Draw Things

🔍 SeedVR2 is now officially supported in Draw Things! 🔄 Please update to the latest version v1.20260430.0 in the iOS / macOS AppStore. This version brings: 🔹Support SeedVR2 as one of the diffusion models for 1-step upscaling work.（Check the comments for SeedVR2 DT parameters.） 🔹Fix Anima LLM adapter related issues. 🔹Switch to better and newer causal / mask attention kernels. 🔹Support export LoRAs for Qwen, Z Image, FLUX.2 and ERNIE Image models. 🔹Imported Z Image models now run at BF16 precision. gRPCServerCLI and draw-things-cli both are updated to 1.20260430.0 with above related updates.

@KKar_nan check this >> x.com/drawthingsapp/… for 2511, it has built-in controlnets, u dont need control-pose tab, Just use Moodboard.

@drawthingsapp This cool. In DT this control-pose tab, any guide to use for qwen image edit 2511 model, Thanks.

🚗 Create Your Own Dream Car Poster With Draw Things. 📝 Check the Prompt Template in the ALT or Comments. 🤖 Just feed the prompt to an LLM and tell it which car you want to replace it with. 📐All the posters below were generated using Ernie Image Turbo with 8 steps in a 4:5 aspect ratio.

🆚Some Comparisons and Draw Things Configs for SeedVR2 3B/7B are provided for your reference, no prompt needed： (feel free to set the dimensions according to your needs) 📝SeedVR2 3B —————— {"faceRestoration":"","seedMode":2,"controls":[],"loras":[],"cfgZeroStar":false,"upscaler":"","width":768,"batchSize":1,"tiledDiffusion":false,"height":1152,"shift":1,"refinerModel":"","cfgZeroInitSteps":0,"maskBlur":2.5,"maskBlurOutset":0,"steps":1,"batchCount":1,"strength":1,"hiresFix":false,"preserveOriginalAfterInpaint":true,"tiledDecoding":false,"sharpness":0,"guidanceScale":1,"model":"seedvr2_3b_i8x.ckpt","causalInferencePad":0,"sampler":10,"seed":3016078078} —————— 📝SeedVR2 7B —————— {"faceRestoration":"","seedMode":2,"controls":[],"loras":[],"cfgZeroStar":false,"upscaler":"","width":768,"batchSize":1,"tiledDiffusion":false,"height":1152,"shift":1,"refinerModel":"","cfgZeroInitSteps":0,"maskBlur":2.5,"maskBlurOutset":0,"steps":1,"batchCount":1,"strength":1,"hiresFix":false,"preserveOriginalAfterInpaint":true,"tiledDecoding":false,"sharpness":0,"guidanceScale":1,"model":"seedvr2_7b_i8x.ckpt","causalInferencePad":0,"sampler":10,"seed":2757549182} ——————

🔍 SeedVR2 is now officially supported in Draw Things! 🔄 Please update to the latest version v1.20260430.0 in the iOS / macOS AppStore. This version brings: 🔹Support SeedVR2 as one of the diffusion models for 1-step upscaling work.（Check the comments for SeedVR2 DT parameters.） 🔹Fix Anima LLM adapter related issues. 🔹Switch to better and newer causal / mask attention kernels. 🔹Support export LoRAs for Qwen, Z Image, FLUX.2 and ERNIE Image models. 🔹Imported Z Image models now run at BF16 precision. gRPCServerCLI and draw-things-cli both are updated to 1.20260430.0 with above related updates.

2️⃣ Cinematic Atmospheric Location Editorial Prompt： —————— A 4:5 vertical social media poster, ultra-high resolution, 8K. Subject: A BMW M3 Touring (G81) finished in Isle of Man Green Metallic. Presented in a rear three-quarter view, emphasizing its aggressive quad exhaust system, large rear diffuser, and the signature muscular wheel arches of the M lineup. The car sits low, capturing dynamic tension as if cornering on a mountain road. License plate detail: The plate area clearly and prominently displays white text reading ‘Draw Things’. Background: A misty alpine mountain road at dawn. The winding asphalt surface shows slight motion blur, with layered dark green pine forests and drifting fog in the background, creating a cold, deep performance-oriented atmosphere. The lighting carries a cool early-morning tone, subtly outlining carbon fiber details on the vehicle. Typography layout: 1. Top headline: ‘M3 TOURING’ in a bold, slanted, sporty sans-serif font, colored matte white. 2. Subtitle: ‘THE FIRST-EVER / M POWER’. 3. Tagline: ‘Performance Without Compromise’, highlighting its all-around performance capability. 4. Decorative element: Within the misty background, the iconic blue, purple, and red ‘///M’ stripes are faintly visible. Mood: Extreme performance, driving excitement, cold mountainous environment, motorsport aesthetics, cinematic action shot, motion blur, moody lighting, ultra-realistic automotive rendering. ——————

