Flake Generate

Flake Generate is the sink at the end of the pipe: drop a flake_data bundle in, pick a seed, and it runs KSampler, VAE Decode, and Save Image internally - no output pins, no sampler node, no filename node. The result renders inline with a small 2×2 grid and a "Generation Data" overlay showing what actually went into the image: checkpoint, VAE, LoRAs and their strengths, the BREAK-separated prompt segments, resolution, steps, CFG, sampler, and scheduler. It's the pack's answer to the complaint that presets hide what you're doing - here the receipt is printed right on the node.

The seed input is worth calling out because it's a native ComfyUI seed widget, with control_after_generate and the full 64-bit range. That means it behaves exactly like every other seed in your graph - randomize, fixed, or fixed-then-increment - instead of whatever half-baked custom control most pack terminals ship.

The two optional post-processes are where it gets interesting. adetailer offers Off, Face, or A/B. Face runs a face-detail pass via the Impact Pack's FaceDetailer - detect the face, regenerate it at higher resolution to fix the mushy-SDXL-face problem, composite it back. A/B generates the image twice, once with and once without the detail pass, so you can see exactly what the pass bought you; the filenames get an _ad marker so the two variants don't collide. adetailer_denoise defaults to 0.4, and adetailer_steps of 0 means "match the first pass's step count" rather than guessing a number. Then there's the upscale toggle: enable it, optionally pick a real ESRGAN-style model from models/upscale_models/, set upscale_factor (default 1.5), and every variant gets upscaled with an _up marker appended. Leave upscale_model blank and it's a plain Lanczos rescale to the factor - more pixels, no invented detail, which is the right default given how much the upscaling discourse has split between "more pixels" and "more detail."

The honest fine print. The ADetailer and upscale paths are soft dependencies, resolved at call time. ADetailer needs ComfyUI-Impact-Pack plus the Impact Subpack's UltralyticsDetectorProvider (the default bbox model is face_yolov8m.pt); if they're missing, you get a clear error telling you exactly what to install, and the rest of Flakes keeps working untouched. The pack's own source notes that Impact Pack's FaceDetailer signature varies between versions, so if you hit a signature error on a fresh update, that's a known-fragile seam, not you doing something wrong. Filenames are built from the preset name plus the stack's output_stem values, dropped under output/img/<family>/..., and the workflow is embedded in the PNG metadata - drag the image back into ComfyUI and the graph restores.

Install. ComfyUI Manager → search ComfyUI Flakes → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/JeyzerMC/comfyui-flakes

then restart. The pack itself only needs PyYAML (already in ComfyUI core); the Impact Pack and any upscale model are extras you install on top.

If you want to keep your own sampler in the graph, don't use this node - split the flake_data and wire your own KSampler. But if you just want presets to render, Flake Generate is the least-friction endpoint the pack offers.