MOSS-Image is now open source: a 5B DiT with Chinese & English text renderingn. 图中文字渲染 and controllable layoutsn. 可控布局
MOSS-Image (Base / SFT)OpenMOSS·
MOSS-Image is now open source on Hugging Face. Built on a 5B Diffusion Transformern. 扩散 Transformer（DiT 架构） (DiT), MOSS-Image brings Chinese & English text rendering and controllable layouts to image generation — from posters and infographics to creative visuals (OpenMOSS)
Long prompts & text rendering — up to 4K-token prompts, including multi-line text
Experimental JSON layout control — specify text, objects, and their positions using bounding boxesn. 边界框 (a `summary`/`items` layout pasted as the prompt); editing text, element descriptions and boxes also enables controlled editing and layout tools
Flexible generation — multiple aspect ratios, resolutions, and customizable inference settings; takes natural-language prompts, tag sequencesn. 标签序列（如 Danbooru 标签） and JSON layouts. Repo highlights: high-quality anime and photorealistic styles, easy to fine-tune, compact and efficient 5B backbone, trained on heterogeneousadj. 异构的（多种硬件混合） / alternative GPU platforms
Two checkpointsn. 模型权重检查点 released together: MOSS-Image-Base (base text-to-image model) and MOSS-Image-SFT (fine-tuned from Base on curated data; clearer text grouping and more organized layouts). Model weights, inference coden. 推理代码 and a technical report are public
License: Apache-2.0 for the model weights and bundled inference code (Hugging Face card)
Developer workflows — Python (Diffusers), Gradio, a native ComfyUI plugin (install via ComfyUI Manager; integration into official ComfyUI in progress) and a Prompt Enhancer Skill for agents; Day-0 supportn. 发布当天即支持 from SGLang, vLLM and ModelScope; also on ModelScope and a Hugging Face Spaces demo
Default sampling: flow-matching Euler, 50 steps, CFG 3, shift 5, negative-branch layer skipping at layer 22; bundled VAE uses the FLUX.2 small decoder. A commented-out repo note mentions a 42-layer DiT with a Qwen3.5-9B text encoder (未经官方确认)
Community (社区, 未经官方确认): Ostris AI Toolkit added MOSS Image SFT and Base training the same day (plus training it as an edit model); kohya says sd-scripts support is coming; a WIP 8-step DMD distilln. 蒸馏版（少步数加速模型） by lodestones appeared; Japanese testers describe it as "Krea 2 with Danbooru tags and JSON layout"
Weights: huggingface.co/OpenMOSS-Team/MOSS-Image-Base and MOSS-Image-SFT · Code: github.com/OpenMOSS/MOSS-Image
Krea 3: "a new frontier for image editing" (beta opens to everyone)
Krea 3 (Beta)Krea·
today we open the beta for Krea 3, a new frontier for image editing — "Krea 3 is now available for everyone" at krea.ai/krea-3
End-to-endadj. 端到端（从提交到出图全程） 2K edits in under 8s (native 2K); matches Nano Banana Pro or GPT Image 2.5 quality (Krea's claim); bounding boxn. 边界框 control and precise multi-turn editsn. 多轮编辑
Pricing: $0.06 per image at 2K and $0.03 at 1K. Krea's comparison (median edit time on Krea, Sep 24–Oct 1, queue included; list prices): Krea 3 under 8 s / $0.06 vs Nano Banana Pro 33 s / $0.135 vs GPT Image 2 90 s / $0.22; 100,000 edits a month ≈ $6,000 vs $13,500 vs $22,000
Keeps what you didn't ask to change: products stay intact down to the label and the logo, the same person/character stays consistent across images, and text is preserved; works from up to 6 reference images per edit (blog; the API accepts up to 8 reference URLs)
Follows detailed briefs (e.g. a line drawing + a six-color palette → a finished room with the camera and architecture unchanged), keeps paintings, vector art and ink drawings in their own medium ("Krea 2's eye for style"), turns sketches/moodboardsn. 情绪板（风格参考拼贴） into photos; also handles text-to-image
Availability: live today on krea.ai for 30 million users, and through fal, Hedra, Together AI, Nous Research, OpenRouter and Replicate. Krea API is in early accessn. 抢先体验 for Krea 3 beta workspaces (contact sales@krea.ai): `fast` / `high` thinking modes, regional editsn. 局部区域编辑 that keep the rest pixel-exactadj. 像素级不变的, 1K/2K output, Zero Data Retentionn. 零数据留存 for enterprise
"final release will come next week, together with more benchmarks." LoRAs and srefs are ready to experiment with (early access by DM)
What's next: not real time yet (fastest setup ≈ 4–5 s per edit); more resolution for less compute; understanding circled or written-on-image instructions
FLUX 3 Image Klein is here: open weightsn. 开放权重（可下载的模型参数）, 5.6B parameters
FLUX 3 Image KleinBlack Forest Labs·
FLUX 3 Image Klein is here. Open Weights with the quality of a large frontier model. Customizable and fast.
