PVL Flux.2 Camera Control (fal.ai)

Camera control is one of those "wow" features that usually demands a ControlNet, a camera-motion LoRA, or a video model. This node gets you a version of it with none of that: it wraps fal.ai's fal-ai/flux-2-lora-gallery/multiple-angles endpoint, which takes one image plus camera angles and returns the same scene rendered from different viewpoints. It's a member of the pvlprk/comfyui-pvl-api-nodes pack - the "ComfyUI Assistant Node" grab bag - and like every fal node in it, it runs on the cloud and costs credits per generated image.

Where this shines: turnaround sheets for character concepting, product shots where you want the object seen from three sides, or just "what does this composition look like from the left?" The Flux 2 family (per the KB) is BFL's most capable line, and its LoRA gallery ecosystem on fal is what supplies the multi-angle behavior here. One input image in, a handful of angled renders out.

How it works

The node encodes your image to a base64 data URI and submits a request to the fal queue API (https://queue.fal.run/fal-ai/flux-2-lora-gallery/multiple-angles), then polls until it completes and fetches the result. It fans out: set num_images and it runs multiple angle requests, with retries (retries, default 2) and a per-call timeout_sec (default 30) - which is low for cloud inference, so don't be surprised if a slow day triggers a retry.

The inputs that matter

• image is required - the single source image the camera moves around.
• horizontal_angle (degrees, -360 to 360) and vertical_angle (-90 to 90) are the orbit controls. zoom (0-10, default 5) pushes in or pulls out.
• image_size defaults to auto (keep input resolution), with presets like 1:1, 9:16, 16:9, or custom (then width/height, 512-2048, apply).
• num_images (1-8, default 1) is how many angled variants you get. With it > 1 the node fans out into parallel requests.
• guidance_scale (2.5), num_inference_steps (40), lora_scale (1.0) tune the generation; seed of -1 means random.
• output_format (png/jpg/webp) and enable_safety_checker (off by default) round it out.

Output is a single IMAGE - with num_images > 1 you'll get a batch, so a Batch To Image-style node is useful if you want to inspect each frame separately.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/pvlprk/comfyui-pvl-api-nodes

Restart ComfyUI and set FAL_KEY in the launching environment (fal.ai/dashboard). Same story as every fal node in the pack: the requirements.txt drags in diffusers and friends you don't need for this one, but Manager installs it all anyway.

Where people get burned

• The 30-second default timeout is tight. Heavy Flux 2 renders can exceed it on a loaded queue; bump timeout_sec toward 120 and rely on retries.
• More images = more money, in parallel. num_images 8 fires 8 paid requests at once. Start at 1-2.
• It's a LoRA-gallery endpoint, not a physics simulator. Angles and zoom are the model's interpretation - don't expect perfect perspective consistency across the batch. For rough turnaround/preview work it's great; for pixel-accurate multi-view you'll want a proper 3D/render pipeline.