Draw Things App: 2025 Feature Digest & Release Guide
June 5, 2025
A clear, practical summary of all the major 2025 Draw Things app updates—explaining every new feature, supported model, and technical term in simple language for creative workflows.
Executive Takeaway
Draw Things evolved from a free, offline image generator into a hybrid image + video studio supporting the latest open-source models (Hunyuan-Video, Wan 2.1, HiDream, FLUX Redux). You can now off-load heavy jobs to a secure cloud—without losing the local control that artists love. Recent updates bring first-class video tools, new models, RAM-friendly features, and smoother workflows for LoRA, ControlNet, and batch management.
What Draw Things Can Do Now
| Pillar | 2024 Baseline | 2025 Upgrades & Benefits |
|---|---|---|
| Generative Core | SDXL-class image diffusion | Adds video diffusion with Hunyuan, Wan 2.1 & SkyReels—480p/720p text-to-video, inpainting |
| Model Zoo | FLUX.1 image model | FLUX fixes, Flex.1 Redux, HiDream, Chroma, Transparent-Image LoRA, 8-bit conversion |
| Performance | fp16-only inference | SVDQuant 6-bit models, TeaCache (up to 4× faster DiT), CPU/GPU off-load, Universal Weights Cache |
| Remote Compute | Experimental gRPC server | Cloud Compute toggle + shared-secret auth, local control, remote power |
| Editing Tools | ControlNet, img2img rails | New Control panel: merge ControlNets, tiled overlap fix, zoom-in img2img, batch video |
| UX & Language | English-only UI | Japanese, Korean, French, German support, onboarding improvements |
| Export | PNG/JPG sequences | ProRes 4444 video export, codec selector, batch video history preview & save |
Release-by-Release Highlights
Feb–Mar 2025
- Hunyuan video model: Best-in-class text-to-video, LoRA support for video, seamless tiling & inpainting, bugfixes for model imports.
- Cloud Compute: Offload jobs to Linux servers with all model settings available.
- FLUX.1 & ESRGAN improvements: Small quality gains, upscaling is up to 2× faster.
Mar–Apr 2025
- Wan 2.1 models: Fast, high-quality text-to-video and image-to-video at 480p/720p; 6-bit quantization means less RAM, faster loads.
- TeaCache: Up to 4× faster generation for DiT models, now available for HiDream, Wan, FLUX, Hunyuan.
Apr–May 2025
- HiDream series: State-of-the-art image generation, multi-language tokenization, CJK character fixes, transparent PNG LoRA support.
- Universal Weights Cache: Slashes RAM requirements on 8–18GB Macs/iPads.
- New codecs: ProRes 4444 as default for video export on macOS, codec selection in settings.
- Batch & video management: Videos grouped in history, right-click/long-press for fast export.
Late May 2025
- Chroma models, BYO-LoRA in cloud: Bring your own LoRA for cloud jobs, private storage, up to 10GB.
- Causal Inference for Wan 2.1: Improved video realism for cause/effect scenes.
- General speed & stability: 2–5% speedup on DiT/CUDA; more robust chunked gRPC for large jobs.
Key Acronyms, Explained
| Acronym | Meaning | Role in Draw Things |
|---|---|---|
| gRPC | Google Remote Procedure Call | Offloads jobs to remote server, with full local-like control |
| LoRA | Low-Rank Adaptation (style/character fine-tunes) | Lets you change styles or subjects with small adapter files |
| ESRGAN | Enhanced Super-Resolution GAN | High-quality, fast image/video upscaling |
| VAE | Variational Auto-Encoder | Compresses images for faster, better generation |
| CFG | Classifier-Free Guidance | Directs models to stick closer to your prompt |
| SVDQuant | Singular Value Decomposition Quantization | Allows big models to run with less RAM (6-bit weights) |
| TeaCache | Timestep Embedding Aware Cache | Skips redundant compute for up to 4× faster generations |
| T2V/I2V | Text-to-Video / Image-to-Video | Animate from text, or from your own still images |
| DiT | Diffusion Transformer backbone | Newer, more powerful models use this architecture |
| SigLIP | Sigmoid-loss Language-Image Pre-training | Advanced vision-text encoder (better prompt/keyword handling) |
| LLaVA | Large Language-and-Vision Assistant | Lets you guide with text and images |
| ProRes 4444 | Apple’s high-fidelity video codec | Default for Mac video export, perfect for professional workflows |
Practical Tips for Creatives
- Use Cloud Compute on low-RAM Macs/iPads—no settings lost, remote power gained.
- Enable TeaCache for fast generation (especially on big models).
- Try 8-bit model conversion on heavy checkpoints to cut RAM in half.
- Group videos and export them in the history panel—ProRes and PNG supported.
- HiDream/Wan 2.1: best for multilingual and CJK prompts after recent patches.
Why These Updates Matter
Together, these features flatten hardware barriers, open cinematic workflows, and cut creative wait times—turning Draw Things from an “offline Stable Diffusion toy” into a production-ready local + cloud studio for stills and motion.
For more technical details or the full changelog, see Draw Things Wiki.