Creative AI Briefing November 2, 2026 morning edition 10 min read

A generated draft now costs about as much as a pencil sketch, so the work that separates good AI projects from bad ones has moved to the judging step, where creators score, reject and redraw.

TBC mischief 1 MAI Image 2.7 huashu-bird-sticker Anima RGBA Patina teardown video-gen image-gen creative-workflows comfyui design-tools licensing-provenance

Creative AI Briefing: Monday, November 2, 2026

A storyboard artist can now turn forty still frames into moving animatic shots for about a dollar: TBC's mischief 1 animates a still into a five-second clip at half a cent per second of 480p video (720p doubles it). That price, and a week of community releases that lean hard on review, point one way. When a draft costs pennies, the scarce skill is the review. The most instructive project this week is a set of 11,064 AI bird illustrations whose author published exactly which 3,394 failed inspection, and why.

New models

TBC mischief 1, a sketch-priced image-to-video model. The Biological Computing Co. launched mischief 1 on October 30, an image-to-video model built on the open Wan 2.2. You give it a starting image and a prompt and get a three- or five-second clip, which makes it a fit for animatics, pitch boards and testing whether a shot idea moves at all before you spend real credits. On fal it costs $0.005 per second at 480p and $0.01 per second at 720p, so a five-second clip runs 2.5 or 5 cents. It is also on OpenRouter from $0.005 per second through a single provider. The specs set the ceiling: 16 frames per second, landscape only (832x480 or 1280x720), five seconds maximum. fal labels it for commercial use but publishes no terms beyond that label, and no weights are released. The catch: TBC says the speed comes from software derived from lab-grown neurons and claims it is "5x faster than its base model" with the same quality, but cites no independent benchmark. By default fal rewrites your prompt against the image (switch prompt_processing to original to keep your wording), and fal's docs say TBC's policy service screens every prompt and image. Free demo: none; the fal playground needs a sign-in and bills per run.

Image

Microsoft's MAI Image 2.7 is an editor first. MAI Image 2.7 (announced October 27, in preview) handles restyling and relighting a scene, editing text and symbols inside an image, and restoring and cleaning up existing images. Microsoft says it debuted at No. 3 on the Arena edit leaderboard and No. 2 on Artificial Analysis's image-editing board, with its biggest lead in scene and style changes. Those are third-party leaderboards, but the ranking claim is Microsoft's. Text-to-image is "coming soon." The catch: no price, API or usage terms are published, so this is a model to test, not to build a client pipeline on. Try it free on Arena (the link opens the text-to-image board), Artificial Analysis or MAI Playground (the playground link still names version 2.6).

Tiling textures from a sentence, inside ComfyUI. ComfyUI v0.43.1 (October 28) added three Patina nodes that return the full set of maps a 3D material needs: base colour, normal, roughness, metalness and height. Patina Material makes a texture from a prompt (or restyles an image) that tiles across, down or both, at up to 2048 pixels a side. Patina Material Extract pulls one surface out of a photo when you name it, such as "the wall." The photo-to-maps node has no tiling control. These are paid partner nodes that call a fal-hosted service; per the node source, a one-megapixel texture from the Material node works out to roughly 6 cents with one map and 7 cents with the default two; the photo-to-maps node starts near 3 cents, and the Extract node's base charge is ten times the Material node's. The catch: your materials depend on a hosted service and its pricing.

Transparent anime characters with one swapped file. localai0301/Anima-RGBA (created November 2, 00:28 UTC) gives the Anima illustration model a real transparency channel: a 508 MB VAE (the encoder-decoder that turns the model's internal image into pixels) that outputs alpha and a 184 MB add-on that teaches Anima to draw for it. No custom nodes: load that file in place of the usual one, add the add-on at strength 1.0, and put "transparent background" in your prompt. The author's limits are honest: the decoder trained on 512-pixel crops, so fine semi-transparent detail (the card names hair strands, glass and smoke) is the weakest area, edges are mostly hard rather than soft, and close-ups sometimes keep a white background. The license catch matters: the add-on is tagged Apache 2.0, but the Anima base model carries the CircleStone Labs non-commercial license, so sprites made this way are for personal and portfolio work unless you clear that.

