Genfocus Bokeh (Native)

GenfocusBokeh is the payoff node of the whole pack. Give it a sharp image plus a defocus map - the grayscale "how much to blur each pixel" map - and it generates realistic depth-of-field using the trained BokehNet LoRA. This is the "shallow DOF portrait" and "tilt-shift miniature" effect, done by a model that learned what lens bokeh actually looks like rather than by blurring with a filter.

Why does that matter? Because fake bokeh is one of the biggest tells in AI images. The KB's photorealism doc flags it directly: "Background blur: doesn't match real lens optics. AI bokeh is often too uniform or has wrong falloff patterns." A Gaussian blur of a depth map is exactly that kind of tell. Genfocus is generative - FLUX re-renders the scene with the defocus map steering where blur accumulates, so you get the characteristic falloff, edge behavior, and structure of real lens blur instead of a smudge. That's the whole point of spending 23GB on FLUX for this.

How it works

This is the "Native" (diffusers) sibling of the BokehNetApply utility node - and the one that actually applies the LoRA. Internally it mirrors the paper's demo: your sharp image becomes one condition, the defocus map becomes another (with no_preprocess=True, since it's already a normalized tensor), both tagged with the bokeh adapter, and only the bokeh LoRA is activated. The stock prompt "an excellent photo with a large aperture" runs at guidance 1.0. It even auto-resizes the defocus map if it doesn't match the image dimensions - so wiring mismatches won't crash it, though you should still aim to match sizes upstream.

Inputs

• pipeline - GENFOCUS_PIPELINE from either Genfocus loader.
• image - a sharp image. The README is blunt: feed it blur and the results are soft and unpredictable. Deblur first.
• defocus_map - a plain IMAGE. This is the defocus_map output from Compute Defocus Map (the normalized one, not the preview).
• steps (28) - recommended; more if you want refinement.
• seed (42).
• use_custom_prompt / custom_prompt - off by default. The stock prompt is calibrated for the LoRA.

Output: bokeh_result.

Getting the effect you want

The blur intensity lives upstream in Compute Defocus Map, not here. Two knobs there do the real work:

• blur_strength (K) - how fast blur grows with depth distance. Subtle portrait: 10–20. Standard bokeh: 20–30. Dramatic tilt-shift: 40–60.
• max_coc - the cap on blur radius. 64 for gentle softening, 100+ for obvious separation.

One thing to set expectations about: even if your focus point matches the original photo's focus, the result won't be byte-identical - FLUX re-generates the image, so already-sharp areas get slightly softer. If you want no change in already-sharp regions, the defocus map needs to be genuinely zero there; check the preview output from Compute Defocus Map to confirm.

Install

Part of the Refocus pack - ComfyUI Manager ("Refocus - Generative Refocusing") or git clone https://github.com/EricRollei/comfyui-refocus - with the full diffusers stack (pip install diffusers transformers accelerate peft), the Genfocus LoRAs in models/genfocus/, and a FLUX-dev model via one of the two Genfocus loaders. Remember the constraint that applies to every native node: FLUX-dev architecture only, never Schnell.