FLUX.1 Fill Dev LoRA · Edit

No purpose-trained open inpainting checkpoint has shipped since FLUX.1 Fill dev in November 2024. Everything since is either a fine-tune of it or a general model with a mask bolted on - and the difference shows exactly where it matters, at the edges of what you painted. Which is why a node that gives you Fill dev, a mask, and up to three LoRAs for six cents an image is worth knowing about before you try to bend an edit model into the job.

When Fill is the right tool

The honest 2026 rule from the KB's controlnet reading: one condition, one shot, describable in a sentence - hand it to an edit model. Fill is what you reach for when the mask is the spec. Product shots, removing an object cleanly, repainting a background behind a preserved subject, extending a canvas. Fill dev was trained on image-fill and image-extend, so it sees outpainting as a normal task rather than a hack.

Its defenders are blunt about why: it's trained specifically for inpainting, the edge blending and context awareness beat a masked general model, and you can't replicate that with strength dials. FLUX.1-Fill-dev-OneReward (ByteDance's RL fine-tune) is the only meaningful new option since, and people report it working about the same.

The inputs

mask_url is required here - not optional like on the Dev LoRA edit node. Black and white, source size, white areas get redrawn. If those match, this node is predictable; if the mask is cropped or scaled wrong, you'll get a plausible image with the mask misaligned and no error telling you why. Check dimensions before you blame the model.

prompt should describe what belongs in the white area, in the light and style of everything around it. Not "remove the bin" - describe the wall that should be there instead. The KB's warning about grounding-based masks applies: a tight mask needs the prompt to describe the surroundings, or the model invents clutter to fill the space.

resolution is 1k or 1.5k on the long side, aspect kept from the source. num_outputs gives 1 to 4 images per run.

Then the field that catches everyone out: guidance_scale defaults to 30, range 28 to 35. That is not a typo and it is not the 3.5 you're used to from Dev. Fill runs a different guidance regime; treat the default as the setting and stop touching it. This is what people mean by "read the tooltips" - cranking it to 3.5 because that's what Dev uses will hand you a mush of noise.

steps is 1-50, default 28. seed starts fixed. loras accepts up to three weights trained for FLUX.1 dev, each with a strength, via a PopNow LoRA node - local .safetensors (uploaded once, 12-hour cache), Hugging Face repo id or page, Civitai page, or direct link, with HF_TOKEN / CIVITAI_API_KEY read from the environment for gated files. Style LoRAs transfer fairly well to Fill since it shares dev's backbone; expect to run them a little lower than you would on Dev.

Outputs: images, and task id - the string worth pasting into the API console when a run goes sideways.

Cost and install

6.3 credits per image, about six cents, LoRAs included.

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

Or Manager → PopNow → install → restart. Create a key at popnow.ai/api/console/keys, then Settings > PopNow or POPNOW_API_KEY, then one more restart so the node list refreshes from your account. No GPU, no downloads, no dependencies - the 12B Fill checkpoint stays on PopNow's disks, not yours.

Where people get burned

Uncomposited passes. This is the oldest inpainting mistake and it's worth repeating: three or four Fill passes stacked without compositing the original back in turns an image to trash. If you're doing multiple corrections, keep the source and composite each masked result over it.

Then the guidance range. Its floor is 28. People see a slider that won't go below 28 and assume the node is broken; it's the model telling you what it accepts.

And remember what the mask is for. Fill repaints what you painted. If you want "change the shirt colour", an edit-model node is cheaper and more forgiving than painting a shirt-shaped mask.