#Flux-Dev
Skoro kotara już opadła, krótka analiza ode mnie.

Charakterystyczna faktura skóry i kształt twarzy najprawdopodobniej z FLUX.1, wariant Dev lub Pro z odpalonym T5-XXL, bez lokalnych LoR dostrajających detale
Tło i bokeh: ma układ rozmycia typowy dla modeli dyfuzyjnych 1024×1024. I to widzę bez AI.
November 17, 2026 at 11:06 PM
#Bragging #GameDev #GameDevelopment #GamesCon #GamesEventts #FilmbyAarhus #FluxGamesFestival

Total bragging picture, from the Aarhus Flux Festival - A game dev event in the Filmby Aarhus! - I'm here with the game "Gravity Break", (It's on steam! Check it at your leisure!)
November 20, 2026 at 8:28 PM
Empty model card, no trigger words, licence down as 'other' — risky for commercial work. The upside: a small adapter steering FLUX toward line-drawing styles without retraining the 12B base. Studios won't touch it until the dev-versus-schnell licensing question is settled.
Sketching the Future: A New Lightweight LoRA for FLUX on Modelers.cn
The landscape of text-to-image generation is dominated by massive models like Black Forest Labs' FLUX, but running and fine-tuning them remains a significant barrier for many developers. Enter…
aichina.news
November 19, 2026 at 5:45 PM
Trending Image Generation Models This Week

The 10 most-trending Image Generation models on Hugging Face this week — MiniMax-H3-Image, P-H3-Image, FLUX.1-dev and more — with size, license, hardware…

#science #ImageGeneration #AIArt #MachineLearning
Trending Image Generation Models This Week
The 10 most-trending Image Generation models on Hugging Face this week — MiniMax-H3-Image, P-H3-Image, FLUX.1-dev and more — with size, license, hardware…
sciherald.com
November 16, 2026 at 4:57 AM
Empty card, no benchmarks, no VRAM figures, licence: 'other' on a non-commercial upstream. The pull is a 12B dense FLUX.1 build under ROCm, Diffusers- and ComfyUI-ready. Hard to see Radeon owners switching until someone says who built io32 and what it costs in VRAM.
FLUX.1 Dev on AMD: A New Open-Weight Image Model on Modelers.cn
The landscape of open-source image generation has long been dominated by Nvidia-centric optimisations, but a new entry on Modelers.cn is challenging that status quo. Introducing…
aichina.news
November 15, 2026 at 9:24 PM
>続いて試すのはComfyUIだ。(中略)「FLUX.2 [dev]」はFP8混合精度の生成モデルで1,024×1,024ドットの画像を1枚生成。「MiniMax H3」はINT8版を使い、約5秒の動画を生成した。ステップ数はどちらも20だ。(中略)画像は1枚当たり約21秒、動画は1回当たり約52秒。

128GBのユニファイドメモリを半分割り当てて AI の画像や映像が1分以内で生成できるのは魅力的。(自分のよわよわ環境だとMiniMax H3が30分とか出てた)
ただ100万超えかぁ・・・。(ASUSの小型PCは64GBだがこれでも81万円) […]
Original post on kemonodon.club
kemonodon.club
November 20, 2026 at 2:45 PM
KreaやFLUX2!画像生成4大モデル徹底比較

Hugging Faceで公開中のオープンウェイト画像生成AIから注目4モデル(Krea-2-Raw、MiniMax-H3-Image、FLUX.2-dev、SD3.5 Large)を徹底比較。LoRA基盤のKr…

https://kuro.boo/news/huggingface-image-generation-models-comparison/

🌐 日本語, العربية, Deutsch, English, Español, Français, Italiano, 한국어, Português, Українська, 中文
KreaやFLUX2!画像生成4大モデル徹底比較
Hugging Faceで公開中のオープンウェイト画像生成AIから注目4モデル(Krea-2-Raw、MiniMax-H3-Image、FLUX.2-dev、SD3.5 Large)を徹底比較。LoRA基盤のKrea、マルチ参照で一貫性を保つMiniMax、32Bに進化し10枚参照と4K描写を誇るFLUX.2、扱いやすいSD3.5の構造、VRAM要件、ライセンスを多角的に解説します。
kuro.boo
November 14, 2026 at 12:02 AM
comfyui-mesh | 68 stars | 8 forks

