FervorCreative AI
Live Latest 18.11.26 · morning 401 tools tracked 992 workflows indexed 626 topics Hot: Gaussian splatting, ComfyUI, ZONOS3-Mini

Generated work is starting to arrive as parts you can fix one at a time, layers, words and recorded settings, so a revision costs an edit instead of a reroll.

Seedream 5.5ZONOS3-MiniRecraftMiniMax H3 Imageh3metalmondo-harvestVIDEO.mdimage-genaudio-genvoice-clonedesign-toolslocal-creative-aicreative-workflows

Creative AI Briefing: Friday, November 13, 2026

You can now hand ByteDance's Seedream 5.5 a finished poster and get it back as a base image plus a stack of separate transparent PNG layers, each with a name, a position and a stacking order, through hosted services that start at about two cents an image. The same week, Zyphra released a free voice model you can steer word by word, with a companion model for fixing a single word in a take, and a small open tool appeared that reads the settings hidden inside your old ComfyUI renders and turns them into a searchable prompt library. The pattern: generated work is starting to come back as parts instead of a flat file. That changes what a revision costs. You fix the headline layer, the one mispronounced word, the one prompt you want to rerun, and leave everything else exactly where it was.

New models

Seedream 5.5 (ByteDance) is an image generation and editing model built around two things designers have been asking for: edits that leave the rest of the picture alone, and output you can take apart. ByteDance's official page carries no date but promises "Multi-turn editing, consistent every time," with unedited regions kept "intact at the pixel level," and "native support for transparent and semi-transparent backgrounds." Third-party hosts date the release to November 9 (Segmind, Maverickframe); Runware's changelog added it on November 12. The layer feature is the news for designers: Runware says it "splits one reference image into a base image and up to 99 transparent PNG layers, each with its position, stacking order, name and description," with optional text recognition that reports the words, color and closest font. fal sells it as a separate layerize endpoint. Prices (all third-party or reseller): BytePlus advertises "starting at $0.02/image"; Segmind charges about $0.021 at 1K and $0.037 at 2K; fal charges $0.03 per returned image at 1K, so every layer is a billed image (Segmind's own nine-image layer test cost $0.199). Up to 10 reference images; output tops out around 2K (Runware allows up to about 4.6 megapixels). The "100 rounds of editing" figure is ByteDance's claim as relayed by third parties; Segmind tested 12 rounds and found the background held but the frame darkened about half a level per round. The catches from Segmind's hands-on: glass came back opaque when asked for see-through, the layer split returned 2 to 8 layers in its tests rather than dozens, text recognition missed a title, and small layers came back upscaled. No weights; hosted only. No license text on ByteDance's page, so commercial terms are each host's. Free tries: ByteDance links a BytePlus playground and Dola; AlphaSignal reports it in Dreamina and CapCut (secondary).

ZONOS3-Mini (Zyphra) is a free, Apache 2.0 text-to-speech model you can direct like a voice actor. Hugging Face API createdAt is 2026-11-09T05:03:57Z (four days old, not covered here before). It clones a voice from a 5 to 10 second clip, designs new voices by age, accent and texture, covers 43 languages, and steers emotion, pace, projection, liveliness and whisper down to individual words, per the model card and Zyphra's post. Zyphra also released a separate edit model based on VoiceCraft, the approach for changing words inside an existing recording. It is small: Zyphra says 800 million parameters with about 100 million active per word, and the compact download file is about 663 MB. Zyphra quotes under 150 ms to first audio on an M3 MacBook Air and under 570 ms in a browser on a Pixel 9; no memory figure is given. Try it free in the official Space or the in-browser playground. The catch: Apache 2.0 covers the weights, not consent. Clone only voices you have permission to use.

Image

Recraft added the MiniMax H3 lineup to Recraft Studio on November 12. The post lists H3 Image (generation and editing at 2K or 4K, 6 to 13 credits) and H3 Image Highspeed (same settings, faster, 10 to 25 credits). Recraft notes H3 Image cannot output transparent backgrounds, and these models need a Recraft plan that includes external models. For designers already in Recraft, this puts a new 4K model next to the vector tools without a new account.

MiniMax H3 Image now runs on newer Macs, fast. audenvenn/h3metal (created November 12, Apache 2.0 code) is a custom Metal engine for the matrix units in M5 and M6 chips. On the author's M6 Mac mini with 24 GB, H3 Image makes a 2048 by 2048 image in 41 seconds, which the author says is 2.8 times faster than the earlier h3-image-mlx port, and FLUX.2 [klein] 4B makes a 1024 square in 7.4 seconds. These are the author's own numbers on one machine. M1 to M4 Macs are untested and may not run. H3 Image needs a roughly 125 GB checkpoint download to convert, and its license limits where and how you can use it.

