WorldCrafter camera control: Base weights, adapter and LoRA

WorldCrafter-Base is a set of weights for TencentARC's inference code, not a hosted service: the README describes base transformer weights with their matching camera adapter and LoRA, tagged image-to-video and camera-control. Sume has no adapter input. To steer the camera on Sume, describe the movement in the video prompt and start from a still with first_frame.

Weights facts are from the README, read 2026-10-01; Sume facts are from Video generation.

What does the WorldCrafter-Base README say?

It says the directory holds "Base transformer weights and their matching camera adapter and LoRA for the WorldCrafter inference code". Base is placed beside a WorldCrafter-Fast directory and reads the shared repencoder/, text_encoder/, tokenizer/, vae/ and scheduler/ folders from it, and the Base-specific transformer/ and adapter/ folders must stay together. The run command is python inference.py --model-type base. The linked paper is titled "WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory". The README does not describe how the camera adapter is driven, so I do not either.

How do I steer the camera on Sume?

Sume's docs advise including "details about motion, camera angles, lighting, and scene composition" in a video prompt. Name one camera move per shot, such as a slow dolly in or a static frame, and keep the subject description short. Examples are in AI video camera movement prompts.

What if I need an exact camera path?

A text prompt gives direction, not coordinates. If the README's adapter gives you the control you need, run it on your own hardware. If a described move is enough, check GET /v1/videos/models for the model's supported_frame_images and durations, then iterate on a short clip first. Billing is reserved on submit at provider list x 1.25, so short test clips are cheap to repeat.