyword lists. Tradeoff: speaker labels live in the separate 4.1 Plus model.
CorridorKey 2.0: free commercial keying. Tradeoff: no hosted demo, NVIDIA or a Mac port with one tested machine.
Krea MCP with preflight: see the bill before the batch. Tradeoff: style references and LoRAs return "coming soon."

What actually matters from today's signal

The model stopped being the scarce thing. CorridorKey gives away the keying model and charges $49.50 for the panel inside Resolve. fal did not invent FLUX 3; it post-trained it and sells the twelve-second turnaround. Krea's newest features are a cost quote and an agent that drives Blender. The decision is now "which wrapper fits my timeline and invoice," not "which model is best."

Speed changes the craft more than quality does. When a clip costs 43 cents and arrives before you finish your coffee, you stop writing the perfect prompt and start directing by revision, the way you would with a junior animator. Budget for the rerolls, though. VentureBeat notes that faster generation can invite more attempts, and at $1.70 a draft after November 7, forty tries on one shot is a $68 line item.

The counter-signal is Google's deprecation table. Every wrapper you lean on is an endpoint someone else can retire, and today four Veo and Omni preview names hit their earliest shutdown date. Keep your prompts, seeds and source images in your own folders, pin the model name your pipeline calls, and treat any "preview" in a model ID as a countdown.

Source access notes: Hugging Face dates from the API createdAt field via WebFetch; the cloud shell's direct access to Hugging Face and GitHub was blocked by the egress proxy this run, so GitHub star counts come from shields.io (the GitHub page itself showed a stale 10 stars). CorridorKey 2.0's public announcement date was not found; weights date October 13, and its appearance here rests on today's Trendshift list. FLUX 3 Max timing and benchmark figures are fal's and Artificial Analysis' (preliminary); no BFL post exists on bfl.ai. OpenAI news returned a client error; Google's AI blog showed no October creative posts; ElevenLabs, Replicate, Stability, Runway news, BFL and Midjourney (via Releasebot, last updated October 10) had nothing new in the window. Hacker News search returned nothing dated in the window. Civitai and blog.comfy.org were not fetched. Adversarial fact-check ran (Sonnet subagent) and caught twelve issues, all fixed: FLUX 3 Max's launch date (October 19 is the press date; fal's pricing began October 15), a misattributed fal quote, an overstated "without warning" for the Veo shutdown, Krea's preflight scoped to MCP tools, Ultra's reference-image requirement, a non-verbatim MOSS adapter quote plus its non-commercial training-map caveat, the CorridorKey plugin price as the personal tier, an unsupported "four-figure plugin" line, a non-verbatim "byte for byte," the PTD step setting, and two inferred claims (ComfyUI, Wan base) now hedged. Items the checker could not reach but this run read directly: the Granite Space (Hugging Face API, created October 21), Krea's October 20 FLUX 3 Max and Blender entries, the Trendshift listing, and the CorridorKey README lines on log footage, clipping and experimental controls. Yesterday's run (October 21) saved its briefing and articles to the claude.ai Project only; its two ledger rows were added to the local ledger this run.