Forbidden Vision 🛠️ Builder

The Builder is Forbidden Vision's replacement for the plain KSampler in the first pass of a workflow. It's a sampler plus three things a KSampler doesn't give you: resolution presets, a self-correction polish pass, and CFG scheduling that doesn't just hold one value for the whole run. If the flagship Fixer is the pack's showpiece, the Builder is the everyday workhorse that makes its example workflows feel less cluttered.

What it does

Standard sampling inputs - model, positive, negative, seed, steps (default 15), cfg (default 5.5), sampler_name (euler_ancestral), scheduler (sgm_uniform) - plus the extras:

• Resolution presets. resolution_preset lists 19 named options covering SDXL portrait/landscape ratios, SD1.5 sizes, and squares, so you pick "📱 2:3 Portrait (832x1216)" instead of typing two numbers. Custom unlocks custom_width/custom_height (must be divisible by 64 - latent resolution is image/8, and the node enforces sensible bounds of 256–4096). batch_size up to 64.
• Self-correction. self_correction (on by default) runs a final 2-step pass at 0.05 denoise when sampling finishes. That low-denoise polish stabilizes shapes and scrubs out the overcooked artifacts that high-CFG first passes leave behind. The README's demo shows 5 active LoRAs at CFG 7 coming out visibly cleaner with it on. Recommended on for most workflows.
• Adaptive CFG. cfg_mode is the interesting bit. Constant holds CFG for all steps (the KSampler behavior). Linear transitions straight from cfg to cfg_finish. Ease Down drops fast toward cfg_pivot early, then settles gently down to cfg_finish - the idea being you get away from high CFG quickly (avoiding burn artifacts) while still landing on your target value. For a normal SDXL run starting at CFG 5–9, the author suggests a pivot of 3–4; for turbo/distilled models, drop it to 1–2. CFG at 1 is worth remembering here because ComfyUI skips the negative prediction entirely at exactly 1, so low-CFG workflows get a speed win for free.

Outputs are LATENT and IMAGE - the image is decoded only if you connect the optional vae (if you don't, the latent stays valid and you get a placeholder preview).

Where it fits

The pack's suggested flow is: Builder for the first pass → Refiner to fix tone and color → a second pass (often the Rebuilder) → the Fixer if faces need it. The Builder's job is to find your composition without fighting settings, which is why the README tells you to start a workflow with just the Builder enabled and switch on the later stages one at a time. If you already live on core KSamplers, this is a mild convenience rather than a capability jump - but if you're assembling a Forbidden Vision workflow, it's the node that keeps the first pass from becoming four separate nodes.