Creative AI Briefing October 2, 2026 morning edition 12 min read

Two image labs shipped edit models within a day of each other that promise to leave untouched pixels alone, which turns AI image editing from a reroll into something a retoucher can stack ten changes deep.

Ideogram 4.5 FLUX 3 Image Suno Speech Comfy Agent fal Recast MAI-Voice-2.1 image-gen ai-editing audio-gen comfyui creative-workflows design-tools

Creative AI Briefing: Friday, October 2, 2026

Take a finished product shot, recolor the bottle, swap the label text, then relight the background, and the shadow under the bottle and the grain in the tabletop come back the same on the third pass as on the first. That is the promise of Ideogram 4.5, which shipped on Wednesday, and of FLUX 3 Image, which Black Forest Labs opened to everyone on Thursday. Two labs, inside 27 hours, built their headline around the same unglamorous thing: an edit that changes what you asked for and copies everything else. For anyone who retouches, that is the difference between AI editing as a slot machine and AI editing as a layer stack. Meanwhile Suno taught its music model to talk, and ComfyUI got an assistant that builds node graphs from a sentence.

New models

Ideogram 4.5 is an edit model first, and its pitch is that nothing else moves. Ideogram announced it on X on September 30 at 16:01 UTC, and its model page says it reduces the pixel shifts, color changes and texture artifacts that build up over repeated edits. The API docs make a stronger, more checkable claim for the new Precise Edit endpoint: "pixels the edit doesn't touch are copied exactly from your image, and the result comes back at your image's own width and height." You can add a mask to fence the edit and up to four reference images for a product, material or style. The model page shows an edit on a 4,016 × 6,016 photo (24.2 megapixels) at native size, and lists color and lighting changes, text replacement, restoration, sketch to image, reframing and depth to image. On fal, an edit in the default "regular" mode costs $0.008, $0.03, $0.06 or $0.22 depending on quality tier, and image size does not change the price; the pixel-restoring behavior is a separate "high" precision setting with its own quality tiers, whose prices fal does not list. The catch: the only evidence is Ideogram's own side-by-side comparison against GPT Image 2.5 Sunburst, Nano Banana Pro and Nano Banana 2, whose outputs it says become "unusable within a few edits"; no drift numbers are published. An open-weights release is described as forthcoming, and no 4.5 weights were public at press time. It is live in Ideogram's web and iPhone apps, the API, and partners including fal, Higgsfield and Replicate; Recraft says it is coming to Recraft Studio.

FLUX 3 Image lets you lay out a picture with boxes, then edit it one box at a time. Black Forest Labs posted the launch on October 1 at 19:00 UTC, more than two months after it first announced the FLUX 3 family in July. The product page describes the interface plainly: "drag out a box for every element that matters, then describe what goes in it," on a 0 to 1000 grid that ignores aspect ratio, and then "edit a finished image one box at a time. Everything you didn't touch stays exactly where it was." You can combine up to ten reference images. BFL's own editing docs are more honest than the marketing: "Pixels outside the boxes usually stay the same. Shadows, reflections, or nearby lighting may still change," and "very small boxes can fail." List prices per BFL's pricing docs run $0.041 for 768 square, $0.048 at about 1 MP, $0.100 at about 4 MP and $0.607 at about 16 MP. A half-price launch rate through October 8 is reported by Kingy AI and AlphaSignal and appears on BFL's pricing calculator, but not in the docs. BFL takes commercial licensing inquiries through a contact form; a public open-weight version is reported as promised "for the coming weeks" (Kingy AI, secondary). Try it in the BFL playground; Higgsfield added it the same day.

Suno Speech puts a spoken voice and an original score in one generation. Suno's October 1 post calls it "the first audio model that generates voice and music together as one cohesive track": you "type an idea, a poem or something you've written, then describe the voice and musical style you have in mind." Suno lists dramatic readings, meditations, poetry and pep talks among the uses its testers found; picture a poem read over a cello bed without recording narration and mixing it under a cue yourself. It opened to all users in beta after a month of testing with select users. The catch is in Suno's own words: "Occasionally, British accents can wander off to Australia and back. Dramatic pauses may be very dramatic." Suno's post gives no price, length limit, language list or rights terms specific to Speech; Unite.AI reports it is on web, iOS and Android (secondary). Because voice and music arrive baked into one file, you cannot re-level the narration against the score afterward.

Image

ChatGPT now does virtual try-on. TechCrunch reports a global October 1 launch: upload a selfie or full-body photo plus a garment and ChatGPT shows it on you, built on the ChatGPT Images 2.5 model. Useful for fashion mood boards and quick client comps. Secondary sourcing only; plan limits were not stated.

