arkennemasis fal · Flux General Image to Image · $0.075/MP

Flux Dev image-to-image is $0.03/MP and takes a source photo and a prompt. This node is the same base model at $0.075/MP, and what you're paying the extra four and a half cents for is the ability to attach a LoRA, a ControlNet, an IP-Adapter or an EasyControl to that img2img pass.

If you don't need one of those, close this page and use ArkFal_fal_ai_flux_dev_image_to_image. That's not a knock - it's the whole selection criterion for this endpoint.

What's here

image is required: the initial image, and the one that sets the colour and content. strength defaults to 0.85, defined exactly as you'd expect - 1.0 completely remakes the image, 0.0 preserves the original. The tooltip adds "only used if the image_url is provided", which on this node is always.

Then the reason you're here, all as optional JSON fields:

• loras - extra LoRAs merged into the generation.
• control_loras - LoRAs that consume a control image.
• controlnets and controlnet_unions - each with the tooltip's own warning: only one is supported at the moment.
• ip_adapters - IP-Adapter entries.
• easycontrols - EasyControl inputs.

Empty leaves each out, and empty is the correct value until you have something specific. The gotcha that surprises everyone the first time: these take URLs fal can download from. A LoRA in your local models/loras folder can't be reached. It has to be on Hugging Face, your own storage, an S3 bucket - somewhere with a public URL.

The knobs inherited from the general endpoint

Because this is the general pipeline, you also get the sampling and conditioning controls that Flux Dev img2img doesn't have:

• guidance_scale 3.5 and real_cfg_scale 3.5 with use_real_cfg (false). Enabling real CFG applies classical SD-style CFG, and the tooltip warns it increases both generation time and price for the same output.
• use_cfg_zero (false), which applies CFG-Zero-star initialisation - the arXiv id is in the tooltip.
• negative_prompt with the NAG family: nag_scale (3), nag_tau (2.5), nag_alpha (0.25), nag_end (0.25). This is the honest way to use negatives on a guidance-distilled model - the tooltip says NAG is the default path for the negative prompt, and the community advice is not to push nag_scale past 3.
• reference_image with reference_strength (0.65), reference_start (0.0) and reference_end (1.0), plus six optional fill_image_* sockets.
• scheduler (euler), base_shift (0.5), max_shift (1.15), use_beta_schedule (false).
• image_size (square_hd default, custom with width and height), num_inference_steps 28, seed -1, num_images (capped at 1 - the tooltip notes it's fixed for streaming), enable_safety_checker, output_format png.

Outputs

images, image_paths (files land in output/fal/fal-ai_flux-general_image-to-image/), timings, has_nsfw_concepts, prompt, seed, info - request id, estimate, saved files, fal's full response.

Install

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

Restart ComfyUI, or ComfyUI-Manager → Install via Git URL. The pack's only declared dependency is httpx and that's for the Codex nodes; the fal provider uses Python's standard library. No weights download and no GPU load.

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

in the .env next to run_nvidia_gpu.bat (portable build) or your ComfyUI folder - never a node widget, so the key can't be saved into a shared workflow.

Cost discipline

At $0.075/MP: 1024×1024 ≈ 7.5¢, 1536×1536 ≈ 18¢. Add use_real_cfg and it climbs for the same picture. max_cost_usd (default $20; 0 disables) blocks the run before anything is uploaded if the estimate is over. The node's price badge follows your image_size and num_images.

reuse_identical_run (default on) returns the previously paid result when the settings and image bytes match - genuinely useful while you're testing the wiring of the rest of the graph, and confusing if you forgot it's there. max_concurrent (default 1) keeps paid fal calls serialised across the whole run, which is the sane default when each one costs real money.

Troubleshooting

It says "is not valid JSON". One of the six JSON fields is malformed. Validate it somewhere else first; the node can't tell you what the field was expecting.

Nothing changed. strength at 0.85 with a vague prompt gets you a re-render. Say what should differ.

My LoRA did nothing. Check that the URL is publicly fetchable. Fal's downloader doesn't have your session or your API key for arbitrary hosts.