FLUX 3 Klein, MOSS-Image, Krea 3: creator news, Oct 20

If you post the same face every day, the week of October 13 gave you three new ways to keep it, and one new license to read before you use any of them. FLUX 3 Image Klein, MOSS-Image and Krea 3 arrived within four days of each other, and ComfyUI shipped three releases around them. This briefing sorts them by what they change for a character feed, and says plainly where we could only read a vendor's own claims.

The week at a glance

FLUX 3 Image Klein: the one to test first

Klein is a 5.6B-parameter model for text-to-image, editing and multi-reference work, released October 15. Black Forest Labs ships a base checkpoint for fine-tuning and a step-distilled one for speed, plus FP8, INT8 and NVFP4 builds. The generator weights need 11.2 GB in BF16 and 5.6 GB in FP8, and BFL says the INT8 build runs on GeForce RTX cards with as little as 8 GB of VRAM.

For a recurring face the headline is multi-reference: up to 10 reference images in one request, which is the feature that decides whether a character survives a change of scene. ComfyUI's changelog lists native support for both checkpoints with the Qwen3-VL-4B text encoder and the existing Flux2 VAE, so no custom node pack is needed. Our missing nodes guide still applies if a template asks for something your install lacks.

bfl.aiOpen the page

The license is the catch. The weights use the FLUX Kommunity License, which BFL says allows free commercial use for individuals and companies with annual revenue under $5M. Larger organizations need self-hosted licensing or the API. VentureBeat notes that the related FLUX 3 Action license counts revenue together with affiliates, so a small studio owned by a large company may not qualify; read the license that ships with the download before you rely on it.

One caution on the numbers. The speed and quality charts are BFL's own. VentureBeat reports 4.1 seconds for a 1024 by 1024 image on an RTX 5090 and 8.9 seconds on an RTX 4070 for the distilled INT8 model, and says the preference charts give no detail on prompts, judging or sample size. We have not reproduced them.

MOSS-Image: open, permissive, and slow so far

OpenMOSS introduced MOSS-Image on October 16. It has a compact 5B backbone, two checkpoints (Base for fine-tuning, SFT tuned on curated data), and weights under Apache-2.0, the most permissive license of the three. Its pitch is layout: it can take bounding boxes and text descriptions as structured prompts. The model card recommends 50 steps, CFG 3 and shift 5, and says a native ComfyUI plugin installs through the Manager while integration into the official ComfyUI repository is still in progress. The 0.40 to 0.41.1 changelog entries we read do not mention it. We have since published a full MOSS-Image review with file sizes, speed tests and the face verdict.

The Apache-2.0 line covers the weights. The project page adds that the pipeline code and third-party components keep their own licenses, so check those if you ship anything built on the code. File sizes are a separate story, and our download size study found MOSS-Image's text encoder is the biggest file in its stack.

That is three people, not a survey, and the mood is mixed: promising variance and a permissive license, against slow generation and unconvincing first results. Treat the Base checkpoint as a training target and the SFT checkpoint as the one to try for generation.

Krea 3: fast, hosted, not something you download

Krea 3 is an editing model, released October 15 and available on krea.ai and through hosts including fal, Replicate, Together AI and OpenRouter. Krea says most edits finish in under 8 seconds at native 2K and prices it at $0.06 per 2K image and $0.03 per 1K image. Krea's post does not mention open weights, and neither does anything else we found, so assume it is API only.

Reference limits differ by source: Krea's post says up to 6 reference images per edit, while ComfyUI's 0.41.1 notes for the partner node say up to 8, with 1K or 2K output and a precise edit mode that leaves the rest of the picture unchanged. Both can be true of different surfaces, so check the one you call. If you used Krea 2 for faces, our Krea 2 versus Qwen-Image 2.1 comparison is still the open-weights baseline.

The three side by side

What to do this week

Test Klein with your own references before you change anything else: it is the only one that is both open and built for multi-reference. If it keeps your face, check the revenue line in its license against your account, then decide whether the distilled checkpoint is fast enough for daily use. If your results vary run to run, our distilled versus base explainer covers when to train on one and generate with the other.

Leave MOSS-Image for a weekend of testing unless you need Apache-2.0 or layout control, and keep Krea 3 as a paid fallback for hard edits rather than a pipeline dependency. Short on VRAM? Rent a card for the test instead of buying one, as described in the rent-a-GPU help page. For the earlier releases, see October's first briefing.

Musetta studioDemo data

Where Musetta fits

Musetta runs on your own ComfyUI and keeps each character's reference photos, so trying Klein on a character means installing what the workflow is missing rather than rebuilding the character. It does not change any license above; those apply to you. See how Musetta compares with working in ComfyUI directly.