KSampler (Complex)

A stock KSampler is one pass over the latent: one prompt, one seed, one sampler, one scheduler. KSampler (Complex) breaks that into three legs - structure, style, specifics - each with its own model, prompt, seed, steps, CFG, sampler, and scheduler, all feeding the same latent in sequence. That's the "finer control over the diffusion process" the pack's README promises, and it's the flagship of this suite.

What it is

The idea is the same staging a photographer or painter uses: first lock in the composition (structure), then apply the look (style), then push detail (specifics). Each leg runs ComfyUI's standard sampler machinery on the previous leg's result - leg one starts from your latent at full denoise, legs two and three work from what came before. The node's web extension even color-codes the widget sections (Structure / Style / Specifics / Extra Parameters) so the wall of inputs is readable.

The settings that matter

The settings that actually matter:

• generation_strategy - how the steps are counted. total_pass (default) runs the three legs as one continuous denoise schedule, so total steps add up across legs. individual_pass restarts each leg from step 0, giving each a fresh full denoise. cumulative_pass is in the dropdown but currently behaves identically to total_pass in the code - don't expect a difference.
• preview_structure - runs only leg one and returns, so you can check composition and color before committing to the other two legs. Cheap way to iterate on a scene.
• low_res_hack (lrh_factor, 0.7) - shrinks the latent for the structure pass, then restores full size. The tooltip's claim is real: some models compose better at small latent sizes, and locking that smaller composition in before upscaling fixes a class of "empty/awkward composition" problems.
• hi_res_fix (hrf_factor 1.5, hrf_method double/single_early/single_late) - upscales the latent mid-generation, the same trick as the classic two-pass hi-res fix the KB covers in upscaling. The tooltip's warning is the practical one: "higher values tend to overcook the image." Trust it.

The share_* toggles (share_seed, share_cfg, share_sampler_name, share_scheduler) make the style and specifics legs reuse the structure values, and the optional style_model / style_positive / style_negative / specifics_* inputs fall back to the structure ones when left unwired. So the "default" experience is one model and one prompt, with the staging bookkeeping handled for you. The three seeds default to randomize each run - for reproducible output, lock the structure seed and tick share_seed.

Outputs and shortcuts

Three outputs come out: STRUCTURAL_LATENT (after leg one), STYLIZED_LATENT (after leg two), and FINAL_LATENT. VAE-decode the intermediate ones to compare stages side by side. A neat trick: set a leg's steps to 0 and it short-circuits to the previous result - style_steps = 0 means "skip style", which is the clean way to dial this back toward a normal sampler.

The honest part

Now the honest part. This is an experimental node from a single-commit pack, and it shows. pass_overlap exists in the UI, but the code says "TODO implement pass overlap" - it currently does nothing, so don't waste time expecting it to change anything. The default 7 steps per leg is very low - fine for drafts and seed farming, but for final images you'll want 20–30 steps and a converging sampler like DPM++ 2M. And the sampler rules from the KB still apply: on flow-matching models like Flux, Karras scheduling and CFG behave differently, so this node is happiest with classic SD 1.5/SDXL-style checkpoints and their fine-tunes.

Installing it

Install is the standard pack route - Manager (search "comfyui-fams") or git clone https://github.com/famellad/comfyui-fams into custom_nodes/, then restart. No model downloads. One thing to fix up-front: the pack needs ImageMagick + wand installed or it won't register at all, even though this node never uses them - check that before you blame the sampler.

If you read this and think "I could just chain three KSamplers," you're right - you could, with worse ergonomics. This node's whole value is the staging bookkeeping: shared settings, per-leg control, and the intermediate latents to inspect, all in one box.