Video

fal Recast swaps the people in a clip and keeps everything else. fal's Recast page says H3 Max Recast "replaces the main people in a source video with people from reference photos, tracking each person across shots," keeping the original motion, gestures, camera work, shot order and audio. Up to four reference photos, 30 seconds total with no single shot over 15 seconds, at $0.30 per second at 768p or $0.45 at 1080p. A 20-second spot recast at 1080p is $9. The page shows no date; The Neuron's October 1 digest lists it as new. The catch is consent: this is a casting tool and a deepfake tool in the same box, and the page offers only a generic report link.

Higgsfield Genjutsu Restyle (per Higgsfield's changelog, September 30) redraws footage in 20-plus animation styles while keeping "original motion, timing, and camera work." Vendor description only; no samples reviewed here.

Audio and music

MAI-Voice-2.1 keeps one voice across 23 languages. Microsoft's October 1 post says one voice holds its identity in "23 languages and 26 locales" while picking up native accents, with a Flash variant at 150 ms latency. Pricing is $22 per million characters, $15 for Flash. For a creator localizing a narrated explainer, that is the same narrator in German and Mandarin. It lives in Microsoft Foundry, MAI Playground, OpenRouter and Vercel, so it is an API product, not an app.

Japanese voice cloning on a Mac or an old GPU. sakasegawa/irodori-tts-ggml (created October 1, 05:27 UTC) converts Aratako's Irodori-TTS v4.1 Small for speech.cpp. The converter measured 0.23 s to first audio on an Apple M5 and 0.13 s on an RTX 2080, with about 2.1 to 2.2 GB of memory. MIT license (Apache 2.0 on the audio codec), plus stated ethical-use rules against impersonation and deepfakes. Japanese only.

Open and local

The local story is small today and mostly about voices: two people packaged voice-cloning models to run in under 2 GB, one made a talking-head model run on a Mac with no PyTorch, and one put an Apache 2.0 image model on ONNX for non-NVIDIA Windows machines. The big image releases are closed for now; Ideogram's 4.5 weights and FLUX 3 Image's public weights are both promises.

sakasegawa/irodori-tts-ggml: Japanese voice cloning that runs on Metal, Vulkan or CUDA in about 2 GB. Created October 1 (card).
Bonenk/OmniVoice-GGUF: multilingual voice cloning and voice design in about 1.62 GB for the CrispASR runtime, cloning gated behind a consent attestation. Created October 2; CC BY-NC, so no paid work (card).
Avdpro/FlashHead-Pro-MLX: photo plus audio to a 512 × 512 talking-head video on Apple Silicon, 6.92 GB, no PyTorch needed, Apache 2.0. Created September 30 (card).
sorge89/FLUX.2-klein-4B-onnx: the Apache 2.0 Klein 4B image model in about 7.84 GB for ONNX Runtime on CPU, CUDA or DirectML, so Windows users on AMD or Intel graphics have a path. Created October 1 (card).
ideogram-oss/ideogram-4: the open Ideogram 4.0 repo (2.9k stars per shields.io) is where 4.5's promised weights should land; watch it (repo).
Comfy-Org/ComfyUI: 136k stars per shields.io; the in-app agent below runs in Comfy Cloud today, with desktop "coming soon" (repo).

Trendshift and GitHub trending showed no creative-media repositories among today's top movers, so there are no star deltas to report.

Creative workflows

1. Stack edits on one photo without drift, using Ideogram 4.5 Precise Edit (endpoint POST /v2/image/precise-edit/ideogram-4-5, or fal's v4.5 edit).

The steps. Start from your full-resolution master. Make one change per pass ("make the bottle cobalt blue"). If the change belongs to one area, paint a mask (on Replicate's listing, black is edited and white is kept). Feed the output of pass one in as the source for pass two. On fal, set precision to "high," which is the setting that "restores unchanged pixels." Keep each pass's output so you can step back.

How it works. Ideogram regenerates the region you asked about and then copies untouched pixels back from your original, so errors cannot compound in areas you never edited.

Why it is good. Ten small edits stop being ten chances to wreck the skin texture. And because output comes back at source size, you are not upscaling a shrunken file at the end.

Where it breaks. The guarantee covers untouched pixels, not the seam: a recolored object can still pick up a halo where new meets copied. With a mask on fal you get three reference images, not four. No independent drift test exists yet, so check with a difference blend in Photoshop before trusting it on client work.

2. Compose a layout with boxes, then fix one element, using FLUX 3 Image's layout editing.

The steps. In the BFL playground, pick an aspect ratio, draw a box for each element that matters (product, headline, prop), describe what goes in each, add a scene prompt, and generate. To revise, draw a box only around the thing that changes and describe the recolor, replacement or move.

How it works. Boxes use a 0 to 1000 grid on both axes, so the same layout works at any aspect ratio. The model treats everything outside your box as an anchor to keep.