Two characters' add-ons in one MOSS-Image picture without their faces blending. shutong-queyue/ComfyUI-MOSS-MultiID (created November 12, MIT, Chinese-language interface) confines each character add-on to its own region of the frame. The README explains why the usual regional-prompt nodes "do not error, but also do not work at all" on this model family. Two routes: a two-pass method that drafts the scene, finds the real faces and redraws each person in place, or a one-pass method with boxes you draw. The catch it states itself: one object spanning two boxes gets one style.

Video

Recraft also added two MiniMax H3 video tiers. H3 Highspeed and H3 Lite make 4 to 15 second clips with sound at 480p or 768p, at 7 to 11 and 4 to 7 credits per second, per the same Recraft post. Sound is always on, both cap at 768p, and you need a plan with video and external models. Good for drafts and social cuts, not finals.

VIDEO.md proposes a standard file for describing a video shot by shot. videomd/videomd (created November 12, Apache 2.0 code, spec under CC BY 4.0) is a folder format: a short card, a timeline with one line per shot, and deeper files per shot, with what was seen kept apart from what was guessed (a made~ key marks inference, with a probability). For editors who already break down reference videos, it is a way to save those notes so an assistant can read them later without re-watching. It is a format with a checker, not a breakdown tool; you or another tool still makes the package.

Audio and music

ZONOS3-Mini leads (see New models). The other audio release is Aria (Abhinav-murugan1/Aria, created November 11, MIT), a Claude Code plugin that writes a script, voices it with inline emotion tags like [whispering], checks every take with speech recognition and re-speaks wrong words, then renders captions, motion graphics and a code-written score cued to the spoken word. Windows only, and voice needs an NVIDIA card with about 6 GB of memory.

Open and local

The local story is records. When an image or clip is cheap to make, the expensive thing is remembering how you made the good one. Two of today's open releases are about that: one reads the recipe out of your old files, the other gives video breakdowns a fixed shape. Star counts are GitHub search API figures at run time (shields.io was unreachable); no daily deltas were available.

  • hodgemann/mondo-harvest: point it at a folder of AI outputs and get one JSON file listing every distinct prompt once, with the model, add-ons, seed and settings behind each output. Created November 11, version 1.3.0 the same day, 2 stars, MIT (repo).
  • audenvenn/h3metal: MiniMax H3 Image, Z-Image-Turbo and FLUX.2 [klein] on M5/M6 Macs with a custom GPU engine. Created November 12, 2 stars, Apache 2.0 (repo).
  • shutong-queyue/ComfyUI-MOSS-MultiID: region-locked character add-ons for multi-person MOSS-Image scenes. Created November 12, 2 stars, MIT (repo).
  • videomd/videomd: an open file format for shot-by-shot video breakdowns, with a checker. Created November 12, 9 stars, Apache 2.0 (repo).
  • yenceesanjeev/2d-character-sketch: an agent skill that builds simple cartoon characters from a small JSON file and renders lip-synced talking sketches on your own machine. Created November 11, 10 stars, MIT (repo).
  • fendouai/ViralForge: breaks a short ad into hook, beats and caption style, then rebuilds the structure around your product as an editable project with ten variants. Created November 11, 73 stars, Apache 2.0 (repo). Its demos use self-made references; copying someone else's video structure is your call, not the tool's.
  • Abhinav-murugan1/Aria: emotion-tagged narration plus code-built video, checked before you see it. Created November 11, 6 stars, MIT (repo).

Creative workflows

1. Find the prompt behind any old AI render (hodgemann/mondo-harvest, single file mondo_harvest.py)

The steps: Install Python 3.10 or newer (add FFmpeg's ffprobe if you want videos read). git clone https://github.com/hodgemann/mondo-harvest.git, then python3 mondo_harvest.py crawl /path/to/outputs. Open harvests/<folder name>.json. Search it for a word you remember from the prompt, and copy the seed and settings to rerun.

How it works: ComfyUI writes the whole workflow into every PNG and many videos. The tool follows that workflow from the save step back to the step that actually drew the picture, so you get the prompt that made the output rather than every text box in the graph. It also reads prompt fields other generators leave in standard image metadata, skips files marked as photographs, and never changes your files.

Why it is good: One entry per prompt, with every file and seed that used it, plus unused prompt text left in the workflow. It saves progress on long crawls. The output is plain JSON, so any editor or spreadsheet can read it.

