Forbidden Vision 💎 Refiner

Half the reason people stop sharing a first pass is that it looks flat: muddy shadows, washed-out highlights, colors that lean too warm or too dead. The Refiner (LatentRefiner) is Forbidden Vision's answer to that - a single node that analyzes an image with a custom-trained neural corrector and fixes exposure, black levels, and tone automatically, then optionally upscales it and even fakes depth of field. It's the pack's second-most-used node for a reason: you can drop it after any sampler, hit go, and the output simply looks more finished.

How it works

The core is the neural corrector, a model trained on curated pairs to reproduce the color-and-exposure adjustments the author used to do by hand. It runs a YUV-based analysis, lifts shadows, protects highlights from clipping, and balances color. Two strength dials blend between your original and the fully-corrected result: corrector_tone (default 1.0, full power) and corrector_color (default 0.7). If the image comes out too bright or too dark in places, pull corrector_tone down; if the "temperature" shifts, lower corrector_color. That's the entire tuning model - two sliders.

On top of that sit three optional extras, all off by default:

• Upscaling - enable_upscale + upscale_factor (the author's pick is 1.2 before a second diffusion pass). It loads any upscaler from your upscale_models folder, with a plain Bicubic fallback.
• AI relight - depth_relight gently brightens the foreground subject using a depth map, useful for lifting a face without flattening shadows.
• Depth of field - depth_dof_enable blurs background and foreground around a computed focus plane. depth_dof_focus (0.5–0.99) sets where that plane sits; depth_dof_model chooses Depth Anything V2-Small (faster, less VRAM) or V2-Base (more stable). The depth-estimation lineage here is solid - Depth Anything V2 is the standard modern monocular depth estimator, not an obscure hack.

Inputs, outputs, and where it goes

The node accepts a latent, an image, or both, plus a vae. The convenience design: if you feed it a latent it decodes it (with caching so repeated runs don't re-decode), processes the pixels, and if a vae is connected it re-encodes the result - so it works mid-workflow and as a final enhancer. Outputs are refined_latent and refined_image_preview.

The two quirks to know about:

• use_tiled_vae - if you're re-encoding a high-res image, tiled VAE encoding avoids seams and cuts VRAM. tile_size default 512; bigger tiles are faster but hungrier.
• maintain_aspect_ratio and enforce_mod32_boundaries - the node may crop a few edge pixels to keep the output a multiple of 32 and unstretched. Leave them on.

The README's suggested slot is a pre-pass before your second sampling stage: refine the tone first, then resample on top of a correctly-graded image. The neural corrector model auto-downloads on first run with the rest of the pack. One honest note: the corrector is marked experimental by the author, and tone corrections are taste - if it's too aggressive for your checkpoint, the tone/color sliders are there to dial it back rather than off.