Creative AI Briefing October 27, 2026 morning edition 12 min read

The add-ons creators installed by hand last week are moving inside the tools they already open, so the work shifts from assembling parts to choosing settings and checking what comes out.

ComfyUI MOSS-Image Anima Reference Adapter CKMoss-Image-RGBA Fibo 2.0 Krea Luma Mini-Apps image-gen local-creative-ai creative-workflows open-weights ai-editing

Creative AI Briefing: Tuesday, October 27, 2026

A compositor can now save a ComfyUI frame as an ACEScg EXR, the colour space most VFX pipelines hand files around in, straight from the stock Save Image (Advanced) node. That option arrived in ComfyUI v0.43.0, published at 02:40 UTC this morning, alongside built-in support for MOSS-Image edit add-ons, a colour-and-depth posing control, and Bria's Fibo 2.0. Krea moved into TouchDesigner the same day, and Luma folded eight single-purpose tools into its agent app. The pattern for anyone making things: pieces you used to hunt down and install separately now ship inside the program you already have open, which means fewer installs and more settings to get right.

New models

Anima Reference Adapter: one picture of your character, any scene you type. ElectronGirl/anima-ref-adapter (created October 27, 05:12 UTC per the Hugging Face API) is a pilot add-on for the Anima anime illustration model. Give it a single reference image and a prompt that describes only the scene; hair, eyes, outfit and accessories come from the picture, while pose, background and framing follow your words. No character names, no appearance tags, no training your own add-on first. It needs its own small node pack (ComfyUI-Anima-Ref-Adapter), because the reference rides along as an extra frame and a plain add-on loader will not pass it in; two example workflows ship in the repo. The card's settings: Euler, 28 to 30 passes, CFG 4.5, reference guidance 4.5 to 6, BF16 (Anima overflows in FP16), and no turbo add-on. License, plainly: the weights fall under the CircleStone Non-Commercial License, which bars commercial or production use of the model file, while Section 2(e) of that same license says you "may use Outputs for any purpose (including for commercial purposes)" and leaves infringement on you; the node code is Apache 2.0. The honest catch, per the card: trained on 300 characters, identity transfers better than drawing style, painterly styles transfer poorly, scars, earrings and unusual skin tones often drop out, large garments can shift colour, and sometimes the character shows up as an object in the scene. No hosted demo found.

CKMoss-Image-RGBA: a second route to transparent MOSS-Image art. ChenkinLab/CKMoss-Image-RGBA (created October 27, 03:15 UTC) is a transparency decoder for MOSS-Image plus an optional add-on, credited to chinoll, the author of the earlier MOSS edit add-on. You can keep the stock MOSS-Image model and swap in only the decoder (workflows/vae_only.json) or run decoder plus add-on at strength 1.0 (workflows/with_lora.json); the card says the two give different edge results without describing the difference. Eight sample PNGs with prompts and seeds sit in samples/. License: GPL-3.0 (the repo also carries an Apache 2.0 file and a NOTICE listing third-party terms), a software license whose terms for generated images the card does not spell out, so ask before client work. Compared with Saturday's paralaif add-on, this one is not built around manga and offers the decoder-only option. The catch: no memory figure, no listed limitations, no evaluation. Test edges over white, black and grey before trusting it.

Image

MOSS-Image editing no longer needs an outside node pack. In v0.43.0, ComfyUI reads MOSS-Image edit add-ons natively (PR #17296, merged October 22). A new reference_latents_method setting matches how each add-on was trained: index_timestep_zero for ai-toolkit add-ons, a separate layout for chinoll's Moss-Image-Edit files. Choose the wrong one and the add-on misreads your photo. The PR reports one layout dropping from 826 to 457 milliseconds per step at 1024 pixels with a 1-megapixel reference. The third-party ComfyUI-Moss-Image-Edit-Native pack has since updated to match the built-in layout; the option name for chinoll's layout changed during review, so check it against your build. A separate PR lowers MOSS-Image's peak memory (#17338) without giving a figure.

Pose a scene with flat colours and a depth map. kijai's PR #17206 (merged October 19, in v0.43.0) adds an experimental node for darask0/MOSS-Image-color-depth-control (created October 19, Apache 2.0). Paint each object a flat colour, add a greyscale depth map with near as white, give each colour its own region prompt, and the image follows your layout. Workflow below. The catch: the depth maps used in training came from Depth Anything V2 Large, which is CC-BY-NC-4.0, and the card's examples run in a separate tool called Hayate, so ComfyUI users start from the PR's moss_image_control_test.json.

A sharpness dial for MOSS-Image. aoleg/MOSS-Sharpness-Slider (created October 26, Apache 2.0) is an add-on with no trigger word: strength around 1.0 sharpens, around -1.0 softens, 0 leaves the model alone. Useful for matching AI elements to a soft lens or a crisp product shot. The card gives no evaluation beyond its own samples.

Krea 3 gets a free allowance. Krea's October 26 changelog says "every plan now includes 9 free Krea 3 images" and removes the login wall from several tools, among them Audio, Seedance Studio, Train and Video Restyle. The changelog does not say whether the nine images reset monthly.