5.6B parameters and runs on a single consumer GPU. One model for text-to-image, image editing and multi-reference generation with up to 10 reference images (Hugging Face card: Qwen3-VL-4B-Instruct as the text encoder)
Retains the larger FLUX 3 Image model's native 4K generation, bounding boxesn. 边界框 (place elements or confine edits to an area) and pixel-perfect editingn. 像素级精确编辑; also renders text in images
What's launching: Base open weights (full checkpoint for fine-tuningn. 微调, further training and research); Distilled open weights (step-distilledadj. 步数蒸馏的（少步数快速出图）, low-latency, for faster generation and self-hosting); Quantizedadj. 量化的（降低精度以省显存） FP8, NVFP4 and INT8 builds, small enough for GPUs and laptops with 8 GB of VRAM; the BFL API; and launch support for Ostris's AI Toolkit, ComfyUI, Diffusers, SGLang Diffusion, vLLM-Omni and TensorRT-LLM, plus NVIDIA fine-tuning recipes (NeMo AutoModel)
Speed (trained on NVIDIA GB200 NVL72, optimized with NVIDIA): 0.55s per image on GB300 (NVFP4), ~4s on GeForce RTX 5090 (INT8) and under 9 s on RTX 4070 for a 1MP image in ComfyUI (distilled, 16 steps); generator weights 11.2 GB in BF16 / 5.6 GB in FP8; serving a LoRA adds about 5%
Internal evals: scores higher than Qwen Image 2.1 (the current open weights leader) in text-to-image and multi-reference generation while costing less; ranks first among the compared open models in T2I, multi-reference and editing, and also above closed Ideogram V4.5 and FLUX.2 [pro]; vs FLUX.2 [klein] 9B: +91 Elo in text-to-image and +70 in image editing with 38% fewer parameters; on the Pareto frontiern. 帕累托前沿（同等速度下质量最优的边界） for quality and speed, with lower API latency than Qwen-Image 2.1, FLUX.2 [pro], Ideogram V4.5 and Krea 2
BFL API: the distilled model as `flux-3-image-klein`, with prompt upsamplingn. 提示词自动扩写 and pixel restoration (keeps unedited regions exact); pricing starts at $0.009 per 1K image. Also in the Playground, on Hugging Face Spaces and on OpenRouter
License: FLUX Kommunity License — free commercial use for individuals and companies with annual revenue under $5M, plus research and non-commercial use; larger organizations deploy via self-hosted commercial licenses or the API. Enterprise: custom fine-tuning on proprietary datan. 自有/专有数据, deployment support, and commercial rights for the larger FLUX 3 Image weights
Second open-weights release of the FLUX 3 family after FLUX 3 Action ("More will follow"). Robin Rombach (BFL): "It literally runs on everything and is ideally suited for local development"
Community first impressions (社区, 未经官方确认): fast and light locally (~3 s for 1024² on an RTX 5090), editing and multi-reference praised, but some testers find text-to-image skin/photo quality weaker than Krea 2 or Qwen-Image 2.1 + LoRA
Weights: huggingface.co/black-forest-labs/flux-3-image-klein (distilled) and flux-3-image-klein-base · Blog: bfl.ai/blog/flux-3-image-klein
Qwen-Image-2.1-Turbo: create and edit images in just 8 denoising stepsn. 去噪步数（扩散模型的迭代次数）
Qwen-Image-2.1-TurboAlibaba Qwen·
Meet Qwen-Image-2.1-Turbo: create and edit images in just 8 denoising steps. Open weights now available!
