FL KSampler SEG Krea

Regional samplers normally work by swapping prompts per region. This one swaps pictures per region. FL KSampler SEG Krea takes the segmented regions your workflow already built and, for each one, encodes that region's crop of the source image as its own Krea 2 visual reference. The model doesn't get told "this tile is a window frame, redraw it" - it gets shown the tile, and the prompt asks it to enhance what it sees without inventing new objects.

That's a genuinely different idea from the usual tiled-sampler stack, and it slots into the pack's SEG pipeline: Regions → (captioner/encoder, optional) → sampler. This node replaces the prompt-encoding step for Krea 2 workflows.

What it does, concretely

For each region, the node takes the padded bounding box the region chain already computed, crops the matching area out of source_image, scales it, and encodes it as a Krea reference with role custom in full reference mode - meaning the visual tokens go through, not just the image-informed text. Then it packs that conditioning onto the region and hands back a model that knows how to blend per-region references.

In other words: the same machinery as FL Krea Reference and FL Krea Reference Guider, wired up per tile automatically. You're not managing a hundred reference nodes; you're feeding the sampler one source image and letting it do the bookkeeping.

The inputs

model, clip, regions, source_image, prompt, reference_strength, reference_resolution.

regions is a SEG_REGIONS socket from the pack's region nodes; prompt defaults to a long instruction worth actually reading, because it defines the whole job: "Enhance natural detail while preserving the reference crop's content, composition, lighting and colors. Do not add objects or extend the scene." That's an upscaling/detailing brief, and it's the right one - change it to something more creative and you're asking the model to redraw, which is a different node.

reference_strength is the dial from the tooltip: 0 uses the prompt alone, 1 uses the matching source crop, intermediate values blend both predictions and require an extra model evaluation. So the middle of the range is the expensive part; pick an end unless you have a reason.

reference_resolution (256/512/1024/1280) is the longest side the crop gets before vision encoding. 512 is the default and a sensible place to stay for detail work.

Outputs and wiring

Three: model, regions, conditioning. You use all three with FL KSampler SEG at CFG 1 - the model comes pre-patched for the reference blend, the regions carry the per-region conditioning, and the conditioning is the baseline. This is the same CFG-1 discipline the Krea reference guider needs; ignore it and the guidance double-dips.

Install

Standard pack install:

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

Restart ComfyUI. You need a Krea 2 checkpoint and a CLIPLoader with type krea2; the node checks both model class and tokenizer and errors out clearly if they're wrong. Fill-Nodes drags in a long pip list, and you need a current ComfyUI for the Krea 2 text encoder imports.

Where people get burned

Size mismatches, twice over. source_image must be a single RGB image at exactly the same size as the regions were built from - the same resized image you're sending to VAE Encode. The node says so in its error, and it means it. Both source dimensions also have to be multiples of the VAE's spatial downscale ratio, or the reference crops won't line up with the latent tiles.

Refinement with denoise. This is regional refinement, not generation from scratch. If you feed an empty latent at a low denoise you get noise - the underlying SEG sampler enforces the rule that empty latents want denoise 1.0, or a real latent with denoise below 1.0.

Cost per region. Every region with a nonzero strength adds model evaluations at sampling time. Six regions is a lot of passes. Start with two or three regions at strength 1.0 and see whether the quality is there before you tile a whole 4K image.

Used well - a handful of regions, strength at 1.0, the default preservation prompt - this is one of the cleaner ways to get a Krea 2 model to add detail without hallucinating new content into your picture.