Video

Krea inside TouchDesigner. The same changelog adds a Krea panel, Agent tools and Realtime inside TouchDesigner, the node-based software many VJs and installation artists use for live visuals. That means live AI frames inside a show patch rather than a browser tab. The catch: the changelog is the only source; Krea's docs index has no TouchDesigner page yet, so setup, latency and pricing are unknown.

Pad to a legal size without a hard edge. ComfyUI's new Pad Image node (PR #17224, merged October 21) pads with mirrored edge pixels instead of a flat colour. The PR's reason: models such as LTX need width and height in multiples of 32, and a solid border can turn into a visible edge. Pad, render, crop back. It also leaves values above 1.0 alone, so HDR frames survive the trip.

Luma Mini-Apps. Luma's October 27 post lists eight one-job tools inside Luma Agents: Face Swap, Camera Angles, Light Transfer, Relight Studio, Virtual Try-On, Background Swap, Ad Variants and Product Shots, built on its Uni-1 image model and Ray 3.2 video model. Plus is $30 a month and Pro $90; tiers set how many credits you get rather than which apps, and paid tiers include commercial rights. The catch: the post gives no launch date for the apps themselves, and its "full ad in about 10 minutes" claim comes from Luma's own tutorial.

Audio and music

Sound effects and cloned voices inside a ComfyUI graph. v0.43.0 adds Mirelo SFX nodes (PR #17265) and HeyGen instant voice cloning plus text-to-speech (PR #17043). A short could now get picture, effects and voiceover from one workflow file. The catch: both are paid partner services, and the release notes carry no prices or voice-consent terms; read HeyGen's before cloning anyone.

Open and local

The local story is consolidation. Last week a MOSS-Image editing setup meant a model, an add-on, a custom node pack and a nightly build; this morning's ComfyUI release makes the add-on loading part of the core program, adds Bria's Fibo 2.0 (Dev for text-to-image at 30 passes and CFG 5, Edit Turbo at 4 passes and CFG 1, per PR #17240, weights via Comfy-Org/Fibo-2.0) and gives MiniMax H3 support for single-head add-ons. It is second-day work in the best sense: nothing new to download for the core, and old workflows may need a setting changed. The community kept posting: the mflux-community account uploaded Mac (MLX) conversions of Iris 3B on October 27 and of Anima on October 26, per Hugging Face createdAt.

Comfy-Org/ComfyUI: v0.43.0 adds native MOSS-Image editing, colour/depth control, Fibo 2.0, ACEScg EXR output and mirror padding. Published October 27; 143k stars (shields.io) (release).
ElectronGirl/anima-ref-adapter: put one anime character into any scene from a single image; the file is non-commercial, the pictures may be used commercially per its license; created October 27 (card).
ChenkinLab/CKMoss-Image-RGBA: transparent PNGs from MOSS-Image, decoder-only or with add-on, GPL-3.0, created October 27 (card).
aoleg/MOSS-Sharpness-Slider: one dial from soft to crisp, Apache 2.0, created October 26 (card).
filipstrand/mflux: runs image models natively on Apple Silicon; the community account added Iris 3B and Anima conversions this week. 2.5k stars (shields.io) (repo).
elayadesign/ai-design-skills: coding-agent skills that build landing pages and turn a static design into a looping MP4 or GIF showcase; on today's Trendshift daily list, plus 111 (repo).
Nutlope/hallmark: a skill that pushes AI coding tools away from generic-looking design, first on Trendshift today (repo).

Creative workflows

1. Hand an AI frame to a compositor as an ACEScg EXR. Source: PR #17023 (merged October 23) and PR #17224, both in v0.43.0.

The steps. Update ComfyUI to v0.43.0. If the model needs sizes in multiples of 32, put a Pad Image node before it with padding_type set to pad-to-multiple and padding_style set to mirror; crop back to the original size after rendering. Route the result through Convert Image Color Space with "linear ACEScg" as the destination (the node also reads "HDR ACEScct", "HDR LogC3" and "HDR PQ" sources), then Save Image (Advanced) as EXR with input_color_space set to "linear ACEScg". Bring it into Nuke or Resolve tagged as ACEScg.

How it works. ACEScg is a wide, linear colour space; per the PR, ACEScg 1.0 maps to the same reference white as ComfyUI's existing linear option, and brighter values stay above 1.0 instead of being clipped.

Why it is good. Before this, saving ACEScct material as EXR meant passing through Rec.709 linear, which clipped colours outside that range. Now the file arrives in the space the comp already uses.

Where it breaks. The PR notes that EXR previews inside ComfyUI treat pixels as Rec.709, so ACEScg files look slightly washed out on screen while the file itself is correct. And a container is not a camera: a normal image model's output becomes an ACEScg file, not new highlight detail. The PR names no specific HDR model.

2. Block a scene with flat colours and depth, then let MOSS-Image paint it. Source: darask0/MOSS-Image-color-depth-control card and PR #17206.