Built on Qwen-Image-2.1, Turbo is an accelerated checkpointn. 加速版模型权重 on the same 7B visual generation architecture
Fewer steps does not mean lower quality: it still generates strong 2K images from text, and supports continued creation through natural-language edits (from adding accessories to changing a scene)
Start directly with Diffusers: load QwenImage21Pipeline; the checkpoint's recommended 8-step sampling schedulen. 采样调度 is ready to go
Qwen-Image-2.1 Pro and Turbo APIs are now officially live (Alibaba Cloud Model Studio). Choose a hosted APIn. 托管 API or use the Turbo weights in your own workflow
Weights: ModelScope (modelscope.cn/models/Qwen/Qwen-Image-2.1-Turbo) and Hugging Face (huggingface.co/Qwen/Qwen-Image-2.1-Turbo)
Oct 10 (Qwen Cloud): Qwen-Image-2.1 is now live on Qwen Cloud — in both Pro and Turbo. A unified model for both generation and editing, delivering top-tier quality in a lightweight package
Compact & exceptionally fast: a lightweight 7B architecture that outperforms most closed-source models, with drastically accelerated inference for multi-image inputsn. 多图输入
Native transparency: natively generates and edits RGBA layers, enabling seamless compositingn. 图像合成 and text editing within transparent images
Versatile, high-fidelity editing: supports up to 10 reference images and precise local control while preserving strict fidelity for portraits and products
Broad coverage & stunning aesthetics: excels at panoramas, infographics, and virtual try-onsn. 虚拟试穿, delivering realistic textures and elegant typography
Turbo takes it further: create and edit images in just 8 denoising steps, same 7B architecture, "same quality, dramatically faster"; try Turbo and Pro on qwencloud.com
Alibaba Cloud: Qwen-Image-2.1-Turbo is the fastest, most affordable model in the Qwen-Image family (an accelerated version of Qwen-Image-2.1-Pro): 1 image in 10–20s, $0.016 / image; API via Model Studio and Qwen Cloud
X news summary (未经官方确认): up to 2048×2048, Qwen Research License (non-commercialadj. 非商用的), community GGUFs of 2.6–16.4 GB
Community local test (未经官方确认): on dual 5090 + ComfyUI, about 22% faster on 1024² text-to-image with nearly identical VRAMn. 显存; the standard version stays more stable for small text and complex edits
Midjourney tests a "thinking mode" and a Midjourney MCPn. Model Context Protocol，让 AI 调用外部工具的协议
Midjourney thinking mode (alpha test)Midjourney·
Oct 8: We're testing a new "thinking mode" for our image generation on our Alpha website (alpha.midjourney.com)
We're finding it boosts prompt accuracyn. 提示词准确度, typographyn. 文字排版 and coherencen. 连贯性、一致性. Help test it on your images and share feedback
A test feature, not a new model version; the current default is still V8.2 (per Midjourney docs)
Oct 10: We want to start testing a Midjourney MCP with a limited community: creative and technical people to push the boundariesv. 突破边界 of what our models can do
Not for Midjourney in SaaS products, only for purely artistic and creative endeavorsn. 尝试、事业; apply via the link in the post
Oct 11 community test (未经官方确认): Midjourney 8.2 "Thinking" mode is very impressive; the result isn't perfect but it makes a huge difference, vs GPT Image 2.5 Sunburst's weird texturesn. 怪异纹理 and noise on the same tricky windstorm prompt; a thinking batch takes about as long as an HD batch
Nano Banana 2.1: Google's latest image generation and editing model
Nano Banana 2.1Google·
Meet Nano Banana 2.1, our latest image generation and editing model
This upgraded version outperformsv. 胜过、优于 our previous models across the boardadv. 全面地、各方面
Notable leaps in visual design, mask-based editingn. 基于蒙版的局部编辑, and subject consistencyn. 主体一致性 to help you create more natural-looking images
Logan Kilpatrick (Google): higher quality images, lots of bug fixes from previous models, and a new lower price pointn. 价格档位 (no figure given in the post)
