SeFi Turbo (MLX, all-in-one)

SeFi-Image got buried at launch. The quality was genuinely good - strong text rendering, real realism, a 5B model trained on a tenth of Z-Image's compute - but launch week was mostly people asking "will we even get ComfyUI support?" and grumbling about the CC-BY-NC license. Then SeFiTurboMLX shipped: the whole SeFi-Image turbo pipeline in one node, running end-to-end in Apple MLX. On Apple Silicon, this is the reason the "semantic-first diffusion" family is finally a real option.

What the node actually does

The name is doing a lot of work. One node downloads the gated SeFi-Image-*-turbo checkpoint, converts its DiT to MLX under models/mlx/sefi/, runs the dual-time turbo denoise loop, and VAE-decodes - no CLIPLoader, no KSampler, no VAELoader wiring required. The one thing it can't do is run on your Nvidia box: MLX is Apple Silicon only, and the code checks for it up front.

Under the hood it's a faithful MLX port of the Flux2-style dual-stream backbone SeFi uses, with the Qwen3-VL-4B language tower as text encoder and a 32-channel AutoencoderKL decoder. Everything converts once and caches - first run pays for download plus conversion, reruns are fast.

The semantic-first bit, briefly

SeFi's trick is "dual-time" diffusion: instead of one latent, you get a semantic stream and a texture stream, and the semantic one runs ahead - by default delta_t = 0.1 - so composition is decided first, texture fills in behind. That's what the "turbo" recipe exploits: at 4 steps with guidance_scale = 1.0 (no CFG, matching the DMD2-distilled reference) you get fast generations that still compose well. Steps are tuned for 4/8/10; anything else is allowed but you're off the beaten path.

The inputs that actually matter

Beginners can touch these and ignore the rest:

• model - SeFi-Image-5B-turbo by default, with 2B and 1B turbo variants if you want speed over quality.
• prompt, width, height, seed - standard. 1024×1024 default, multiples of 16.
• steps - 4 default; try 8 for a bit more polish.
• quantize - q8 default, q4 to slim the DiT down, bf16 for full fidelity. Same ladder logic as GGUF: Q8 is basically indistinguishable from bf16 here, so if it fits, leave it.
• hf_token - your HuggingFace token. The model repo is gated (CC-BY-NC), so you must accept the license on the HF model page while logged in before anything downloads. You can set HF_TOKEN in your environment instead.

The other two you'll actually meet in the wild: low_memory frees the MLX text encoder after encoding (~2.9GB resident instead of ~5.6GB, at the cost of re-loading it per new prompt), and bn_denorm stays ON - turn it off and output turns washed-out and grainy.

The node outputs image plus texture_latent - the raw texture stream, if you ever want to re-decode it yourself. And yes, you can override the MLX encoder and VAE with torch ones: wire a CLIP into the clip input or a VAE into vae, or flip mlx_encoder/mlx_vae off. Just don't point the VAE at a generic 16-channel Flux VAE - SeFi's texture VAE is 32-ch, and the node picks the right file by slug for a reason.

Installing it

Via ComfyUI Manager, search "SeFiMLXComfyNode" and install. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/jwooldridge234/SeFiMLXComfyNode
cd SeFiMLXComfyNode
pip install -r requirements.txt

Then restart ComfyUI. Dependencies are light - mlx, huggingface_hub, safetensors, transformers - and no other custom nodes are required. The real heavy dependency is the model itself: the first generation triggers a multi-GB gated download plus a one-time torch→MLX conversion, so don't panic when the first run crawls. It's cached after that.

Where people get burned

• 403 / "not authorized" - the gated-repo classic. Accept the license at huggingface.co/SeFi-Image/SeFi-Image-5B-turbo with the account that token belongs to, then pass the token.
• Washed-out output - bn_denorm flipped off, or pad_to_max off (the DiT was trained on the padded sequence; turning it off is out-of-distribution → noise).
• Outputs that look "off" at a given seed - this is a fresh MLX port. The key mappings are verified statically, but the numeric forward was never run on Apple hardware before release. The README ships a parity harness for exactly this: python custom_nodes/ComfyUI-SeFi-MLX/dev/parity.py builds a tiny model and checks torch vs MLX on identical weights. Run it before you judge the model. Images also won't bit-match the native torch node at the same seed - the initial noise deliberately uses torch's RNG, since MLX's normal RNG gives worse compositions here.

Honest caveat: it's non-commercial by license, and the corner is weeks old. But as the one-stop shop for SeFi-Image on a Mac, it's the one you'd reach for.