Where it breaks: It can only read what was saved. Images re-saved through most editors or uploaded to social platforms usually lose this data, and many hosted services write none. Trigger words are stripped by guessing from file names, so check the lora_triggers list. It is two days old with two stars.

2. Split a flat design into layers with Seedream 5.5 (fal layerize endpoint, Runware docs)

The steps: Upload one finished image (Runware allows one reference image when splitting). Turn on text recognition if the design has type. Download the base image and the PNG layers, then stack them in Photoshop, Figma or Affinity using the returned positions and order.

How it works: The model separates the objects it recognizes, fills in the base image behind them, and returns each object as its own transparent PNG with a name, a description and a bounding box.

Why it is good: Swapping a product shot or rewriting a headline becomes a layer edit instead of a full regeneration.

Where it breaks: The background behind each object is invented. Segmind found placed layers differed from the original by 9.4 percent of pixels, so the restacked file is close, not identical. Each returned layer is billed as an image. Output is PNGs, not a PSD. Glass and fine edges are weak.

3. Direct a voice line word by word with ZONOS3-Mini (Zyphra Space, code)

The steps: Record 5 to 10 seconds of the voice you have permission to use. Generate the line. Mark only the words that need different delivery (a whisper, more projection, a slower pace) and regenerate.

How it works: Zyphra says word-level control comes from tracking which audio lines up with which word, so a direction can apply to one word instead of the whole sentence.

Why it is good: It runs free, under a license that allows commercial use, and small enough for a laptop or a browser tab.

Where it breaks: No memory figure is published, cloning quality depends on the clip, and the edit model is a separate download.

Worth testing

  • Seedream 5.5 layerize on a social graphic you already own. Tradeoff: every layer is billed, and the result is close to the original, not pixel-identical.
  • ZONOS3-Mini in the browser playground with your own voice. Tradeoff: no published memory needs for running it yourself, and consent is on you.
  • mondo-harvest on your ComfyUI output folder. Tradeoff: anything that passed through an editor or a social upload is probably blank.
  • MiniMax H3 Image in Recraft Studio if you already pay for Recraft. Tradeoff: no transparency, and you need a plan tier with external models.
  • h3metal on an M5 or M6 Mac. Tradeoff: older Macs untested, and the H3 checkpoint is about 125 GB.

What actually matters from today's signal

The reroll is losing its grip on the workflow. For two years the honest answer to "can you change just the headline" was "I'll generate it again and hope." Seedream's layer split, ZONOS3's word-level control and its edit model all say the same thing from different media: the output should be a set of parts you can open. That saves money, but mostly it saves the approved parts of a job from being redrawn by accident.

The bill moves to bookkeeping. A layer split costs you nine images, not one. A prompt library only works if you stop stripping metadata from your files. A voice you can edit by the word still needs a consent trail. Creators who keep originals, keep records and track who owns what will get the most out of this week's tools.

The counter-signal is that the strongest numbers today are all vendor or reseller claims. "100 rounds" and "99 layers" were tested by nobody we could find at that scale; Segmind's real test found 12 good rounds, a slow darkening and up to 8 layers. Budget for what the hands-on tests show, not the headline.


Source access notes: Hugging Face's createdAt-sorted lists for text-to-video, image-to-video, text-to-image, image-to-image, text-to-speech and text-to-audio returned stale results (newest between August 25 and October 27), so discovery leaned on the trending list, GitHub search and vendor pages; ZONOS3-Mini's date comes from its API createdAt. shields.io and the GitHub repository API were unavailable, so star counts are GitHub search figures. openai.com/news, Adobe's Firefly topic page and Ideogram's blog failed; midjourney.com returned an empty page; seed.bytedance.com's blog index was empty and the Seedream 5.5 page is undated, so its release date is third-party. runwayml.com redirects to runway.com, whose news and changelog showed nothing after October 27. Nothing in the window from BFL (October 15), Stability (August 25), Suno (November 4), Replicate (November 3), fal's blog (October 20), Krea (November 5), Pika (October 29), HeyGen or ComfyUI releases (v0.43.0, October 27). Luma's November 12 posts are tool-comparison guides, omitted. Google's Gemini 4 Hydrogen (November 12) is a text and coding model with no generation features, left to the general briefing. Apex Neural Rendering (game frame generation) is trending but not a creative tool. Recraft's Vectorizer Pro and Adobe's Sneaks were covered yesterday. Writing section omitted.

Articles from this briefing