Alisa Fortin (Google AI Studio): model id gemini-nano-banana-2.1; better at instruction followingn. 指令遵循 and handling complex prompts; improved subject consistency across turns
Starts rolling outv. 逐步推送上线 today (Oct 6) across Gemini App, AI Mode in Search, Google AI Studio, Gemini API, Flow, Stitch, Google Ads, and Gemini Enterprise Platform
X news summary (未经官方确认): built on Gemini 3.6 Flash, about $0.0336 per 1K image (half the predecessor), up to 4K, up to 14 reference images; seen in Flow around Oct 5 under the codenamen. 代号 "beluga"
Make precise multi-turn editsn. 多轮编辑 without changing any other pixel
Lay out the image exactly how you want using bounding boxesn. 边界框（框定位置的矩形）
Generate in up to 4K to preserve details
Use up to 10 referencesn. 参考图 to compose an image
Commercial Weightsn. 商用授权权重 available for companies running image generation at scale
Open Weightsn. 开放权重 version of FLUX 3 Image is launching in the coming weeks (no date given)
Community reports (未经官方确认): built on the multimodal FLUX 3 backbonen. 主干模型 introduced in July; live on partner platforms incl. Higgsfield, Krea, OpenRouter, Design Arena; third-party pricing quotes of about $0.048 per 1K image / $0.607 per 4K image
Oct 6 (BFL): FLUX 3 is #1 on Physics-IQ, Google DeepMind's benchmark for physical world understanding in video models; claims to substantially outperformv. 优于、胜过 Seedance 2.5, MiniMax H3, Gemini Omni 1.1 Flash, Veo 3.1, Sora 2 and Cosmos3 in most cases; the base model leads all foundation models, and FLUX 3 with a physics verification layern. 物理校验层 scores higher still. The backbone also powers FLUX 3 Action (robotics) and FLUX 3 Video (mainly a video result)
Oct 15: the open-weights member arrived as FLUX 3 Image Klein (5.6B, FLUX Kommunity License) — see its own item
Introducing Ideogram 4.5, the most precise edit model
With each edit, leading models add artifactsn. 伪影（生成/编辑带来的瑕疵）, pixel shiftsn. 像素偏移, and color changes. Ideogram 4.5 eliminates artifact buildupn. 伪影累积, making multi-turn editingn. 多轮编辑 possible
Also supports Text to Image
Live in Ideogram, the API, and launch partnersn. 首发合作伙伴
Open weightsn. 开放权重（可下载的模型参数） soon (no date given)
Artificial Analysis (Oct 3, evaluated at "high" quality): #23 on AA-Image-Editing v2.0 (Ideogram's first model on the board; just behind HiDream-O1-Edit-1.5 and HunyuanImage 3.0 Instruct, ahead of Seedream 5.0 Lite) and #34 on AA-Image-T2I v2.0 (up from #40 for Ideogram 4.0)
Pricing per Artificial Analysis (未经官方确认): High quality $0.22/image ($220 per 1k) for edits and $0.10/image ($100 per 1k) for Text to Image; lower tiersn. 档位 start at $0.008/image for edits and $0.03/image for Text to Image
Oct 9: Ideogram reshared the updated Artificial Analysis leaderboardn. 排行榜: "Ideogram 4.5 is the most precise edit model"
Artificial Analysis multi-turn test (Oct 9, 社区): GPT Image 2.5 Sunburst, Ideogram 4.5, FLUX 3 and Nano Banana 2.1 each made 30 consecutive edits on the same photo (a real estate stagingn. 房产样板布置（虚拟软装） sequence: light the fire, add a sofa, repaint the walls, swap day for twilight, ...). Ideogram 4.5 keeps most of the room intact, while the others driftv. 漂移、逐渐走样, GPT Image 2.5 Sunburst most visibly. On small edits, Ideogram 4.5 and FLUX 3 left 95% or more of the image essentially untouched; GPT Image 2.5 Sunburst re-rendersv. 重新渲染 most of the image on every edit, leaving only about a fifth unchanged; Nano Banana 2.1 sits in between (edits stay local, but the rest of the image shifts slightly and gradually darkens)
Oct 15 (Ideogram): Ideogram 4.5 very low pushes the Pareto frontiern. 帕累托前沿 for image editing on both price and speed, according to @DesignArena — every edit model rated higher costs more and runs slower; no artifact buildup, edit after edit. Native 2K. 0.8¢ per image. 8s per edit.
First image model from Meta Superintelligence Labs: agenticadj. 智能体式的（可自主调用工具） generation, precise edits, multi-reference composition; #2 on Image Arena at launch