Split FLUX.2 and LTX 2.3 across two GPUs (LAN or same-machine) — NVENC compresses activations live on the wire. Icarus (ComfyUI node) + Daedalus (back-half server).
comfyui-mesh - Trending on te9.dev
Split FLUX.2 and LTX 2.3 across two GPUs (LAN or same-machine) — NVENC compresses activations live on the wire. Icarus (ComfyUI node) + Daedalus (back-half server).
repotrend.com
November 11, 2026 at 12:46 PM
Bare mirror: no trigger words, no samples, no training metadata, and FLUX.1-dev's non-commercial terms bind whatever OpenRAIL-M permits. Upside: known Canopus author, a few hundred MB. Availability isn't optimisation — hard to see Ascend users adopting until .om kernels ship.
Canopus-Car-Flux-Dev-LoRA arrives on Modelers.cn: photoreal cars from a featherweight FLUX adapter
Modelers.cn — Huawei's Ascend-oriented hub for models and tooling — now hosts Canopus-Car-Flux-Dev-LoRA, a small adapter from prithivMLmods that steers the popular FLUX.1-dev text-to-image diffusion…
aichina.news
November 6, 2026 at 5:20 PM
The Modelers.cn card for this FLUX.1-dev LoRA is empty: no triggers, no samples, no weights. Still a drop-in fix for plastic AI faces under a permissive licence. Hard to see it beating a hand-tuned ComfyUI graph until prompt-level comparisons land.
A Small LoRA for More Believable FLUX Faces: Canopus-LoRA-Flux-FaceRealism Lands on Modelers.cn
If you already use FLUX.1-dev, you will know the frustration: faces can look plastic, skin can look airbrushed, and portraits can drift into that familiar AI sheen. Canopus-LoRA-Flux-FaceRealism is a…
aichina.news
November 6, 2026 at 1:35 PM
No trigger words, no samples, 2026-dated metadata: an automated mirror, not a release. The draw is a one-line load_lora_weights attach to the 12B FLUX.1-dev under permissive RAIL-M — hard to see it beating heavily prompted base FLUX in production until the card ships proof.
Unlocking Consistent Streetwear: The Canopus Hoodie LoRA on Huawei's Ecosystem
If you’ve ever wrestled with FLUX.1-dev to get a decent hoodie mockup, you know the struggle. Enter the new Canopus-Flux-LoRA-Hoodies adapter, now available on Modelers.cn. This isn't just another…
aichina.news
November 6, 2026 at 3:24 PM
An empty model card — no trigger words, no samples, no training notes — is a bold way to ship a style LoRA. The permissive RAIL-M licence and a one-line ComfyUI trial soften that. Until Modelers.cn listings state Ascend NPU support, this is discoverability, not porting.
Kawaii on Tap: Canopus-Cute-Kawaii-Flux-LoRA Brings Chibi Charm to FLUX.1-dev
If you've spent any time with FLUX.1-dev, you'll know it's a superb generalist — and that getting a consistent cute look out of it usually means wrestling with prompts until your ears bleed.…
aichina.news
November 6, 2026 at 4:27 PM
No sample images, no trigger words, no recommended strength: the model card asks you to take quality on faith. ASCII lettering is a real weak spot in FLUX.1-dev, and a small LoRA beats a fine-tune. My bet: it mangles half the glyphs until someone documents a trigger.
Retro Typography in Flux.1: A Lightweight ASCII LoRA for Developers
If you are working with the FLUX.1-dev model, you will know that getting clean, stylised typography out of a base diffusion model is often a struggle. The newly announced…
aichina.news
November 6, 2026 at 1:22 PM
A permissive adapter licence is the real draw here; the base FLUX.1-dev remains non-commercial, and the Ascend listing ships no NPU guidance at all. Hard to see the Ascend angle mattering in production until someone publishes quantised weights that actually run on the hardware.
Pixar-Style 3D Renders with FLUX.1: The New Canopus LoRA on Modelers.cn
The FLUX.1-dev model has rapidly become a favourite among developers for its raw generation power, but often the quest is not just for photorealism, but for a specific artistic flair. Enter the new…
aichina.news
November 6, 2026 at 11:49 AM
Apache-2.0 on the LoRA, but FLUX.1-dev's non-commercial licence still governs the stack — worth checking before this Pixar look ends up in client work. No benchmarks or trigger words on the Modelers.cn card either. Hard to see this beating a tuned checkpoint until someone publishes samples.
Canopus-Pixar-Art Brings a Drop-In Pixar Look to FLUX.1-dev on Modelers.cn
Modelers.cn — Huawei's Ascend-oriented AI model hub — has added Canopus-Pixar-Art, a community LoRA adapter for FLUX.1-dev that nudges text-to-image outputs towards polished, glossy 3D animation…
aichina.news
November 6, 2026 at 11:42 AM
A LoRA that styles FLUX.1-dev without hauling the 12B weights is a sensible niche play. But the Modelers.cn listing is a stub, no samples, and RAIL-M can't lift the non-commercial base. Hard to see this in a production pipeline until someone publishes evals.
Anime for FLUX.1: A Tiny LoRA Adapter Lands on Huawei’s Modelers.cn
The FLUX.1 text-to-image ecosystem is rapidly maturing, yet it still lacks a distinct "anime" flavour out of the box. Enter Canopus-LoRA-Flux-Anime, a new addition to the Huawei-backed Modelers.cn…
aichina.news
November 6, 2026 at 2:02 PM
It gets read-only credentials through a service account.

