arkennemasis fal · Flux LoRA Fill · $0.035/MP

Flux Fill is BFL's purpose-trained inpainting and outpainting model, and it's the last one that was actually trained for masks - nothing comparable has shipped since, which is why it's still the answer when an edit model's whole-frame rewrite isn't acceptable. This node exposes the LoRA flavour of it, and it goes one step further than a normal inpaint: you can hand it pictures to place inside the hole.

What makes it different

Most inpainters generate something plausible in the masked region. Fill can be given a reference element and told to fit that in - the six fill_image_fill_image_1…_6 sockets on the node are exactly that, and each can carry a batch. The tooltip is precise about ordering: socket 1 first, then socket 2, and within a socket, in order. That's how you tell it "put this lamp here and that rug there" instead of hoping.

The second thing to know is guidance_scale, which defaults to 30 here with a range of 28–35. If you've internalised Flux's usual 3.5, that number will look like a typo. It isn't - it's the Fill endpoint's own convention, and it's why trying to be clever with a low guidance value produces mush.

The inputs that matter

• image - the picture being edited. mask - a ComfyUI MASK; the node converts it to the RGBA convention fal expects (transparent where you painted) and resizes it to your image.
• prompt - tooltip says empty leaves it out, so a mask-only fill with no text is legitimate.
• paste_back (default true) - composite the original back over everything outside the mask. This is the setting that gives you bit-identical surroundings, so leave it on unless you specifically want the whole frame re-rendered.
• fill_image_in_context_fill (default false) - uses the provided fill image in context with the base image "to fill in more faithfully", and the tooltip warns it increases the price. This is the one to flip when the inserted element keeps landing in the wrong lighting.
• fill_image_use_prompt (default false) - use the prompt as well as the reference image.
• resize_to_original (default false) - turn on to get the exact original dimensions back. Handy when the endpoint's chosen size keeps nudging your framing.
• loras - multiline JSON of LoRA entries with weights and scale, merged together. Empty leaves them out.

Plus image_size / image_size_width / image_size_height, num_inference_steps (28), seed, num_images, enable_safety_checker, output_format, acceleration, and the pack's shared max_cost_usd, reuse_identical_run, extra_json and max_concurrent.

Outputs

images → preview or save. image_paths → output/fal/fal-ai_flux-lora-fill/. Then has_nsfw_concepts, prompt, seed, timings and info.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

or search comfyui-arkennemasis in Manager. Restart, then find it under arkennemasis/fal/Image/Flux. Keys go in .env next to run_nvidia_gpu.bat:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

Realistic problems

Guidance at 3.5. A common first mistake, carried over from local Flux habits. The Fill range starts at 28 for a reason.

Prices that jump when you connect a fill image. fill_image_in_context_fill explicitly costs more, and it's the kind of setting you enable for a hard composite and forget to switch off. Also note the whole image's megapixels are billed, not the masked area.

Mask and image disagreeing. The mask is resized to the image, so it lines up - but paint the region you want changed, not the region you want to keep.

Empty output region. If the fill produces something blank or nonsensical, check whether paste_back is fighting you, and whether a stray reference socket is carrying a placeholder tile. The pack treats placeholder tiles as "no image", which is the right behaviour and also means a connected-but-empty socket quietly does nothing.