FLUX 3 Klein Base LoRA · Edit

Here's the interesting one. LoRA models are the reason the local crowd puts up with 20GB checkpoints - and this node puts a LoRA slot on a hosted edit model. You supply a .safetensors you trained (or downloaded), it gets loaded server-side against FLUX 3 Klein Base, and your instruction edit runs with the LoRA's style or subject baked in.

If you've been training Klein LoRAs - and plenty of people have, since Klein's newer VAE is what made the fine-tuning front move away from Z-Image Base - this is how you use one without owning the hardware it wants.

How the LoRA gets there

The loras input is a POPNOW_LORAS socket, and you fill it with the pack's PopNow LoRA node. That node's source accepts four things:

• a file in ComfyUI/models/loras, e.g. my_style.safetensors - uploaded once, up to 2 GiB, and reused for 12 hours
• a Hugging Face repo id or page
• a Civitai model page
• any direct https:// link (or an existing popnow://f/... URI)

PopNow imports the file once through its own API, so the second run with the same LoRA doesn't re-upload. You can chain several PopNow LoRA nodes into the one socket - the model takes up to three, each with its own strength, which is roughly the same stacking you'd do locally.

One design decision I actually respect: Hugging Face and Civitai tokens are read from HF_TOKEN / CIVITAI_API_KEY in ComfyUI's environment and never from a widget. Widget values get written into every workflow you share and every PNG you post. The pack refuses to make that mistake.

The inputs that matter

prompt - same rules as the non-LoRA version, with one addition the tooltip spells out: put the LoRA trigger phrase first. "sksstyle, put the jacket from image 2 on the woman in image 1, keep her face and the background" is the shape. Refer to sources as image 1, 2, 3.

image_urls - one to three references, in the order your prompt names them. Autogrow: connect more, get more sockets.

loras - optional, but it's the point of the node. Wire the LoRA node in and set its scale; a strength that's too high is the usual cause of a burnt, over-saturated result, and the fix is turning the dial down rather than rewriting the prompt.

resolution (1k or 2k) and aspect_ratio (auto keeps the first source's shape) are the two settings that shape your bill and your framing. seed starts on fixed so re-queueing the same graph doesn't quietly bill you twice; switch control_after_generate to randomize for a new take. output_format is jpeg by default, png if the result is going further down the pipe.

Outputs

images for Save Image or downstream nodes, and task id for the API console or support. Keep the id when you're debugging a LoRA that "isn't doing anything" - it tells you exactly which render to look at.

Installing it

Manager → PopNow → Install → restart. Or:

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

Add a key from popnow.ai/api/console/keys in Settings > PopNow (or via POPNOW_API_KEY), restart again so the catalog loads, and if you're importing gated weights set HF_TOKEN or CIVITAI_API_KEY in the environment before you queue. There are no model downloads for the pack itself - the only "model" you install is your own LoRA file.

Where people get burned

Architecture lock-in is the big one. A LoRA trained on SDXL or Flux.1 will not do anything useful here; the tooltip is explicit that these must be trained for FLUX 3 Image Klein, Base or distilled. If your edit comes back looking exactly like the base model, check that first.

A 2 GiB upload cap is real, and a failed import can look like a silent no-op. Watch the node's status line - it shows the estimate, then queue state, then the final charge, so you can see whether it actually picked your LoRA up.

And price: this node costs more per image than the plain Base edit, because it's Base plus your weights. The rate badge on the node updates as you change resolution, and every run is quoted before anything is held. Set Max credits per run in Settings and let it be your seatbelt.