arkennemasis fal · Flux Control LoRA Depth Image to Image · $0.04/MP

The depth sibling of the Canny image-to-image node, and the difference matters: depth is the loosest of the structural conditions and image-to-image is the loosest of the generation modes, so this pairing gives the model the most freedom of anything in the control family. Composition from a depth map, appearance driven by a real photo, and a denoise dial deciding who wins.

That's either exactly what you want or a recipe for mush, depending on whether you have a specific change in mind.

The two references and the dial

control_lora_image is required and is your depth source. One absence worth knowing before you plan around it: unlike the plain depth node, this variant has no preprocess_depth toggle, so nothing estimates depth for you - what you connect is what the model treats as the scene's geometry.

image is required here too, and it's the initial image: the real pixels the render starts from.

strength defaults to 0.85, and the tooltip is explicit - 1.0 completely remakes the image, 0.0 preserves the original. Practical bands:

• 0.3–0.5 - colour and light shift, subject intact. The "same shot, better mood" range.
• 0.5–0.75 - materials and surfaces change; you keep the object, not its texture.
• 0.85–1.0 - the source is a reference for layout and palette only.

control_lora_strength (default 1.0) is the separate grip of the depth condition. Keeping the two apart is what makes this node useful: low control_lora_strength with high strength means "new everything, whatever shapes it lands on", while high control at low strength holds the geometry hard while changing little.

Where it earns its keep

The workhorse case is interiors. One depth map of a room, one photograph as the palette and content reference, and you generate the same space in a different finish - oak to concrete, daylight to dusk - without re-modelling or re-lighting anything.

Second case: product and architecture visuals where you need five variants of one scene for a review. Depth keeps the arrangement honest across all five, which is the only way variants look like the same room rather than five different rooms.

Skip it when the thing you care about is a specific edge - a logo, a machine part, a building facade. That's Canny's job; depth will drift the outline and you'll spend the afternoon chasing it.

The rest of the settings

Then: prompt, image_size (square_hd by default, which does not preserve your source aspect ratio; use custom and enter dimensions), num_inference_steps 28, seed -1, loras (JSON, normally empty), guidance_scale 3.5 (Flux guidance, not SDXL CFG), num_images, enable_safety_checker, output_format defaulting to jpeg.

Outputs

images, image_paths (saved under output/fal/fal-ai_flux-control-lora-depth_image-to-image/), has_nsfw_concepts, timings, seed, prompt, and info with the request id, cost estimate, saved files and fal's raw response.

Install

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

Restart. ComfyUI-Manager users can do Install via Git URL instead. The pack's only declared dependency is httpx - used by its Codex nodes, not the fal ones - and nothing is downloaded locally because the model runs on fal's hardware.

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

in the .env beside run_nvidia_gpu.bat (portable) or in your ComfyUI folder. No key box exists on the node, which is the point: a key typed into a widget gets stored in every workflow you export.

Cost and caching

$0.04 per megapixel of output - 1024×1024 ≈ 4¢, a 2048-wide render ≈ 16¢ - tracked live by the node's price badge. max_cost_usd (default $20, set 0 for no cap) refuses the run before upload if the estimate crosses it.

reuse_identical_run defaults on. Same settings, same picture bytes, same files still in output/fal means you get the earlier result back instead of a new charge. While you're sweeping strength, remember each value is a distinct run and a distinct 4¢ - cheap, but figure out your sweep range before you start rather than after.

What usually goes wrong

The output is a different room. With control_lora_strength at 1.0 and strength at 0.85 you've given the model both references and a lot of rope. Reduce one of them.

The output is the same room with nothing changed. Same cause, opposite end: depth alone under-constrains materials, so if the prompt is vague you get a re-render of the original. Name the surfaces.

A batch on either image socket. Both take a single picture; the node stops and tells you to pick a frame instead of silently choosing.