1️⃣ High-purity, minimalist flat poster aesthetic Prompt： —————— A 4:5 vertical social media poster, ultra-high resolution, 8K. Subject: A Porsche Taycan Cross Turismo in Frozen Berry Metallic. Presented in a rear three-quarter view, emphasizing its distinctive crossover estate proportions, wide wheel arches, and full-width glass-like LED taillight strip. The lighting flows smoothly across the body, conveying the futuristic feel of a top-tier electric sports car. License plate detail: The plate area clearly and neatly displays white text reading ‘Draw Things’. Background: A dreamy, elegant gradient style. The color palette transitions softly from light lavender to misty rose, creating a high-fashion studio aesthetic reminiscent of a runway editorial. The ground features soft diffuse reflections, resulting in a clean and futuristic atmosphere. Typography layout: 1. Top headline: ‘TAYCAN’ in a modern, widely spaced sans-serif font, colored deep purple or charcoal grey. 2. Subtitle: ‘CROSS TURISMO’. 3. Tagline: ‘Porsche — Soul, Electrified.’ 4. Decorative element: A very faint ‘ELECTRIC’ text appears subtly in the background as an accent. Mood: Avant-garde fashion, electric performance, ultimate elegance, CGI cinematic rendering, fashion photography style, soft pastel palette, high-end automotive visualization. ——————

Some Comparisons👇🏻

📝 Chinese Prompts: —————— 一幅电影级黑白监控抓拍风格的画面，具有极高清晰度，带有粗糙的胶片颗粒质感以及极高对比度。画面采用高角度拍摄，模拟安装在天花板上的广角监控摄像头俯视一个熙熙攘攘的车站大厅。背景中的人群呈现出强烈的动态模糊效果，营造出一种冷峻压抑的情报监视氛围。光线来自头顶冷白色荧光灯直射，在主体的眉骨和下颌边缘投下深邃阴影，强调他抬头直视镜头的瞬间。 画面中央是一位气质冷硬的男性特工，正直视上方摄像头。他眉头紧锁，目光锐利且警觉，脸部轮廓棱角分明，下颌带有短而灰的胡茬。他头戴深灰色羊毛贝雷帽，身穿一件厚重、质感清晰的深色战术风衣，具有明显的双排扣设计，领口微微竖起。 画面叠加了典型的中情局风格战术情报界面。一组完全对称的亮绿色人脸识别方框精准地框住特工的面部，方框上方显示加粗且对齐整齐的绿色文字：‘IDENTITY CONFIRMED’。 从绿色方框右侧延伸出一条细而明亮的蓝色数据线，连接至画面右侧一个纵向排列的深灰色半透明悬浮信息面板。面板顶部是一张用于对比的证件照，画面中的人物与特工完全相同，五官一致、短发，但未戴贝雷帽，表情中性，构图干净用于识别比对。其下方以清晰无误的蓝色等宽字体显示： ‘SUBJECT: JASON BOURNE’ ‘PROGRAM: TREADSTONE’ ‘THREAT LEVEL: EXTREME’ ‘PROXIMITY: ACTIVE’ 特工的颧骨与鼻梁区域覆盖着极其精细的浅蓝色生物识别三维线框网格，紧密贴合面部结构，具有明显的空间立体感。 画面左上角显示极简白色监控坐标信息： ‘CIA_TRACKING_NODE_402’ ‘GPS: 48°51′24”N 2°21′03”E’ 画面底部中央有一个小而高亮的红色圆形录制指示灯，其右侧对齐显示红色文字 ‘REC’。在其下方，居中排列三行白色时间戳文字： ‘SATURDAY’ ‘APRIL 25, 2026’ ‘10:13:22’ 整体画面达到8K级别的超高细节表现，UI界面元素与真实物理空间自然融合。 ——————

🚨 Identity Confirmed! 🏃 You Need To Run, Now!! 👇 (Prompts in the Comments Below)

📸 One Prompt to Remove The JPEG Artifacts feel of your images. 📝 You need to put this in your “custom_prompts.txt”: 🔧 We often encounter real-world or AI-generated images with obvious compression artifacts or banding. By using Flux.2 Klein 9B KV paired with the following prompt, you can effectively remove these imperfections while keeping everything else intact.

