h4 - Double Sampler (Advanced)

H4_DoubleSampler is the monster truck of samplers, and its display name ("Advanced") is earned. Where a stock KSampler does one pass, this node can do two (primary plus a refiner stage), slide your CFG down across the steps, and - if you're feeling chaotic - inject prompt mutations via a "chaos engine." It's h4_Live's answer to "I want a fancy sampling pipeline without wiring six nodes." The trade is that it's a big node with a lot of knobs, and most of them are optional, gated behind three toggles.

The parts you'll actually use

The core is a normal sampler: model, positive, negative, latent_image, seed, steps, cfg, sampler_name, scheduler, denoise. Those behave exactly like a KSampler, so start there and pretend the rest doesn't exist. euler with a normal scheduler at 20 steps is a fine default.

The interesting stuff is gated:

• enable_stage_2 - turns on the refiner pass: its own stage_2_steps (default 10), stage_2_denoise (0.35), stage_2_cfg, sampler, and scheduler. Classic primary-then-refine: run the main pass, then a light low-denoise pass on top. This is the two-pass engine the name promises.
• cfg_sliding_scale - when on, CFG shifts from your starting cfg down to cfg_end (default 4) across the steps. The idea, per the tooltip: lower CFG near the end gives the image more "air." It's a cheap way to get crisper early structure with less color-fry at the end.
• enable_chaos_engine - the weird one. It mutates prompts across the run (or spawns chaos_batch variants), with chaos_mode choosing the injection pattern (Pure Chaos, Odds, Evens, every Nth, Random Pulse), chaos_range the weight spectrum (negative values "try to delete concepts"), and chaos_denoise how strongly variants rewrite the original. This is exploration/dada territory - great for finding happy accidents, terrible if you want reproducibility.

The outputs

Beyond the obvious latent and preview, it hands back telemetry you can actually use: processed_prompt (the prompt after stutter/chaos editing), seed_out, steps_out, cfg_out, sampler_str, scheduler_str. So you can log exactly what ran - which matters when the chaos engine was on and the "real" prompt is whatever came out the other end.

There are also optional text inputs (positive_text, negative_text, wildcard_text) and optional clip/vae - wire those in and the node can do its own prompt encoding instead of taking pre-built conditioning. And prompt_stutter randomly repeats tokens to force emphasis - the tooltip calls it "deep-fried emphasis levels," which is a fair warning.

Installing

ComfyUI Manager → search "h4_Live", or:

cd ComfyUI/custom_nodes
git clone https://github.com/m3rr/h4_Live

restart. Pure PyTorch, no extra deps.

Where people get burned

The traps are all about the toggles. Stage 2 with default stage_2_denoise 0.35 is gentle; if you set it high, the refiner rewrites the image and you'll wonder why your output stopped matching your prompt. The chaos engine and stutter are off by default for a reason - turn them on, and seeds stop being reproducible, which confuses anyone who treats this like a normal sampler. And the honest caveat: this is a young, in-development node in a pack the author calls a "Massive project." The two-pass and CFG-slide features are genuinely useful; treat the chaos engine as a toy until it's been battle-tested. Start with stage 2 only, get comfortable, then decide if you want your samples to grow tentacles.