Genfocus Model Loader (Components)

Here's the situation this node exists for: you want the real Genfocus refocusing quality, but the flagship loader wants a full diffusers-format FLUX.1-dev folder - a ~23GB download most ComfyUI users never made, because they load FLUX as single-file checkpoints through Load Checkpoint. GenfocusComponentLoader is the answer. It assembles a complete diffusers FluxPipeline from the component files you almost certainly already have lying around: a single-file FLUX transformer, the standard ComfyUI T5-XXL and CLIP-L text encoders, and the FLUX VAE. No 23GB folder, no extra download.

The output type is identical to the Native loader's GENFOCUS_PIPELINE, so every downstream Genfocus node - Generate, Deblur, Bokeh, Switch Adapter - works unchanged. This was the pack's headline May 2026 addition (v0.1.2), and it's the loader most people should reach for first.

What you wire up

Four required dropdowns, all of the form folder:filename, which the node resolves against ComfyUI's configured model folders:

• transformer - your FLUX .safetensors from checkpoints/, unet/, or diffusion_models/. The tooltip is blunt: Genfocus LoRAs were trained on FLUX.1-dev, fine-tunes will load but quality is untested, and don't use FLUX.1-schnell (timestep distillation differs).
• t5_encoder - a t5xxl_*.safetensors from text_encoders/ or clip/. Use bf16 or fp16 variants; fp8 is not supported by this loader.
• clip_l_encoder - a standard clip_l.safetensors (77 positions). The gotcha here: clip_l_hidream and other extended-context variants will not work - that's the single most common "incompatible with FLUX" error in the README's troubleshooting table.
• vae - the FLUX VAE, e.g. ae.safetensors.

Then the usual precision (bf16 recommended), optional deblur_lora / bokeh_lora from models/genfocus/, and offload_to_cpu.

How it works

Under the hood a FluxComponentAssembler builds the diffusers pipeline piece by piece, then loads the Genfocus LoRAs the same way the Native loader does - so the LoRAs still land on the exact architecture they were trained on. The one automatic download: on first use it fetches ~5MB of tokenizers and component configs from FLUX.1-schnell (Apache 2.0, non-gated) into Refocus/.flux_cache/, then reuses them forever.

Install and gotchas

Same pack install as everything else - ComfyUI Manager ("Refocus - Generative Refocusing") or git clone https://github.com/EricRollei/comfyui-refocus - plus pip install diffusers transformers accelerate peft. No diffusers FLUX folder needed, but the diffusers stack itself is still required, because the assembled pipeline runs diffusers under the hood.

Two failure modes from the README worth memorizing:

• "incompatible with FLUX" → you picked clip_l_hidream (for HiDream, not FLUX) or a t5-base/large. Use a vanilla clip_l.safetensors and a t5xxl_*.safetensors.
• Stale behavior after updating the pack → delete Refocus/pipeline/__pycache__/ and Refocus/nodes/__pycache__/, then restart ComfyUI. Bytecode cache has a habit of pointing at old code.

If the node isn't showing up at all, you're missing diffusers - the component loader imports separately from the native nodes and hides itself on failure.