4K posters in about 75 seconds from Krea 2 Turbo. lvladikov/Krea2Turbo-PiD-SuperFast-HighRes (created November 1, 12:57 UTC) is second-day work: the same author's small-render, big-decode trick from Friday's MOSS-Image version, now on Krea 2 Turbo. A 512x896 render becomes 2048x3584 through NVIDIA's PiD decoder, with no separate upscaler. Workflow below.

Audio and music

A thin day. The one creator-facing item is ComfyUI v0.43.1's VEED Dubbing node (October 28): it translates the speech in a video, clones each speaker's voice and re-syncs the lips to the new language. It is a partner node that calls an outside service; the changelog lists no price.

Open and local

The local story this week is agent skills and asset libraries that ship with their own quality control. The ones worth a look either render their own output and inspect it, or publish their failure rate.

alchaincyf/huashu-bird-sticker: a transparent sticker for every one of the world's 11,064 living bird species, drawn with GPT-image and individually scored against a reference photo; created October 30, 548 stars per GitHub's search API (repo).
elvarofficial/teardown: a Claude Code skill that builds a take-apart 3D machine as one offline web page, then renders a 60 fps film with generated sound mastered to streaming loudness; MIT, created October 30, 1,094 stars per the search API (repo).
kevinraymond/florilegium: Delft-blue watercolour floral margins grown by code at typesetting time, seeded by page number so each page gets its own garden, exportable to SVG and PostScript; MIT, created October 30, 1,327 stars, with a browser playground (repo).
caru-ini/stack-diagram: an agent skill that turns one sentence about a tech stack into a clean architecture diagram with official logos, as HTML, PNG or a React component; MIT, created November 1, 796 stars (repo).
patch-gg/pixelfx: the 244 code-based pixel-art effects from yesterday's briefing, now at 3,113 stars on the search API, up from 1,889 a day earlier (repo).
lvladikov/Krea2Turbo-PiD-SuperFast-HighRes: 4K text-to-image in two stages for ComfyUI or Comfy Cloud; workflows Apache 2.0, model files under their own licenses, created November 1 (repo).
localai0301/Anima-RGBA: real transparency for Anima; created November 2 (repo).

Star totals are from GitHub's search API on November 2. No daily deltas were available apart from the pixelfx comparison.

Creative workflows

1. Make a large illustration set that holds up, the huashu-bird-sticker way (README.md, species.json)

The steps. For each subject, find a reference photo (the example record links to Wikimedia Commons; 453 species needed a board of several photos). Draw the sticker with GPT-image, upscale it and cut it out. Have a separate reviewer score each drawing against its reference on a fixed checklist: bill, eye, legs and feet, plumage pattern and colour. Anything under 80 is redrawn, up to a limit, and the best version is kept. Borderline scores (80 to 82) get a second, blind review; 211 of 852 failed it and were redrawn. Record the score and a pass flag (reviewScore, meetsQualityBar) in the data file.

How it works. The image model is fast and cheap but confidently wrong about small details. The checklist turns "does this look right?" into five yes-or-no questions a reviewer can answer the same way ten thousand times.

Why it is good. You get a consistent library (transparent WebP, about 1,000 pixels on the long side, white outline, names in 43 languages) and an honest map of which pieces to trust: 7,670 pass, 3,394 do not.

Where it breaks. The README does not name the reviewer, the upscaler or the cutout tool, and gives no cost or time. The license is PolyForm Shield 1.0.0: commercial use, editing and redistribution are allowed with a credit link, except in any birding, bird-ID or field-guide product. The author calls them "illustrations, not field-guide plates."

2. Transparent character sprites with Anima-RGBA (vae/anima_rgba_vae.safetensors, loras/anima_rgba_lora.safetensors)

The steps. Open any working Anima workflow. Put the decoder file in models/vae and select it in Load VAE. Add the add-on file with a standard LoRA loader at 1.0. Add "transparent background" to the prompt; if the background stays solid, raise the phrase's weight, for example (transparent background:1.3), rather than the add-on strength. Save Image writes a PNG with alpha. Author's example settings: 30 steps, guidance 4.5, 1024x1024.

