◎ Radiance Sampler Pro

The default KSampler is fine until you need any of the things it can't do: phase-shift sampling, Align Your Steps schedules, perturbation attention (PAG), tiled sampling on a big latent, or sampling a WAN video with the right sigma schedule. ◎ Radiance Sampler Pro is the pack's one-sampler-to-rule-them-all: it auto-detects the model architecture (Flux, SD3, SDXL, WAN, LTX, HunyuanVideo, Lumina2, and friends), then exposes everything the modern sampler toolbox has, in one node.

If you only ever generate with a stock KSampler, this is overkill and you should keep your simple workflow. If you've ever stacked a second sampler for a phase shift, or wished your sampler understood your Flux shift, or wanted tiled sampling without a separate patch node - this is the consolidation. It also pairs with the rest of the pack: wire latent_format from Radiance Resolution / VAE Encode in and it validates channels before you waste a run.

How it works

Under the hood it wraps ComfyUI's core sampling machinery (it uses the stock KSampler.SAMPLERS and SCHEDULERS lists, so every sampler/scheduler you know is here) and layers on production features. The headline is sampler_mode:

• Standard - a normal sample with your chosen sampler/scheduler.
• Phase-Shift (Euler→DPM) / Phase-Shift (Euler→SGM) - start with one sampler and hand off to another at phase_split (0–1), with optional sigma_blend_steps to smooth the transition. This is the community's "detail then smooth" trick, baked in.
• CFG++ (Perpendicular) - the CFG++ variant that separates the score direction; helps at higher CFG.

Then the per-architecture dials. For Flux-style models there's flux_shift and flux_guidance (with a Dynamic (Creative Start/End) profile option). For video there are normalization factors. ays_schedule uses the Align Your Steps research-optimized sigma schedule (best at 8–15 steps). noise_type swaps the initial noise - Gaussian, Perlin (coherent structure), Spectral (pink/1f), Brownian (video-correlated), Uniform. pag_scale adds perturbation attention for prompt adherence (0 = off). guidance_rescale_phi applies Imagen-style guidance rescale (0.7 recommended for SDXL to stop oversaturation at high CFG).

There's also multi_cond_mode (average/weighted/sequential blending of a second positive via positive_2), tile_mode with tile_size/tile_overlap/tile_blend for memory-efficient high-res sampling, a preview_method (TAESD/Latent2RGB) for mid-run previews, and a preset dropdown that pre-configures whole recipes (Flux txt2img, Flux Fast 12-step, Flux Schnell 4-step, SD3.5 Turbo, WAN txt2vid, and ~24 more).

The inputs that matter

For a beginner, honestly: model, positive, negative, latent_image, then steps, cfg, sampler, scheduler, seed. That's a KSampler's worth of familiarity. Then add, in order of payoff:

• preset - pick a model recipe and skip the rest of the thinking.
• model_type (default auto) - set it when auto-detect gets an exotic architecture wrong.
• sampler_mode / phase_split - the phase-shift upgrade.
• flux_shift / flux_guidance - the Flux-specific dials.
• denoise - for img2img; 1.0 is full denoise.

Outputs: latent (→ VAE decode), sigmas (the schedule used, chainable), sigma_report (human-readable timing/schedule), and latent_meta (JSON telemetry: arch, steps, scheduler, noise type, tile mode, seed, time - wire it to Radiance Show Text for debugging).

How to install it

It's in the radiance pack. ComfyUI Manager → "Radiance", or:

cd ComfyUI/custom_nodes
git clone https://github.com/fxtdstudios/radiance.git
cd radiance
pip install -r requirements_windows.txt

Restart, and it's under FXTD Studios/Radiance/Generate.

Common issues

The realistic gotchas: (1) force_exact_steps and force_full_denoise_steps are blank-tooltip power switches - leave them alone unless you're chasing a specific schedule behavior, they exist for edge reproducibility. (2) Phase-shift with sigma_blend_steps can add a visible hitch if you pick wildly different samplers; start with the two built-in phase-shift presets. (3) tile_mode at small tile_size with aggressive overlap will noticeably slow sampling - it's a VRAM saver, not a speed boost. (4) Because it auto-detects model_type, an unusual checkpoint can sample with subtly wrong defaults; latent_meta shows you what it chose, and the manual model_type override is right there.

One sampler to replace several, and the presets mean you can be productive before you understand half of it. That's a good deal for a node this deep.