Video walkthrough · 20:58
DiffusionBear Video Tutorial: Local AI Art Studio on Mac
Watch the complete DiffusionBear workflow on Apple Silicon: find a Civitai LoRA, generate with Juggernaut XL Lightning and FLUX.2-klein 9B, refine prompts with the on-device Qwen 1.5B enhancer, upscale 2×/4×, and submit to the Civitai Daily Challenge, all offline on a 16 GB M1 Mac.
Privacy: this page does not contact YouTube until you press Play. The player uses the privacy-enhanced youtube-nocookie domain; YouTube may set cookies once playback starts. YouTube privacy policy.
Watch on YouTube →
What you will learn in 20 minutes
The walkthrough follows a complete real project: pick a Daily Challenge concept, install the matching Civitai LoRA, compare models, refine the prompt, upscale the result and publish it with embedded generation metadata.
Register a LoRA in one click
Browse Civitai, download the trigger words and register a LoRA without leaving the studio.
Model choiceCompare 20 local configurations
See why Juggernaut XL Lightning is the fast SDXL option and where FLUX.2-klein 9B wins on detail.
MemoryRun FLUX.2 on 16 GB
Understand the 4-bit MLX, fast-VAE and memory-budget choices that keep a 9B model usable.
EditingEdit with FIBO 2.0
Use instruction-guided edits, multi-reference images and masked inpainting after the tutorial render.
DistillationTry the MOSS DMD pipeline
See how the 4-step and 8-step DMD adapters turn a 25-step engine into a fast local workflow.
MigrationBring your DiffusionBee models
Import .tdict LoRAs and existing checkpoints with a read-only migration path.
Chapters
Jump directly to the part you need. Timestamps open the video on YouTube.
- Introduction and model architecture overview
- Browsing Civitai and selecting the Daily Challenge LoRA
- Universal LoRA downloader and one-click registration
- Prompting with trigger words and SDXL Lightning setup
- Advanced controls: samplers, guidance scale and DeepCache
- Local Qwen 1.5B prompt enhancer, on-device LLM
- Post-processing: 2× and 4× high-resolution upscaling
- Reference image conditioning with the DiffusionBear logo
- Open-source MIT licensing and the local tech stack
- Final workstation render and outro
How the local workflow works
DiffusionBear is a native macOS studio built on Apple's MLX and Metal stack. Every model shown in the video runs on the Mac's unified memory; there is no cloud worker, no queue and no per-image billing. The reference machine is a 16 GB M1 MacBook Pro, and the default 4-bit quantisation keeps the heavier 9B and 13B engines inside that budget.
The studio handles the full loop: model and LoRA discovery from Hugging Face and Civitai, prompt crafting with a local Qwen 0.5B/1.5B enhancer, generation, gallery metadata, 2×/4× Lanczos upscaling, reference-image conditioning and drag-and-drop export to Civitai with Automatic1111-compatible parameters. The source is MIT-licensed, and the benchmarks are published with reproducible prompts and seeds.
Want the numbers behind the video? Start with the 220-run benchmark arena, then read the FLUX on 16 GB guide for the memory model. If you are switching from DiffusionBee, the migration guide covers .tdict LoRAs and read-only import.
Frequently asked questions
What is DiffusionBear?
DiffusionBear is a free, MIT-licensed local AI image studio for Apple Silicon Macs. It runs FLUX.2, FLUX.3, SDXL, FIBO 2.0, MOSS-Image, Krea 2, Qwen-Image and Z-Image models with 4-bit MLX quantization, entirely offline.
Is DiffusionBear really free and offline?
Yes. There is no account, no cloud generation and no per-image fee. After models are downloaded, generation runs locally on the Mac's Metal GPU. The source code is MIT-licensed on GitHub.
Which Macs can run the tutorial workflow?
Any Apple Silicon Mac from M1 onward. The reference machine is a 16 GB M1 MacBook Pro. Heavier models are tuned for 16 GB, and larger configurations are available on higher-memory machines.
Can I install Civitai LoRAs in DiffusionBear?
Yes. The studio includes a Civitai and Hugging Face browser, a one-click LoRA downloader, automatic key conversion for SDXL and FLUX, and read-only import of LoRA folders already used by Draw Things and DiffusionBee. See the Pony V6 XL and Civitai guide.
Does the tutorial cover FLUX.2 and image editing?
It covers FLUX.2-klein 9B generation, Juggernaut XL Lightning, the local Qwen prompt enhancer, gallery filtering, 2×/4× upscaling and reference-image conditioning. Separate pages cover FIBO 2.0 instruction editing and the MOSS-Image DMD pipeline.
Continue with DiffusionBear
Ready to try it on your Mac?
Download the standalone Apple Silicon app, run the same workflow locally, and keep every image and prompt on your own machine.
Download DiffusionBear v0.5.0 Installation & Gatekeeper guide