FLUX.1 Dev LoRA · Text to Image

FLUX.1 Dev is a two-year-old model that refuses to die, and the reason is boringly practical: an enormous LoRA library, an aesthetic everybody recognises, and settings nobody has to think about. Ask the corpus what people actually generate with and "Flux Dev" still shows up in hundreds of threads while Flux 2 shows up in thousands of mentions and very few LoRAs - because training Flux 2 required a card most people don't own.

So this node is the value play: the model with the most community weights attached to it, running on hardware you don't have, with a socket for three of those weights.

The knobs, and the one trick worth knowing

prompt is subject, composition, lighting, with any LoRA trigger word first.

guidance_scale defaults to 3.5, range 0-20, and it is not CFG - FLUX.1 Dev is guidance-distilled, so there's no negative prompt and no CFG pair. Here's the tip that ages well: locals have been running Dev at lower guidance, around 1.8-2.0, specifically to get less "Fluxmaxxed" faces out of it. That uniform, over-lit, slightly waxy look the model is famous for softens when you stop pushing guidance. If your LoRAs on this model keep producing the same face, try 2.0 before you blame the LoRA.

steps is 1-50, default 28. The conventional local range is 20-30; past that you're paying latency for noise you won't see.

aspect_ratio gives you nine shapes including 21:9 and 9:21, which is more generous than the Z-Image nodes in this pack. output_format is jpeg, png or webp. seed starts fixed so queueing twice is one charge, not two.

loras takes up to three files trained for FLUX.1 dev, each with a strength, fed from a PopNow LoRA node. Sources it accepts: a .safetensors already in ComfyUI/models/loras (uploaded once, reused 12 hours), a Hugging Face repo id or page, a Civitai page, or any direct https:// link. Gated downloads read HF_TOKEN or CIVITAI_API_KEY from the environment, deliberately not from a widget.

Flat 2.7 credits an image - about $0.027 - LoRAs included. That's the cheapest way to run somebody else's FLUX.1 fine-tune stack without downloading 12 billion parameters.

Mechanics

The node is generated from black-forest-labs/flux-1-dev-lora/text-to-image's published schema, so its fields are the model's actual fields. On queue: resolve and upload LoRAs, quote the job before any credits are held, submit with an idempotency key (retries reuse it, never a second task), poll, download, decode into an IMAGE batch. Outputs are images and task id. The node shows the quote estimate, then the final charge.

Install

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

ComfyUI Manager, search PopNow, is the same thing with fewer keystrokes. Restart, get a key at popnow.ai/api/console/keys, paste it into Settings > PopNow or set POPNOW_API_KEY, restart again so your account's catalog is read. Nothing else installs - no checkpoints, no Python packages, no GPU. If you had the old single-file popnow_api.py node, delete it; workflows still load.

Where people get burned

Fine-tunes are not always LoRAs. Sites are full of merged checkpoints for this model, and those don't load as LoRAs - if your file is 11 GB, it's a checkpoint, not a LoRA, and the import will do nothing useful. LoRA files for Flux Dev are typically 20-200 MB.

Second, base confusion cuts both ways here: a Schnell LoRA and a Dev LoRA are not interchangeable, and the node doesn't check. Read the Civitai page.

Third, the plastic look. Guidance at 3.5 plus a realism LoRA at strength 1.0 frequently produces the same model-pretty face; drop guidance, and consider stacking a second LoRA at 0.3-0.5 rather than ever higher strengths.