Legacy SD 1.5 (DiffusionBee)
CLIP Tag Soup + Negative Prompt
watchmaker, old man, magnifying loupe, brass gears, masterpiece, best quality, 8k, sharp focus
Negative: blurry, bad hands, extra fingers, watermark
Migration & Studio Workflow Playbook
DiffusionBee's last release was v2.5.3 in August 2024 (which added FLUX.1 support for macOS); while there has been no official announcement of its discontinuation, there have been no updates since August 2024. Here is how to move your workflow into the DiffusionBear v0.5.0 studio (615 MB pure-MLX PyTorch-free bundle, 5 core studio workspaces plus 5 specialized sub-views): convert your existing .tdict LoRAs read-only, upgrade to 4-bit Apple MLX engines across M1–M4+ / all M-series chips, and adapt your prompts with 9 canonical family-specific Qwen 2.5 profiles in one click.
~/.diffusionbee folder & convert LoRAsDiffusionBee stores all downloaded and imported weights inside ~/.diffusionbee/downloaded_assets and ~/.diffusionbee/imported_models using a custom .tdict binary container. Over time, old 4 GB to 6 GB checkpoints quietly consume 30 GB to 80 GB of macOS SSD space.
.tdict file without copying or modifying your folder, classifies each file by tensor structure, and shows which LoRAs are ready to import..safetensors file into its own LoRA library..tdict base checkpoints or legacy SD 1.5 files, DiffusionBear displays a direct recommendation chip (→ Juggernaut XL Lightning, → Pony Diffusion V6 XL, or → FLUX.2-klein 4B) so you can grab the 4-bit MLX equivalent in one click.du -sh ~/.diffusionbee/* 2>/dev/null to see how much SSD space your old .tdict checkpoints occupy.
Every model below runs natively on Apple MLX and Metal within a 16 GB unified-memory budget (measured on a base 16 GB M1 Mac at 512×768).
| Old DiffusionBee Model / Task | DiffusionBear MLX Replacement | M1 16 GB Latency | Why it is a major upgrade |
|---|---|---|---|
| SD 1.5 Realistic Vision / Photon | Juggernaut XL Lightning (4s) | 8.80s (512×768 median) | Native 1024-class SDXL anatomy and lighting in 4 steps via 4-bit UNet quantization. |
| SDXL 1.0 Base (25–30 steps) | RealVisXL V5.0 / Juggernaut XI (8s) | 38.78s – 41.22s | Hyper-SD 8-step distillation delivers studio skin texture 3× faster than unquantized SDXL. |
| General Creative & Multi-Reference | FLUX.2 Klein 4B (4-bit MLX) | 44.09s | Rectified-flow DiT prompt adherence with up to 10 reference images and FLUX.2 LoRAs. |
| Inpainting & Outpainting | Bria FIBO 2.0 Edit Turbo (4-bit) | ~45.0s | Instruction-guided editing ("change jacket to red"), 1–4 references, and unmasked pixels preserved. |
| ControlNet / Multi-Subject Layout | FLUX.3 Klein 5.6B / MOSS-Image 5B | 54.8s (FLUX.3) | Draw interactive Virtual Boxes right on the canvas with zero ControlNet VRAM overhead. |
| DreamShaper / Anime Checkpoints | Pony Diffusion V6 XL / Illustrious | 110.6s (25 steps) | Auto-converts single-file .safetensors checkpoints with adjustable CLIP Skip and Civitai LoRAs. |
The #1 surprise for users leaving DiffusionBee is that old Stable Diffusion 1.5 prompts like masterpiece, best quality, 8k, photorealistic, 1girl, cyberpunk, (neon lights:1.3) produce flat or overly literal images on modern Diffusion Transformers (FLUX.2, FLUX.3, MOSS-Image, Bria FIBO 2.0).
Legacy SD 1.5 (DiffusionBee)
watchmaker, old man, magnifying loupe, brass gears, masterpiece, best quality, 8k, sharp focus
Negative: blurry, bad hands, extra fingers, watermark
FLUX.2 / FIBO 2.0 (DiffusionBear)
Click Enhance Prompt in DiffusionBear. On-device Qwen2.5-0.5B/1.5B-Instruct-4bit (~350 MB RAM, <1s) uses 9 canonical family-specific profiles to rewrite your 4 words into rich camera/lighting prose or a structured { } JSON scene layout, automatically keeping your LoRA trigger words intact before unloading from RAM.
Because FLUX.2 Klein and Bria FIBO 2.0 Edit Turbo are guidance-distilled flow-matching models, they do not require a negative prompt at all; describe the lighting, lens, materials, and spatial composition you want to see.
In DiffusionBee, sharing an image on Civitai required manually typing out the model name, steps, seed, and LoRAs. In DiffusionBear, every generated PNG (and even its gallery thumbnail) embeds:
tEXt parameters chunks and EXIF UserComment tags.DiffusionBee stores its downloaded and imported weights in a hidden folder inside your home directory at ~/.diffusionbee/downloaded_assets and ~/.diffusionbee/imported_models, packaged in a proprietary .tdict binary format.
Never. DiffusionBear opens ~/.diffusionbee strictly read-only, inspecting only the 288-byte binary header and JSON tensor key index. Converted LoRAs are saved as new .safetensors files inside DiffusionBear's own asset directory.
Click the Enhance Prompt button in DiffusionBear. It runs a local 4-bit Qwen2.5-0.5B or 1.5B model on MLX with 9 canonical family-specific profiles that expand short tag lists into natural descriptive prose for FLUX.2/3 or structured JSON scene layouts for Bria FIBO 2.0 while preserving any active LoRA trigger words.
Once you have converted your SDXL and FLUX LoRAs into DiffusionBear (which stores standalone .safetensors copies) and downloaded your preferred 4-bit MLX base models, DiffusionBear no longer needs ~/.diffusionbee. You can archive or remove legacy .tdict checkpoints if you no longer use DiffusionBee. Compare model timings in our Apple Silicon benchmark, read the DiffusionBee vs DiffusionBear comparison, or follow the 16 GB FLUX setup guide.