Automatic CFG - Warp Drive

Here's the deal that makes this node famous inside the pack: real CFG runs the model twice per step - once with your prompt, once without - and the second pass is where roughly half your render time goes. Warp Drive keeps the quality story of Automatic CFG but stops paying for the negative pass most of the time, replacing it with a fake negative built from the previous step's positive prediction. The author claims average speeds around 160% of a normal run when paired with the AYS scheduler. In practice: several steps rendered in about half the time, and results that stay surprisingly coherent.

The idea traces to the author's own experiments with delayed guidance (his update thread credits the "start guidance later" idea to arXiv 2404.07724, guidance interval). The twist: instead of just not having a negative, it averages the current positive prediction with the previous step's - a "previous_average" - so the stand-in negative is a plausible self-portrait of where the image is going, not a zero tensor.

The inputs that matter

Three floats, all sigma thresholds, and the defaults matter more than the labels:

• uncond_sigma_start (default 5.5) - the real negative only runs while sigma is between uncond_sigma_end and this. Above 5.5, the fake negative takes over.
• uncond_sigma_end (default 1) - below this, no negative at all.
• fake_uncond_sigma_end (default 1) - where the fake-negative trick itself stops.

The console helpfully prints the author's cheat sheet when you queue a render:

• 1 / 1 / 1 - maximum speed, uncond fully disabled (no negative, ever). "Fasten your seatbelt!"
• 3 / 1 / 1 - risky "space-time continuum distortion," awesome for prompts with a clear subject.
• 5.5 / 1 / 1 - the default: "Frameshift Drive Autopilot… should work with anything but do it better and faster!"

So: leave it at defaults, or push all three to 1 if you want the pure speed run. It's effectively the "boost" idea from the main node, but running for most of the render instead of just the tail.

Output and wiring

One patched MODEL into your KSampler. Pair it with the AYS scheduler and a 10–12 step SDXL workflow, per the pack's example images (the README shows a 10-step SDXL AYS warp-drive run at 2.5s on an RTX 4070). Set CFG to 8, the pack's usual reference.

Caveats that matter

• The negatives you skip don't exist for 3/4 of the render, so your negative prompt only influences the phase between ~5.5 and 1 sigma. Write it for the late "what to avoid in details" job, not the whole composition. The author's own framing: this replaces the negative, so results differ from a normal render - check them before you assume a bug.
• Community reports on this pack from 2024 flagged black-image failures after install even on unrelated workflows, traced to the global sampling-function patch going stale. If Warp Drive (or anything in the pack) misbehaves, update ComfyUI first.
• It's tuned for SDXL-era workflows and the author has since moved the concept into Skimmed_CFG - worth a look before you commit long-term.

cd ComfyUI/custom_nodes
git clone https://github.com/Extraltodeus/ComfyUI-AutomaticCFG

Restart ComfyUI. Warp Drive is the pack's best speed-vs-quality bargain, and the easiest win if you want the anti-burn benefits and faster renders instead of just one of them.