How it works. The stock decoder only knows three colour channels. This one adds a fourth for transparency, and the add-on nudges Anima toward images that decode cleanly into it. Without the add-on the alpha comes out hazy.

Why it is good. No cutout pass, no halo from background removal, and it drops into a workflow you already have.

Where it breaks. Hair, smoke and glass; see-through clothing comes out solid; stacking other add-ons makes the white-background failure more common. And the base model's non-commercial license governs what you can sell.

3. 4K images in two stages with Krea 2 Turbo × PiD (local and Comfy Cloud workflow JSONs, krea2-pid-generate.py)

The steps. Load the workflow, keep the Krea render at 512 pixels or more on its long side (any multiple of 16), run Krea 2 Turbo for 4 steps with the author's speed add-on, then let PiD decode at four times the size in 4 more steps. The repo includes 100 example prompts with seeds.

How it works. Instead of rendering big or upscaling afterward, the small render is handed to a decoder that paints the fine detail as it enlarges.

Why it is good. The author reports about 75 seconds total on an RTX 5090 with offloading, the PiD stage taking 3.3 seconds. On Comfy Cloud the README puts it at about 1.4 credits per image when ten prompts share one job.

Where it breaks. The Krea model file alone is 26.3 GB, the script is NVIDIA-only (Mac users use the workflows), and PiD ships under NVIDIA's source code license, which is not open for commercial use. Use it for pitches and personal work.

Worth testing

mischief 1 on fal: turn a storyboard page into moving shots for pennies. Tradeoff: 16 fps, landscape only, five seconds, and a sign-in before the first run.
MAI Playground: try relighting and text edits on your own images. Tradeoff: preview only, no published pricing or terms.
florilegium playground: grow ornamental margins in the browser before touching LaTeX. Tradeoff: one house style, Delft blue, by design.
huashu-bird-sticker files: drop stickers into a poster or app by raw URL. Tradeoff: the full repo is about 700 MB, so sparse-checkout what you need, and check meetsQualityBar first.

What actually matters from today's signal

The price of a draft has collapsed. Two and a half cents for a moving shot, about 6 cents for a tiling texture, 75 seconds for a 4K poster. None of those outputs is a finished product, and the people doing the best work this week have stopped pretending otherwise. The bird sticker project is the model to copy: it generated freely, then spent its real effort on a checklist, a score per image, a blind second look at the borderline ones and a public record of 3,394 failures. teardown does the same thing in motion, rendering review sheets of stills from its own film so the result can be checked.

For a creator, that shifts where your hours go. Generation is now the cheap part of the job, and your eye is the expensive part. If you are building a batch workflow, write the rejection checklist before the prompt.

The counter-signal is licensing, again. Anima-RGBA wears an Apache badge on top of a non-commercial base model. The Krea 4K workflow is Apache, but its decoder is not. mischief 1 says "commercial use" without terms. The permissive label usually describes the glue. Read the license of the model underneath before a client sees the result.

Source access notes: Adversarial fact-check pass ran (Sonnet subagent): it corrected the Patina price (the Material node is about 6 cents per one-megapixel texture, not 1.4), the Anima-RGBA file description (full VAE, not decoder only), the MOSS-PiD timing (Friday, not last week), MAI's restoration wording, the VEED price wording, and an unsupported superlative; it confirmed all dates, licenses, star counts and fal specs. Thin day for lab releases. openai.com/news failed to load. Adobe's blog had no Firefly posts after October 28; ElevenLabs, fal, BFL, Suno and Replicate blogs had nothing newer than October 20; Luma's and Runway's latest items (October 26 to 27) were already covered; DeepMind's newest creative item (Genie 4) was covered October 29. Midjourney, Civitai and blog.comfy.org not attempted (they habitually block); ComfyUI news came from docs.comfy.org. Hugging Face dates are API createdAt; the HF listing endpoints served stale September data without cache busting. GitHub star totals come from the search API; per-repo HTML pages showed stale counts. Trendshift showed no star deltas. shields.io not used.