BlehContrastiveOrthoCFG

Standard CFG scale is one dial that amplifies the difference between your positive and negative predictions equally. This node splits that into two dials - how much of what's unique to the positive prompt gets added, and how much of what's unique to the negative prompt gets subtracted - computed as orthogonal (non-overlapping) features rather than just the raw difference. The node's own description is refreshingly honest about the baseline case: at the default scales of 1.0 on both sides, "this is just a slower version of CFG." The point of the node only shows up once you pull positive_scale and negative_scale apart.

Three patch modes, and the tooltips here are unusually thorough for this pack, worth reading closely rather than skimming. pre_cfg applies the positive change to the conditioned prediction and the negative change to the unconditioned one, then lets ComfyUI's own CFG math do the combining - which has a real consequence: at CFG 1, only the positive scale does anything, because the unconditioned pass isn't part of the computation at CFG 1 in the first place. post_cfg subtracts the negative-unique features and adds the positive-unique ones directly onto the already-computed CFG result, relative to the original cond/uncond generations. post_cfg_base_denoised is flagged by the author himself as experimental and "generally not recommended" - it can behave strangely if you're stacking other CFG-type effects afterward, like CFG++.

Inputs: model is the only required one. Everything else is optional with defaults tuned toward the safest starting point. positive_scale (0.2 default) is deliberately small - in pre_cfg mode it effectively gets multiplied by whatever your actual CFG value is. negative_scale (1 default) is subtracted, similarly scaled by CFG in pre_cfg mode. patch_mode defaults to pre_cfg, which the tooltip recommends explicitly: "for the most predictable effects, use positive scales and pre_cfg mode with use_noise disabled." start_dim/end_dim (1/1 default) pick which tensor dimensions count as "features" when figuring out what's orthogonal - the default targets the channel dimension, which is right for image models; video models insert an extra frames dimension after channels, so that's worth adjusting if you're patching a video model. use_noise (false default) applies the effect to the noise prediction instead of the clean image - the author calls this the harder case to reason about, since separating orthogonal features is murkier in noise space. force_uncond_generation (false default) overrides the usual CFG-1 shortcut that skips computing the unconditioned pass entirely, useful if you're running at CFG 1 but still want this node's negative-side effect to actually compute.

Output is a patched MODEL.

This is another node with no README section and nothing findable in wider community discussion - everything above comes from the node's own tooltips, which happen to be some of the most detailed in the whole pack, likely because the mechanism genuinely needs that much explaining to use correctly. Start with the recommended combination - pre_cfg, use_noise off, small positive_scale - before pulling any of the other knobs.

Install is standard: ComfyUI Manager, or a manual clone of blepping/ComfyUI-bleh into custom_nodes plus a restart. No extra dependencies.