FLUX.1 Schnell LoRA · Text to Image

Six fields total, one of them optional. That's the entire node: prompt, aspect_ratio, output_format, seed, loras. No steps, no guidance, no resolution tier, no image input, nothing to tune. It's the closest thing in this pack to a text box with a render button, and at 0.9 credits an image - nine tenths of a cent - that's exactly what you want it to be.

Deliberately dumb, on purpose

FLUX.1 Schnell is BFL's step-distilled, Apache-2.0 model: run it in a handful of steps, get quality approaching Dev at the top of that range and degradation past it. Distillation ate the dials. There's no step count to argue about and no guidance scale to chase because there's no real CFG loop left in there to chase - guidance is baked in.

Compare that with the Dev node in this pack, which hands you steps and guidance_scale and invites you to spend an afternoon on 1.8 versus 3.5. Different jobs. This one is for when the question is "does this idea work", not "is this the best version of this idea".

The fields, such as they are

prompt - subject, composition, lighting. Trigger word first if a LoRA needs one.

aspect_ratio - the full nine, 1:1 through 21:9 and 9:21. Genuinely more flexible than the Z-Image Turbo node at a sixth of the price per image, which is not a sentence I expected to write.

output_format - jpeg, png or webp. seed - fixed by default, so re-queueing the same graph reuses the result rather than charging again.

loras - up to three weights, each with its own strength, from PopNow LoRA nodes. Sources: a .safetensors already in ComfyUI/models/loras (uploaded once, kept 12 hours), a Hugging Face repo id or page, a Civitai page, or any direct https:// link. Gated files need HF_TOKEN or CIVITAI_API_KEY set in ComfyUI's environment; the pack refuses to store tokens in widgets because those end up inside shared workflows and image metadata.

The honest caveat on that socket: most FLUX LoRAs in the wild are trained for Dev, and Schnell is a different base. They'll load. They'll also underperform, sometimes badly, and you'll blame the node. Check the base model on the download page before you conclude anything.

Outputs are images - an IMAGE batch into Save Image or whatever comes next - and task id, the identifier for the API console.

How the render happens

Every field is generated from black-forest-labs/flux-1-schnell-lora/text-to-image's published schema, so nothing here is invented by the pack. On queue it resolves any LoRA sources, builds the request, takes a quote before holding credits, submits with an idempotency key so a network retry can't spawn a second billed task, polls until done, downloads and decodes the file into a ComfyUI image batch. The status line ends with the charge.

Nine tenths of a cent a render, LoRAs included. A hundred-image exploration costs under a dollar, which is a different activity from the one where you think carefully before pressing queue. If you're building a prompt library or testing whether a LoRA's trigger phrase does anything, this node makes that cheap enough to be casual.

Install

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

Or ComfyUI Manager → search PopNow → install → restart. Create a key at popnow.ai/api/console/keys, enter it in Settings > PopNow or export POPNOW_API_KEY, and restart a second time so the node list is rebuilt from your account catalog. No model files, no extra packages, no GPU.

Where people get burned

Distillation shows. Faces vary less than you'd like across a batch, and fine detail - hands, small text, jewellery - is where the cheaper model falls over first. If you're judging a LoRA, judge it on Dev before you write it off on Schnell.

Second, the layers of "schnell" out there are confusing: base Schnell, the community Chroma derivative built on it, and various merges. Chroma prompts and LoRAs are not Schnell's. Read the model card.

Third, and this one is free advice: fixed seed plus one edited word is how you learn what a prompt does. Randomize only when you've stopped learning.