The steps. Pose mannequins in any 3D app and export a depth pass (near is white) and a flat ID-colour pass, one colour per object, black for "don't care." The repo's examples/two-people/ folder shows the format, including regions.json. In ComfyUI, load moss_image_control_test.json from the PR. Use moss-control.safetensors, write one region prompt per colour, and end the control at 0.5 of the render for 3D inputs (1.0 for maps taken from real photos).

How it works. The maps steer composition while the image still starts from noise, so lighting and texture come from MOSS-Image.

Why it is good. Art directors can lock a layout with grey mannequins and still get a photographic frame. Inference was about 1.8 seconds per 1024-pixel image on an RTX PRO 6000 with a 4-pass speed add-on, per the card.

Where it breaks. Run the control for the whole render and mannequin surfaces bleed through. Fingers and faces ignore the map. Without region prompts a mannequin map may just draw a mannequin. The card calls the style prior photographic; its training images were Krea 2 Turbo outputs. Evaluation was a handful of scenes and two seeds.

3. A reference sheet for an original anime character, no training. Source: anima-ref-adapter card, workflows/anima_ref_adapter.json.

The steps. Install the adapter into models/loras/, clone ComfyUI-Anima-Ref-Adapter into custom_nodes/, and either install ComfyUI-KJNodes or bypass its attention node. Load the workflow, drop in a clean full-body reference, and write scene-only prompts: quality tags, gender, solo, then place and action. Render a set of scenes at CFG 4.5 with reference guidance 4.5 to 6.

How it works. The reference image enters as an extra frame the model looks at while drawing, so the character's design comes from pixels, not words.

Why it is good. A character designer can test a new design across a dozen situations in an afternoon instead of training an add-on first.

Where it breaks. The file is non-commercial even though the license lets you use the pictures commercially, and only your own character is safe to feed it. Small details drop out, style transfers weakly, and the card lists the character appearing as a prop. Check every frame against the reference before it goes on a sheet.

Worth testing

ComfyUI v0.43.0: free, local. Tradeoff: dozens of merged changes in one release; back up custom nodes first.
Krea Realtime and the new TouchDesigner panel: live frames as you draw. Tradeoff: no TouchDesigner documentation yet.
Luma Mini-Apps: one-click relighting, angles and background swaps. Tradeoff: starts at $30 a month, and claims are Luma's own.
CKMoss-Image-RGBA samples: inspect raw transparent PNGs before downloading the 336 MB decoder. Tradeoff: the author picked them.

What actually matters from today's signal

Native support changes who can use a tool. A MOSS-Image editing add-on that needed a node pack and a nightly build reached a few hundred people; the same add-on in a stable ComfyUI release reaches everyone who presses update. The same goes for a professional detail like ACEScg EXR: it took a merged PR, not a plugin hunt, for an AI frame to land correctly in a VFX comp.

The cost moves from installing to configuring. The edit loader now asks which training layout an add-on used, the colour/depth control wants an end point and region prompts, the reference adapter wants BF16 and no speed add-on. Wrong settings fail without an error, producing a picture that looks almost right. Read the card's numbers before you judge a model by its first render.

The counter-signal is licences, which do not consolidate at all. Today's three community releases carry three different terms: Apache 2.0, GPL-3.0, and a non-commercial file whose pictures you may still sell, and the colour/depth control was trained on maps from a non-commercial depth model. The tool got simpler. What you are allowed to sell did not.

Source access notes: Thin primary-lab window. Checked with nothing new for creators dated October 26 to 27: OpenAI news (GPT-6.1 Terra and research posts only), Adobe blog (latest October 19), BFL (October 15), Stability (August 25), ElevenLabs (September 28), Suno (October 13), fal blog (October 20), Replicate blog (April), Runway news (no dated items), Google blog (no dates rendered), PetaPixel (no dates rendered), CDM (no AI items), No Film School (one sponsored Topview post, omitted). Midjourney via WebSearch only: nothing current. Hugging Face trending list and two createdAt-filtered pipeline feeds returned stale results and were used only for discovery; all Hugging Face dates are API createdAt. The ComfyUI release date is from the GitHub releases API (published 2026-10-27T02:40:46Z). Krea's TouchDesigner integration is sourced to its changelog only. Trendshift figures are shown as daily list placement and daily delta only. VoiceStudio omitted because star sources disagreed sharply (Trendshift 2.4k, shields.io 90k). Civitai, blog.comfy.org and the mflux model cards (fetch error) not read. Adversarial fact-check ran (Sonnet subagent) and confirmed the release time, every PR number and merge date, card settings, licences, prices and star counts. It caught six wording problems, all fixed: an unsupported causal claim about the colour/depth control's photographic look, an overstated claim that the MOSS-Edit-Native pack stands aside (it updated to match; the chinoll layout's option name changed during review), "fan art" added to a non-commercial licence, an incomplete list of Krea's no-login tools, "credits shared across everything" for Luma, and "roughly 60" changes. It could not verify Trendshift placements (taken from this run's Trendshift fetch) or the mflux cards. A separate scoped fact-check of the two articles drawn from this briefing found no factual errors; it tightened the Anima reference-strength range to its APG condition and glossed two technical terms.