Why it is good. A poster becomes a set of slots you can revise one at a time instead of one prompt you rewrite from scratch.

Where it breaks. BFL's docs say shadows, reflections and nearby lighting "may still change," and "Very small boxes can fail. In our tests, a new element in a box about 40 × 25 pixels often did not appear." Give new elements generous boxes.

3. Describe a ComfyUI graph and let Comfy Agent build it.

The steps. Open Comfy Cloud, open the Agent panel, leave run permission on "Ask," and describe the job ("image to video from this still, five seconds, slow push-in"). The agent finds templates, adds and connects nodes, validates, then asks before each run. Open any workflow tab yourself first; per the docs, "The Agent cannot create or switch tabs itself."

How it works. A frontier language model (the docs price it as Opus 4.8) reads your environment's models and nodes and edits the canvas while you watch.

Why it is good. The node-graph learning curve is the biggest wall in local creative AI, and this turns it into a conversation you can watch and learn from.

Where it breaks. Comfy's blog calls it a beta, it runs in Comfy Cloud only for now, and it bills tokens from the same Comfy Credits pool as your renders. "Auto" mode runs generations without asking, which is how you burn a month of credits on an iteration loop.

Worth testing

Ideogram 4.5 on fal: run five chained edits on one of your own photos; in default mode the Low tier is $0.03 an edit. Tradeoff: the pixel-restoring high-precision mode is opt-in and its prices are unlisted, and there is no published drift measurement, so your difference blend is the test.
FLUX 3 Image playground: lay out a poster with boxes, then move one element. Tradeoff: 4K is $0.607 per image at list, and lighting near an edit can shift.
Suno Speech: score a short poem or toast. Tradeoff: accent drift, exaggerated pauses, and voice and music arrive as one mix you cannot rebalance.
Comfy Agent: free starter tokens cover several chats. Tradeoff: cloud only, beta, and Auto mode spends credits fast.

What actually matters from today's signal

The edit is becoming the product. A year of image-model launches sold the first generation: the prettiest single frame. Ideogram and BFL both just led with the tenth edit instead, which is what working designers actually do all day. When untouched pixels are copied rather than regenerated, an AI edit behaves like a non-destructive layer, and suddenly it fits a retouching or packaging workflow that has client sign-off at step three and revisions at step nine.

The money follows. At $0.03 a pass for Ideogram's default-mode Low tier on fal or $0.048 for FLUX 3 Image at 1 MP, a ten-step revision chain costs well under a dollar, and you keep the master at full size. That undercuts both a reroll habit and an hour of manual masking.

The counter-signal: neither lab published a measurement. "Eliminates artifact buildup" and "everything you didn't touch stays exactly where it was" are both claims backed by curated demos, and BFL's own docs already walk the second one back to "usually." Until somebody runs the difference-blend test across twenty turns, treat both as very good tools to verify, not guarantees to put in a client contract.

Source access notes: Hugging Face dates are createdAt from the API listing (/api/models?sort=createdAt, cache-busted); FlashHead-Pro-MLX confirmed on its own /api/models/ record. X post times are decoded from the status IDs (Ideogram 2026-09-30 16:01 UTC; BFL 2026-10-01 19:00 UTC). Ideogram's API pricing page and BFL's pricing calculator are client-rendered and did not fully render; Ideogram prices come from fal's listing. Maxon's Cinema 4D MCP integration (Cinema 4D 2026.4, Claude, ChatGPT and Codex, disabled by default) was reported by Creative Bloq on September 30, but Maxon's own article 404'd, so it is left out of the main sections. Runway's news page and changelog did not render (newest items via Releasebot were September 23); Runway's "Project Continuum" exists only as a labs post on X and is a research preview. Google's AI blog listing did not render dated items; DeepMind's newest creative item (Gemini 3.8 TTS) predates the window. Stability, Luma's news index, Midjourney updates (blocked), Krea (SEO posts only) and Replicate's blog had nothing in the window. OpenAI's news page showed no creative-media posts. ElevenLabs' newest product post is still Eleven v4. Suno Speech pricing and rights are unpublished; platform availability is Unite.AI's report. Adversarial fact-check pass ran (separate agent, primary sources): it caught an unverifiable Ideogram X quote (replaced with the model page wording), unverified "soon" and "nothing else moves" quote marks, an unsourced VentureBeat "coming weeks" line (replaced), a misquoted BFL small-box warning and a misquoted "within weeks" (both now verbatim), a paraphrase presented as a fal Recast quote, a miscount of local speech packagers, the Irodori license wording, and a softened "commercial weights available" claim. All HF createdAt dates, star counts, prices, X timestamps and arithmetic checked out. ComfyUI GitHub releases page did not render. Civitai not attempted. Writing section omitted: nothing qualified.