🚨 Qwen Image 2.1, the compact, efficient, and unified image creation model, is now officially supported in Draw Things! You need this amazing model! 🔄 Draw Things v26.0924.0 was released on the iOS / macOS App Store one hour ago. This version brings: 🔹 Added Qwen Image 2.1 support, including LoRA import. A 4-bit S version is also now available. 🔹 Added Sol Attention support to speed up MiniMax H3. 🔹 Some minor UI fixes.

💡 A few things you should know about MiniMax H3 in Draw Things! 🔷 MiniMax H3 is the first model in Draw Things to support first-and-last-frame generation. The Canvas is used as the first frame, while you can place an image in the Moodboard as the last frame. It supports T2VA / I2VA / FL2VA. 🔷 MiniMax-H3 [ref2va] can be used for reference-based generation. Provide reference images through the Canvas and Moodboard, and identify them in the prompt as , , etc. 🔷 Currently, Draw Things does not support using Audio as a reference input for video generation, nor does it support L2VA (last-frame instruction) generation.

We made something new!

🚨 MiniMax H3 is RIGHT HERE in Draw Things！Go try it NOW！ 🔄 Draw Things v26.0910.1 was released in the iOS / macOS AppStore 2 hours ago. This version brings: 🔹 Support MiniMax H3 series models, including LoRAs and TeaCache; 🔹 Support importing Krea 2 series models. 🔹 Fix some performance issues on M4 Apple Neural Engine. gRPServerCLI and draw-things-cli both are updated to 26.0910.1 with above related updates.

✨ Love this editorial illustration look? Here’s how to transform any photo using Draw Things: 🎛️ Setup: Model: Flux.2 Klein 9B Mode: Edit Mode 📝 Prompt: —————— Based on Picture 1, transform the entire image into a high-end minimalist editorial illustration while preserving the original subject’s identity, pose, silhouette, proportions, composition, action, and core visual relationships. Preserve the original scene and recognizable subject, but reinterpret the photographic image as a sophisticated black-and-white illustration inspired by an old children’s book reimagined by a contemporary high-end design studio. Simplify the subject into rounded, solid, slightly naive generalized forms while keeping it immediately recognizable. Do not mechanically reproduce photographic details; instead, extract the most distinctive contours, gestures, interactions, and visual characteristics from Picture 1. Use an almost pure white paper background with dense black wax crayon, charcoal pencil, and rough dry-media marks. Black areas should be created from many short, energetic, imperfect strokes rather than smooth vector shapes, with visible paper texture, broken marks, rough edges, uneven tonal density, small white gaps, and slightly trembling incomplete contours. Use expressive negative space by carving irregular white areas into black silhouettes to emphasize the most recognizable features of the subject. Simplify environmental elements into a few rough handwritten-like marks and minimal symbolic lines while preserving the essential spatial context of the original image. Keep generous clean white space and let the original pose, action, distance, and relationships carry the visual story. The overall aesthetic should feel warm, quiet, humorous, slightly lonely, childlike yet sophisticated, with strong contemporary art direction and editorial design sensibility. Maintain a tactile handmade quality, balancing rough black marks, clean white space, simplified forms, and expressive paper texture. The final image should look like a carefully designed collectible art-book illustration rather than a conventional commercial cartoon. Absolutely no text, no letters, no numbers, no typography, no logos, no captions, no watermark, no split screen, no duplicated subject, no color, no realistic photographic rendering, no smooth digital vector lines, no 3D effects, no gradients, no glossy surfaces, and no generic sticker or cartoon aesthetics. ——————

4.3x is an understatement for this chip. Will do more fun things once it is generally available!

🔧 The 「Shift」 Value in Draw Things Settings Explained! 🧪 We’ll use the excellent Ernie Image Turbo model to test and explain this. Shift controls how much the image is biased toward higher noise levels. You can think of it this way: high noise controls the overall structure and composition of the image, while low noise controls the details and textures. Increasing the Shift value makes the image structure more stable and less prone to errors. A relatively lower Shift value can produce better details, but if the Shift is too low, the overall structure can start to break down. 🔍 From the examples below, we can clearly see that with Shift set to 0.66 and 1, the girl’s fingers and the phone camera are already broken. Even when Shift is increased to 2.2, the character still has six fingers, and the image has a kind of compression artifacts look. 「4」 is the value officially recommended by Ernie Image Turbo. When we increase Shift to 6 or even 8, the image becomes extremely stable, clean, and smooth overall, but it loses a little bit of realism. ⚙️ So next time you use Ernie Image Turbo, if you want more realism and detail, you can set Shift to around 3–4. If you need more stable structures and a cleaner, smoother image quality, you can increase it to around 6.