Local AI line · stop 01 of 14 · 45 min · members
A full local AI studio on one Apple Silicon Mac: image generation, virtual try-on, 4K upscaling and video with open models that replace Midjourney, Magnific and Kling. The exact tested setup, with real benchmarks and the traps to skip.
Free with an account
Membership is free: an account opens all 86 script pages. The Lab, Studio Canvas and the paid guides need the $99 pass, paid once. Already signed in on this browser? The page opens by itself.
Overview
Image generation, instruction editing, virtual try-on, upscaling, video — all local, all free, no content filters, no credits.
This is the exact setup running on my M5 Max right now: one ComfyUI workspace that covers photorealistic image generation, instruction-based editing, virtual try-on, 4K restoration and video — with open-weight models that match or beat the tools I used to pay for.
Everything below was tested on real agency work: fashion editorials, product shots, ad storyboards and the images on this very site.
Why local wins: full ControlNet and LoRA control, your data never leaves the machine, batch generation is unlimited, and the whole thing is scriptable — Claude Code can drive it end to end through an MCP server. The fashion set on my homepage was generated this way, 15 seconds a frame.
Golden rules
Learn these five rules before downloading a single model. They are the difference between 15 seconds and 15 minutes a frame.
1. Never use fp8 models. There are no fp8 kernels on Apple's MPS backend — fp8 files either crash or silently fall back to CPU at 7x slower. Always pick bf16 / fp16, GGUF Q6_K / Q8_0, or the newer int8 builds (those do work, and at half the memory).
2. Unified memory is your VRAM. On 48GB, budget ~40GB for the diffusion model + VAE + latents. Quit Ollama and LM Studio models before big renders — one resident 26B chat model eats 18GB.
3. Launch with --use-pytorch-cross-attention. It is the fastest attention path on MPS.
4. Generate at 1-1.5MP, then upscale. Direct 2K generation is 3-4x slower on Mac. Render at 1024-1536, then let SeedVR2 take it to 2K/4K in one diffusion step.
5. Lightning LoRAs are mandatory for speed. The 4-step and 8-step distillation LoRAs cut 20-50 sampling steps down to 4-8 with minimal quality loss.
The arsenal
Ten models cover everything. Each one earned its disk space in testing — here is the honest map.
Images: Z-Image-Turbo (6B bf16) is the hero — flawless anatomy, editorial quality, 8 steps, ~15 seconds a frame once warm. Qwen-Image 2512 (20B, GGUF Q6_K) adds human realism and perfect typography — it writes Turkish and English correctly on posters. FLUX.2 klein 9B handles multi-reference composition: lock a face and a product across up to 10 reference images. Krea-2 Turbo copies the style of any moodboard with its Style Reference LoRA.
Editing: Qwen-Image-Edit 2511 (GGUF) does instruction edits — "make the background a beach at dusk" — plus virtual try-on, material swaps and relighting, in ~2.5 minutes with the 4-step LoRA.
Upscale: SeedVR2 (3B for video, 7B Sharp for stills) is the local Magnific/Topaz. One diffusion step, real detail recovery.
The node packs that matter: ComfyUI-GGUF (loads the 20B models on Mac), Impact Pack's FaceDetailer (automatic face and hand repair), controlnet_aux (pose/depth/edge preprocessing), RMBG (background removal), Crystools (memory monitoring — critical on unified memory), and comfyui-ollama for local prompt enrichment.
Benchmarks
Measured on an M5 Max, 48GB, not copied from a spec sheet. Including the failures.
All timings below are end-to-end on ComfyUI 0.34 with torch nightly:
Z-Image-Turbo · 1152×1536, 8 steps → 54s cold, ~15s warm. The int8 build: 46s at half the memory, near-identical output.
Qwen-Image 2512 · 1140×1472, 8-step Lightning → ~5 min a frame, film-camera realism.
Qwen-Edit 2511 try-on · two references → 148s, garment transferred flawlessly.
FLUX.2 klein 9B · single reference re-framing → 289s, identity fully preserved.
LTX-2.5 22B int8 · 3-second 9:16 clip with synchronized audio → ~104s once models are warm. This is the Mac's video champion.
The honest failure: the Wan 2.1/2.2 family is currently broken on Apple MPS — chroma streaking regardless of precision (a confirmed upstream kernel bug, not a settings issue). The files can wait on disk; use LTX-2.5 or HunyuanVideo 1.5 for video instead. Knowing this saves you a wasted day.
Agency workflows
The stack is only worth building if it ships client work. These are the pipelines it runs for mine.
Fashion editorial: Z-Image-Turbo hero frames → DWPose ControlNet to lock a pose across ten variations → Qwen-Edit try-on to put the real product on the model → FaceDetailer → SeedVR2 to 2K. No SaaS content filter rejecting swimwear shots.
Ad storyboards: depth-ControlNet keeps framing consistent shot to shot; Qwen-Image writes the campaign line directly into the key visual, correctly, in any language.
Video: pick the best still → LTX-2.5 image-to-video with synchronized sound → RIFE interpolation to 48fps → SeedVR2 video upscale.
Full automation: the whole studio is scriptable. Claude Code drives ComfyUI through an MCP server — my site's 19-image brand set was generated, color-managed and placed by an agent while I reviewed frames. That workflow is its own guide: The Claude Resolve Stack's sibling, for images.
Build order
Do it in this sequence and nothing bites you.
1. pipx install comfy-cli, then comfy install — one workspace, manager enabled.
2. Add the launch flag to comfy-cli's config: --use-pytorch-cross-attention.
3. Install the node packs (GGUF first — nothing 20B loads without it).
4. Download models in a queue, biggest last: Z-Image-Turbo bf16 → Qwen-Edit 2511 Q6_K → SeedVR2 3B → Qwen-Image 2512 Q6_K → LTX-2.5 int8. Skip every file with fp8 in the name.
5. Test each model with its official template before customizing. Measure once warm — the first run always includes model loading.
6. Wire it to your tools: comfy-mcp for Claude, or the plain HTTP API for anything else.
1:1 · two hours