Krea2 Two Style Transfer

What it is

This is the two-image version of the pack's core style-transfer node - take two reference images instead of one and blend their styles into your Krea 2 generation. The author is upfront that this is the experimental branch of the pack: the single-reference Krea2 Style Transfer node is the one that got the actual tuning work, and this one is where you go once that's working and you want to push further. Worth knowing going in: it's capped at exactly two references by design, not by an arbitrary limit the author might lift later. Testing with three or more found the signals start competing instead of blending - weaker transfer, quality loss, one reference randomly dominating, results that are hard to explain. Two is the ceiling on purpose.

How it works

The obvious assumption is that two references get averaged 50/50. That's not how it works, and the author tried the naive average first and rejected it - later sampling steps have outsized influence on the final image's outlines, colors and surface detail, so a straight 50/50 split made whichever reference entered the second stage look overpowering. What you get instead is stage-based blending: primary_reference (choose 1 or 2) picks which of your two prepared references enters the first, longer stage of sampling, and the other one takes over for a shorter second stage. first_phase_ratio controls how that time gets split, and the recommended default of 0.75 means the primary reference gets three-quarters of the run as the main anchor while the secondary reference still gets a real say in the final look.

This also means output is order-sensitive - swap which image is primary with everything else held constant and you'll get a genuinely different result, not just a subtle shift. That's a feature of the design, not noise.

Inputs and outputs that matter

model and ref_conditioning work the same as on the single-reference node - your Krea2 model and the conditioning you're also feeding your sampler. style_refs is the one new required input: a STYLE_REFS bundle, which comes exclusively from the companion Krea2 Two Style References node - you can't wire two loose reference_latents in here directly, they need to be bundled first.

For the settings you'll actually touch: leave mode on recommended first and just play with primary_reference and first_phase_ratio. The README's own testing found 0.70–0.80 to be the useful range, with 0.75 as the strongest default and 0.90 usually landing one-sided. Once you switch to custom mode you get the full spread - style_strength, ref_k_1 and ref_k_2 (per-reference versions of the K-path strength from the single-image node), stage_focus, ref_value_mix, low_scale_end, rf_mode, gamma, beta, the high/low scale start/end pairs, adain_strength, and blocks. That's essentially the single-reference node's advanced controls doubled up for two K paths; treat them the same way - don't reach for them until recommended mode plus primary_reference and first_phase_ratio have stopped being enough.

Outputs are just model (the patched model for your KSampler) and debug (a diagnostic string).

Installing it

Same pack as every node here. ComfyUI Manager, search "ComfyUI Krea2 Style Transfer," or cd ComfyUI/custom_nodes && git clone https://github.com/jieg9341-lab/ComfyUI-Krea2-StyleTransfer and restart. Nothing extra to download for this node specifically, but it needs the same working Krea2 setup (UNet, Qwen3-VL text encoder, Qwen-Image VAE) as everything else in the pack, and it needs two Krea2 Style Reference outputs already prepared and bundled through Krea2 Two Style References before you can wire it up.

Where people get burned

If your blend feels lopsided toward one image, check first_phase_ratio before you touch anything else - pushing it near 0.90 is the documented way to accidentally make the primary reference dominate too hard, and the fix is usually just pulling it back toward 0.75. If it's dominated the other way, you may have the wrong image set as primary_reference - remember that's the one getting the longer first stage, not necessarily the one you'd instinctively call "primary."

Don't reach for a third reference by stacking or chaining nodes to work around the two-image cap - the author tested that route and it degrades rather than improves. And given this whole node is labeled experimental even by its own author, start with the single-reference node until you're confident in what your reference images do individually. Diagnosing a bad two-reference result is much easier when you already know what each image looks like on its own.