FLUX.1 Schnell LoRA · Edit

0.9 credits an image. Nine tenths of a cent. At that price the calculation you make at the queue button changes completely - you stop asking "is this variant worth it" and start queueing six. Schnell is FLUX.1's step-distilled, Apache-2.0 branch, faster and a little rougher than Dev, and for exploratory edits it's the right tool for the same reason a pencil is the right tool before paint.

What Schnell is, and isn't

Schnell came out of Black Forest Labs' first release with the loosest licence of the FLUX.1 line and the most aggressive distillation: a handful of steps instead of Dev's twenty-something, quality approaching Dev at four steps and degrading past it. The KB's read on why BFL was happy to make it Apache 2.0 is unkind and probably correct - it's heavily distilled, so nobody was going to build a better model on top of it.

Practically: it's the lightest of the FLUX.1 family. Fine for a quick restyle, a composition check, a batch of thumbnails, a "which of these six directions is the good one" pass. Not what you use for a final deliverable if you have a cent more to spend.

There's also a well-known side effect of distillation worth flagging: face variety. Dev and Schnell have both been reported as producing less varied faces than they should, and Schnell is the worse of the two. Batch variety is not its strength.

The fields

image_url is the source; the result keeps its aspect ratio.

prompt - describe the whole image you want back, not just the change, trigger word first if your LoRA has one.

strength defaults to 0.8, range 0 to 1. On a distilled model that's a fairly strong push; 0.8 will re-render rather than nudge. Come down to 0.5 for a restyle that keeps the photograph.

mask_url is optional here, and its presence is the nice surprise on the cheap node: black and white, source-sized, white gets redrawn and black is kept. It's a general model with a mask rather than a trained inpainting checkpoint, so edges are less faithful than FLUX.1 Fill, but for "get that thing out of frame" at under a cent it's very hard to argue with.

resolution is 1k or 1.5k long side. output_format is jpeg, png or webp. seed starts fixed.

No steps, no guidance_scale. Schnell doesn't take them - it's distilled to run in a handful of steps with guidance baked in - and the node honestly reflects that instead of showing you dials that do nothing.

loras takes up to three weights, each with a strength, from a PopNow LoRA node: local .safetensors in ComfyUI/models/loras (uploaded once, 12-hour cache), Hugging Face repo id or page, Civitai page, or direct link. Gated files read HF_TOKEN / CIVITAI_API_KEY from the environment, never from a widget - the pack's reasoning being that widget values get written into every workflow and PNG you share.

Outputs: images and task id.

Mechanics

The node is generated from black-forest-labs/flux-1-schnell-lora/edit's published schema. It uploads your image and mask to popnow:// file URIs, quotes the job before holding credits, submits with an idempotency key so retries can't double-bill, polls, downloads, decodes into an IMAGE batch. The status line tells you what it cost.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Manager search PopNow also works. Restart, key from popnow.ai/api/console/keys into Settings > PopNow or POPNOW_API_KEY, restart again. Nothing to download, no GPU, no dependencies - which is a little ironic for the one model in this pack you could genuinely run on almost anything locally.

Where people get burned

The LoRA mismatch is the real one. Schnell is a separate base from Dev, and the overwhelming majority of FLUX LoRAs on Civitai are trained for Dev. They'll load, and then underperform, and you'll conclude the node is bad. Check the base model. Also: Schnell's distillation is exactly why people say it's poor for training - if you fine-tuned on Schnell, expect Dev LoRAs (or anything else) to be a dice roll.

And the usual: a batch on a single-image socket sends one picture, so feed your edits one at a time if the order matters.