SamplerClyb_BDF

"BDF" here stands for backward-differentiation-formula-style, which sounds like a textbook you'd rather not open. What it actually means is friendlier than the name: this sampler does roughly one model evaluation per step, then refines the result using a synthetic prediction at a midpoint, and stabilizes the blend with some linear algebra. It's the author's take on getting high-order-solver quality from a near-one-eval-per-step budget on flow models - Chroma, FLUX, SD3.

How it works

At each step it works with two predictions:

1. A denoised estimate at the current sigma - and here's the trick, it caches the previous step's denoised result and reuses it as the "first" prediction instead of re-evaluating the model. That's how the NFE count stays low.
2. A second, refined prediction obtained by evaluating the model at an intermediate state between the first prediction and the noisy latent.

Then it merges the two with one of three scalar fusions:

• projection - projects the half-step prediction onto the line spanned by the refined prediction, forcing them to agree in direction. Cheap, stable, the "boring but reliable" option.
• atan2sin - blends the two predictions in angle space using atan2(sin(…), cos(…)). This is the Chroma-flavored one; the same family of math shows up in the pack's guidance node.
• atan2sin+projection (default) - does the angle-space blend then projects it back into the line of the refined prediction. Both stabilizations, and why it's the default: on a test image it's the least likely to drift.

For flow-matching models the sampler switches to a flow-aware ancestral update (using alpha_ip1/alpha_down/renoise_coeff instead of the classic sigma_up noise injection). It detects flow models automatically from the model's sampling config, so you don't manage that yourself.

The inputs that matter

• scalar (default atan2sin+projection) - the fusion mode above. When in doubt, leave it.
• eta (default 1.0) - ancestral stochasticity. 0 is deterministic; lower values give cleaner, more repeatable output with slightly less variety. This is the knob to reach for first if you're getting noise artifacts.
• s_noise (default 1.0) - global noise scaling. Leave it unless you know why you're moving it.

Output is a SAMPLER - wire it into a KSampler's sampler input. The pack also registers it as the string clyb_bdf in the standard KSampler dropdown, so you can use it without the node if you prefer names over cables.

Installing

Part of the ComfyUI-ClybsChromaNodes pack; nothing extra to install:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/Clybius/ComfyUI-ClybsChromaNodes.git
# restart ComfyUI

Or ComfyUI Manager → search "ComfyUI-ClybsChromaNodes". No model files are downloaded - the pack is pure code by Clybius (the Chroma-GGUF maintainer), aimed at Lodestone Rock's Chroma and compatible flow architectures.

Common issues

The default fusion is the safest; switch to projection if atan2sin ever looks over-cooked or structurally weird on a subject-heavy image. And remember the ancestral-sampler rule: with eta above 0 this is not a converging sampler, so adding steps changes the image rather than refining it - change one variable at a time on a fixed seed when you're comparing.