MD: PingPong FBG (Full Control)

If you landed on this page, you probably loaded a workflow and found a wall of ~60 sampler knobs staring back at you. PingPong FBG ("Full Control") is the flagship sampler of MD Nodes, originally built for ACE-Step audio generation and later borrowed for image and video texture work. It's a bidirectional ancestral sampler with a feedback-guidance loop, and it returns a plain SAMPLER that drops into a standard KSampler - the complexity is all inside.

What the mechanism actually is

Ping-pong sampling runs a forward and backward pass inside each step (that's the "bounce"), which resolves fine detail that a one-directional solver blurs - the README's pitch is "hyper-detailed textures that standard samplers turn to mush." On top of that sits feedback guidance (FBG): it watches recent steps and tightens or loosens guidance dynamically, and it includes NaN recovery so a bad step degrades gracefully instead of nuking the latent.

The knobs cluster into a few ideas. Step randomization (step_random_mode: off / block / reset / step, with step_size) rotates the seed every N steps so repetitive texture doesn't lock in - block is the recommended default. Ancestral noise injection (first_ancestral_step, last_ancestral_step, ancestral_noise_type: gaussian / uniform / brownian) adds creativity mid-run, with brownian suggested for organic, analog texture. blend_mode controls how new noise mixes with the latent - lerp is fast, slerp preserves spherical structural integrity. enable_restarts plus restart_mode raises noise partway through so the model can fix detail mistakes, at the cost of render time.

What a beginner actually sets

Ignore most of the wall. You'll touch these:

• scheduler (required) - wire in the SIGMA schedule from the pack's Hybrid, GITS, or Noise Decay scheduler.
• seed, step_random_mode, step_size - texture variation.
• first_ancestral_step / last_ancestral_step - leave at 0 / -1 (disabled) for deterministic diffusion; raise the first one to 3–5 if your initial composition is messy.
• fbg_sampler_mode - EULER is recommended for standard ACE-Step audio; PINGPONG for the full bidirectional path.
• enable_clamp_output - leave False; audio latents routinely exceed 1.0.
• enable_restarts + restart_mode - the detail fixer.

For the curious, dynamic CFG lives here too (cfg_start/end_sigma, fbg_start/end_sigma, max_guidance_scale), along with posterior factors (pi, t_0, t_1). Two parameters - fbg_eta and fbg_s_noise - are marked DEPRECATED in the schema; ignore them. You can also override everything by feeding YAML into yaml_settings_str, which is where MD_YAML_Utils earns its keep.

The straight talk

Don't start with this node. PingPongSamplerNodeLite and Basic wrap the same core with sane presets (noise_behavior, a couple of toggles) - use those to learn, then graduate to FBG when you need the production control. It's genuinely the pack's production workhorse for ACE-Step audio, and the concept traces back to blepping's original PingPong sampler, which MDMAchine credits and has also contributed to the HOT-Step-CPP C++ runtime. But for a plain image person it's overkill, and its home turf - ACE-Step music generation - is a real but niche corner of the ecosystem. If that's your corner, this is the sampler to learn.

Installing it

cd path/to/ComfyUI/custom_nodes
git clone https://github.com/MDMAchine/ComfyUI_MD_Nodes.git
cd ComfyUI_MD_Nodes
pip install -r requirements.txt

Or search MD_Nodes in ComfyUI Manager and restart. The solver logic lives in the pack's core/ subpackage (source ships with the repo); if that import ever fails, the node returns a fallback dummy instead of crashing - so if sampling suddenly does nothing, check your ComfyUI console for a core import error.