FL Krea Reference Guider

FL Krea Reference Guider is the half of the Krea reference pair that actually changes your image. FL Krea Reference describes what to borrow from one picture; this node blends all of those references into a single prediction during sampling. It's the one with the interesting mechanism and the annoying cost.

Here's the cost, stated up front because it's the thing that surprises people: each active reference adds a model evaluation. Not a cheap conditioning concatenation - an actual separate forward pass, blended in sigma space. Three references means four model passes per step. On a 12B model that is not a rounding error.

How the blend works

Under the hood this is a CFGGuider subclass, which means it owns the per-step prediction. Each step it takes the baseline scene prompt (weight = 1 minus the sum of active reference weights), calls the sampling function once per active reference with that reference's conditioning and CFG 1, and sums the weighted predictions. Then the sigma envelope zeroes anything outside its window. Start, end and fade aren't decoration - they're how you keep three references from fighting for the whole schedule.

The node hands you three outputs and two ways to spend them:

• GUIDER goes into SamplerCustomAdvanced, the proper route.
• MODEL and CONDITIONING are the convenience route: connect MODEL straight through and put CONDITIONING into a normal KSampler's positive input, with CFG 1. The node patches the model's conditioning-batch function so an ordinary KSampler can do the blending. Set CFG above 1 and you're double-dipping - the whole point of CFG 1 here is that the reference blend is the guidance.

The inputs

model and clip must be Krea 2. The node checks the model class and the tokenizer type and raises Use a Krea 2 model and CLIPLoader with type 'krea2' if you feed it anything else. It also refuses a model that already carries custom CFG or guidance patches, because those would collide with its own.

prompt is your actual scene prompt - the thing the references are modifying. Default "A dog in the park." is placeholder text, replace it.

influence is the master dial: 0 means references are ignored entirely and you're sampling the prompt alone, 1 means full. 0.7 is a sane starting point. references is an autogrow input - click to add sockets, prefix reference_, up to 100. Each takes an FL_KREA_REFERENCE from FL Krea Reference.

blend_mode is add or average. Add sums the contributions; average scales toward dividing by the number of enabled references. average_amount is the dial between the two: 0 keeps the additive blend, 1 fully averages, with individual weights still applying. Disabled references drop out of the count entirely; fading or zero-weight ones keep their share. In practice: with one reference it doesn't matter, with two or three it's how you stop the scene from being pushed too hard.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_Fill-Nodes

Restart, and make sure your ComfyUI is recent enough to ship the Krea 2 text encoder - the guider imports it directly. Everything else here is pure Fill-Nodes; the pack's long requirements list (OpenGL, audio, PDF and API libraries) applies, and the pack itself has no model downloads of its own. If you're on Linux and starting from scratch, the pack ships a helper that installs libgl1-mesa-glx and freeglut3-dev for the shader nodes - you don't need it for this one, but missing OpenGL libraries are a common startup complaint for the pack as a whole.

Where people get burned

Speed. If your generation suddenly takes four times as long, count your references before you blame your GPU. Drop to one active reference while you dial in prompt and influence, then add the second.

Order-of-magnitude fights. With three or more references the community's experience is consistent: they compete rather than compose. Two is usually the sweet spot - one as the primary look, one as a secondary palette or texture contributor - and beyond that results get unstable. If two references both want to own the composition, put the weaker one's weight down and give it a narrower start/end window.

CFG confusion. Both routes need CFG 1. If your images look fried or over-guided, check that first.