arkennemasis fal · Flux 2 LoRA Gallery Multiple Angles · $0.021/MP

This is the most interesting node in the fal LoRA Gallery, because it's the only one with numbers instead of vibes. Give it one photo of an object and three dials - horizontal angle, vertical angle, zoom - and it renders the same object from a position you asked for. Product turntables, character sheets, game asset turnarounds; $0.021 per processed megapixel.

How it works

The endpoint is flux-2-lora-gallery/multiple-angles: Flux 2 with a camera-control LoRA. The node uploads your reference to fal's CDN, submits the angles as part of the request, and saves results under output/fal/fal-ai_flux-2-lora-gallery_multiple-angles/. No local model, no ControlNet, no clever node graph - the conditioning is baked into the LoRA.

That's the strength and the limit. You're not steering a camera in 3D space; you're asking a model to imagine the same object from elsewhere, and it will invent whatever you couldn't see.

Inputs

image_1 is the object; sockets run to image_6 and batches are sent in socket order. Single-object, clean-background references work best - the LoRA has to work out which thing is the subject.

The three camera controls are the reason to use this node:

• horizontal_angle (default 0) - rotation around the object. 0° is the front, 90° the right side, 180° behind, 270° the left, 360° back to front.
• vertical_angle (default 0) - camera height. 0° eye-level, 30° elevated, 60° looking down from above.
• zoom (default 5) - 0 is a wide, distant shot, 5 is a normal medium shot, 10 is a close-up.

lora_scale (0–2, default 1) is the effect strength - the fidelity dial. Lower it if the object starts mutating between angles. guidance_scale (default 2.5) and num_inference_steps (default 40, range 4–50) are the Flux 2 generation settings; acceleration is none or regular. image_size defaults to square_hd, which is exactly what you want for a turnaround - keep it square and the set lines up. num_images (1–4) per call, seed at -1, output_format and the safety checker as usual, with flagged output returned black.

max_cost_usd caps the run pre-upload, reuse_identical_run makes an identical repeat free, extra_json passes through unsupported fal fields, and max_concurrent limits how many paid fal calls run at once.

Wiring it into a turntable

There's no is_list output here - one call, one angle. For a real turnaround you loop the node, one instance per angle, and the pack's own utilities are built for that: a run folder to keep each run's outputs together, and a contact-sheet node to see every still of a run on one board. If you're driving it from a spreadsheet of angles, that's the intended shape.

Outputs

images is the rendered angle. image_paths, seed, prompt and info are the trail - info carries the request id, the estimate and fal's full answer.

Install

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

Restart, then:

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

in the .env next to run_nvidia_gpu.bat. Standard library only - no pip install for the fal nodes.

Where people get burned

Consistency between angles is the whole promise and the usual disappointment: a shiny or patterned object will have its details quietly re-invented, so back-of-pack text and logos drift. Generate the angles you need in one session rather than across days, and consider a consistency LoRA if the object is a character.

Second, the angles are degrees you type, not a slider you scrub, so you find the rotation you want by generating a few - and at $0.021 per processed megapixel with your input counted, an eight-angle set of a 1MP object is roughly $0.34. Third: it renders one object at one angle per call, so a "turntable" is a loop, not this node. Budget accordingly.