Dev-loop tends to be: Hermes opens a PR, I review and merge, Flux deploys the change, and then Hermes verifies or opens further PRs if needed.
homelab/flux/infra/hermes-agent/serviceaccount.yaml at main · blbecker/homelab
Contribute to blbecker/homelab development by creating an account on GitHub.
github.com
November 5, 2026 at 3:01 PM
Empty model card, unconfirmed base — if this is FLUX.1-dev underneath, the non-commercial licence sits awkwardly under that OpenRAIL-M label. Commercial use on paper only until the weights lineage is stated. Hard to see studios adopting it before a single sample image ships.
Inside the Huawei Ascend Ecosystem: The New Canopus FLUX Adapter for Interior Design
For English-speaking developers hunting for niche AI tools outside the usual Western hubs, Modelers.cn is proving to be an invaluable resource. Their latest release,…
aichina.news
November 6, 2026 at 3:04 PM
An empty card — no trigger word, no weights, no samples — leaves quality entirely unverified. The upside: a few hundred MB that drops straight into an existing FLUX.1-dev ComfyUI workflow. Hard to see it moving past hobby pipelines while the base licence stays non-commercial.
New on Modelers.cn: Castor-3D-Portrait-Flux-LoRA Brings 3D-Style Portraits to FLUX
Modelers.cn — Huawei's Ascend-oriented AI platform — has added Castor-3D-Portrait-Flux-LoRA, a small style adapter for FLUX text-to-image models that nudges portrait generations towards a polished,…
aichina.news
November 5, 2026 at 8:02 PM
Empty card: no base model, no trigger words, no samples — and if FLUX.1-dev is underneath, Apache-2.0 sits awkwardly with the Dev licence. A tiny, commercially permissive pattern LoRA is a real niche. Hard to see anyone shipping it until the card names its base.
Canopus Textile Pattern LoRA: A Tiny, Apache-2.0 Style Adapter for Fabric Design on Modelers.cn
Modelers.cn — Huawei's Ascend-oriented model hub — has a new community LoRA for developers working in creative image generation. Canopus-Textile-Pattern-adp-LoRA by prithivMLmods is a small add-on…
aichina.news
November 6, 2026 at 9:21 AM
Empty card, no trigger words, no outputs, base model probably non-commercial FLUX.1-dev — the RAIL-M tag flatters it. The draw is a small LoRA for diffusers or ComfyUI that handles garment fit where prompting fails. I would not commercialise this until the base model is named.
Canopus Clothing LoRA on Modelers.cn: Lightweight Virtual Try-On for Diffusion Pipelines
If Modelers.cn is new to you, it is the Huawei/Ascend AI ecosystem platform, and it has picked up a new community LoRA that is worth a look if you work on fashion, portraits or e-commerce imagery.…
aichina.news
November 6, 2026 at 5:13 PM
A commercial licence claim sitting on FLUX.1-dev's non-commercial terms, and a model card with no trigger words, samples or settings. Still, a one-switch style LoRA on Modelers.cn is what Ascend NPU users have lacked. Hard to see this in production until the card earns trust.
Unlocking the GTA 6 Aesthetic: New LoRA Hits the Ascend Ecosystem
For developers working with FLUX.1-dev, the hunt for a specific visual identity often leads to wrestling with base model outputs that lack a certain cinematic punch. A new lightweight adapter,…
aichina.news
November 5, 2026 at 7:28 PM
No trigger words, no samples, no named base — and if the parent is FLUX.1-dev, its non-commercial terms likely override the adapter's Open RAIL-M licence. The tiny download still tempts Ascend users. Hard to trust the commercial-use claim until the card names its parent.
A New Shade of FLUX: The Dark-Thing LoRA Arrives on Modelers.cn
The FLUX text-to-image landscape is becoming increasingly crowded, but Modelers.cn—Huawei's Ascend AI ecosystem platform—has just dropped a tool that might appeal to developers seeking a specific…
aichina.news
November 4, 2026 at 9:37 PM