FLUX.2 Dev LoRA · Edit

The Flux 2 LoRA ecosystem is thin, and the reason has nothing to do with taste. Training them takes an RTX 6000 Pro and many hours - one commenter in the KB put it as plainly as it can be put: "you need an RTX 6000 PRO to train them", against Z-Image Turbo where you can knock out a thousand steps an hour on a 3060. Accessibility decides ecosystems.

So if you have, or find, a FLUX.2 [dev] LoRA, this node is how you run it without owning the machine that made it. Same multi-reference editing as the plain Flux 2 Dev node - one to three source images - with up to three LoRAs and their strengths on top, for 5.4 credits an image.

What it does that the plain node can't

image_urls is the autogrow socket: one IMAGE in, another socket out, up to three sources, ordered the way your prompt refers to them. prompt describes the finished picture - what changes and what must stay the same - and the LoRA stack rides along, which lets you hold a look across a whole batch of edits instead of re-describing it in every prompt.

That's the real use case: a style or character LoRA plus a multi-reference composition. Flux 2 Dev's strength was always composition and understanding; a LoRA gives you the consistency that a base model with no training on your subject can't.

loras takes up to three files trained for FLUX.2 [dev], each with its own strength, connected from PopNow LoRA nodes. source accepts a .safetensors in ComfyUI/models/loras - uploaded once and reused for 12 hours, so iteration costs nothing in bandwidth - a Hugging Face repo id or model page, a Civitai page, or any direct https:// link. Gated files read HF_TOKEN or CIVITAI_API_KEY from ComfyUI's environment. The pack never puts tokens in widgets, on the grounds that widget values are saved into every workflow and PNG you share.

The rest of the fields

aspect_ratio defaults to auto, which keeps the shape of the first source image at about one megapixel. Override it with 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 or 9:21, and then resolution - 1k or 2k - sets the size tier for that framing.

seed starts fixed for the same reason it does everywhere in this pack: queueing twice shouldn't be two charges.

No mask, no strength, no steps, no guidance scale. Flux 2 Dev's editing is instruction-driven, and the field list reflects that honestly. If you want a denoise-style strength dial, you want a Flux.1 node instead; if you want to merge people into one scene, this is the family that does it.

Outputs are images and task id - the latter being what the API console and support will ask for.

What it costs, and how it runs

5.4 credits per image flat, 1 credit = $0.001, LoRAs included. Nothing scales with reference count, so three sources with three LoRAs is the same price as one of each.

The node is generated from black-forest-labs/flux-2-dev-lora/edit's published schema. On run: resolve and upload LoRA files, upload source images to popnow:// URIs, quote the job before holding any credits, submit with an idempotency key so a retried request can't create a second task, poll, download, decode into an IMAGE batch. The node's status line reports the estimate, the queue, and finally what you were charged.

Install

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

Or Manager → search PopNow → install → restart. Create a key at popnow.ai/api/console/keys, add it under Settings > PopNow or as POPNOW_API_KEY, restart once more. No model files. No GPU. No 53 GB download, which - for Flux 2 specifically - is the entire reason to be here.

Where people get burned

Base-model mismatch, and it's worse here than anywhere else in this pack. The Flux 2 line has Klein 4B, Klein 9B, base and distilled tiers, and now Flux 3's own branches. A LoRA trained for FLUX.2 [dev] is what this node wants; a Klein LoRA is a different model. Check before you spend a run.

Second: expectations of a deep library. There isn't one. If your workflow depends on a specific Flux 2 style LoRA you can't find, that's the ecosystem's state, not this node's fault - and it's why a lot of people still edit on FLUX.1 dev, whose LoRA shelves are enormous.

Third: reference ordering. "The woman in the red coat" has to map to a specific connection order, and the prompt is the only place that mapping exists.