FLUX.2 Dev LoRA · Text to Image

Flux 2 Dev has a LoRA problem that has nothing to do with quality: training one needs the kind of card most people don't own. The line going around when people asked why the Dev LoRA shelf stayed empty was blunt - "you need an RTX 6000 PRO to train them, and it takes many hours." That's the gap this node fills. You point it at your Dev LoRA, it runs on PopNow's 32B stack, and you pay about $0.032 per image instead of buying a workstation.

The cheapest part of the deal is the LoRA plumbing

The loras input on this node isn't a file picker. It takes a POPNOW_LORAS value, which comes from a second node in the same pack: PopNow LoRA. That node has a source field, a scale slider, and its own optional loras input so you can chain them. Most PopNow LoRA models accept up to three, and this one's tooltip says the same thing: up to three weight files trained for FLUX.2 [dev], each with its own strength.

source accepts four kinds of reference, and the pack handles them differently:

• A .safetensors in your ComfyUI/models/loras folder - uploaded once (2GiB cap) and reused for about 12 hours.
• A Hugging Face repo id like owner/repo, or a Civitai model page - PopNow downloads it server-side. Gated files need HF_TOKEN or CIVITAI_API_KEY in ComfyUI's environment.
• A direct https:// link to the weights.
• A popnow://f/... URI, which is what the first two turn into.

The environment-variable-only detail is worth internalizing rather than skipping past: widget values get written into every workflow you share and into image metadata. Tokens are read from the environment and nowhere else, on purpose.

Fields you'll actually touch

prompt is the whole image, and the tooltip's first instruction is the one people ignore: put the LoRA trigger word first. Flux 2's encoder is a Qwen3 LLM reading your text as an instruction, so sks_style, a narrow alley after rain, wet asphalt reflecting a red sign behaves better than the same words buried in paragraph three.

aspect_ratio and resolution are the standard set - nine ratios, 1k or 2k. seed defaults to fixed, so a re-queue is a cached re-roll rather than a surprise second charge. Both outputs are plain sockets: images to Save Image, task id to a text node if you want the handle for the console.

What it costs, honestly

Plain FLUX.2 Dev text-to-image runs about 1.74 credits per image ($0.0174). This LoRA lane runs 3.24 ($0.0324) - call it double. That's the price of the endpoints that will load your weights, not a penalty unique to Dev. If you're doing 30 iterations to dial in a style, do them on the cheap Klein nodes and spend the Dev money on the finals.

Install

Manager → search PopNow → Install → restart, then paste a key from popnow.ai/api/console/keys into Settings > PopNow and restart again. The git route:

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

No pip install, no model downloads, no CUDA version to argue about. What you do need is a ComfyUI new enough to have the V3 node API, since the whole pack is built on comfy_api.latest.

Where people get burned

Wrong-generation LoRAs. A FLUX.1 [dev] LoRA will not load on this endpoint no matter how confident you are. Flux 2 swapped the CLIP+T5 encoder pair for a Mistral-3 VLM and the tensors don't line up. If a Civitai page says "Flux.1 D", it's the wrong file.

Scale creep. The scale slider on the PopNow LoRA node goes to 4.0, but a style LoRA at 2.5 doesn't give you 2.5x the style, it gives you a burned image with a melted face. Start at 1.0, move in 0.1 steps, and remember you're paying per roll.

Silent import failures. Pointing source at a Civitai URL behind a login produces an import error rather than a hallucinated image, which is the good outcome - set CIVITAI_API_KEY (or HF_TOKEN for a gated HF repo) before you queue, not after.

Assuming the LoRA is the problem. Dev's weak point is not style, it's hardware adoption, so the LoRA ecosystem around it is thin and uneven. A mediocre result on a thinly-trained LoRA is the LoRA's fault, not the node's - test the same prompt with loras disconnected before you start tuning.