🚨 This update brings support for several awesome new models — GO UPDATE NOW! 🔄 Draw Things v1.20260418.1 was released on the iOS / macOS App Store a few hours ago: 1️⃣ Support for ERNIE Image, including LoRA import, training, and model import. @ErnieforDevs 2️⃣ Support for Anima Preview 3, including LoRA import, training, and model import. 3️⃣ Updated to LTX-2.3 [distilled] 1.1. 4️⃣ Fixed a crash related to Z-Image LoRA training. 5️⃣ Experimental ANE + NAX hybrid inference. gRPCServerCLI and draw-things-cli have both been updated to 1.20260418.1 with the above changes.

@Inzeal1 For M3 Base, M3 Pro, yes. Probably not going to be faster on M3 Max. Will be slower on M3 Ultra.

@drawthingsapp Is 8-bit S recommended on M3 chips?

⤴️Recently, we’ve been mentioning a few KEY TERMS a lot — These are what unlock massive performance gains on Apple devices, especially on newer chips: 🔷ANE (Apple Neural Engine): A dedicated piece of silicon from Apple, designed specifically for neural network workloads. 🔷NAX (Neural Accelerator): Apple’s equivalent of Tensor Cores — small dedicated units embedded inside each GPU core, highly optimized for matrix multiplication. These are only available on the latest chips, starting from M5 and A19 Pro. 🔷8-bit S: An int8 quantization format developed by Draw Things. S stands for fine-grained scaling. It is designed to retain generation quality while using int8 matrix multiplication instructions as much as possible. (This is orthogonal to MQA). 🔷MQA (Metal Quantized Attention): Significantly boosting performance on Apple Silicon when NAX is available.

⚡️The future of on-device AI——Faster. Cooler. More Efficient. 🧠By using ANE as a fine-grained accelerator for 8-bit S models, Draw Things now hits up to 1.8x speed-up on M4. 💎High throughput, lower heat, and total control over memory. 👇🏻 Dive Deeper Into Our Implementation!

🫵 One thing where Flux.2 Klein 9B KV can't replace Klein 9B： 🧪 Based on Draw Things testing, Klein 9B KV struggles with Face Swap tasks, while Klein 9B nailed it in one go. 🍀 Even better, Klein 9B delivers much more impressive results when paired with BFS LoRA.

🎨 What did you create with the Draw Things App today? 🛠️ And what have you built using Draw Things CLI and MediaGenerationKit based on the Draw Things inference stack? ✨ Looking forward to your shares! 💪 Our goal remains the same: Keep Shipping Solid Stuffs!

@jn1ayu Thanks for checking out! We probably won't work on improving SDXL variants performance. You will have better chance migrate to models such as Z Image / Flux.2 [Klein] or Anima for your projects!

@drawthingsapp 先ほどリリースされたバージョンを、Mac mini M4(16GB/256GB)で動かしたところ、ZImageTurbo系なら概ね半分の時間で画像が生成されるようになりました。一方で、Illustrious系では生成開始直後にアプリケーションが落ちる例があります。

🚀 We will get solid gains from ANE — Stay Tuned!

💻Guys with M4-series Macs — Download the 8-bit S models now and run them in Draw Things, you’re in for a nice surprise! 🆕A few hours ago, Draw Things v1.20260410.1 was released on the iOS / macOS App Store. This version brings: 1️⃣8-bit S now will be meaningfully faster on M4 below M4 Max due to use of Apple Neural Engine on these devices. 2️⃣Add LoRA training support for FLUX.2 [klein] 4B / 9B, Z Image and Qwen Image. 3️⃣Fix audio sync issue during in-app video playback. Fix some 8-bit S model generation issue on M5. 4️⃣FWIW, we bumped Draw Things+ LoRA Storage to 20GiB. ✅gRPCServerCLI is updated to 1.20260410.1 with relevant fixes above. ✅draw-things-cli is updated to 1.20260410.1.