For saved masks from a different geometric warping pipeline, use the WorldForge research guide and local mask coverage checker to compare a full frame with a chosen subject region.
World in World, also written WorldinWorld, is a research workflow for re-filming an existing video from a new camera path. Its released warping-based path uses a frozen world model and visual evidence derived from the source. That is a different starting point from generating an unrelated scene from text.
This guide audits source revision 3933f142f4a312f211cea7159c13e2909718dcf6 and supplies an original native Blender freeze-and-camera study. We have not installed the runtime, downloaded weights, estimated model depth or run inference. Every example here is authored geometry and a frame ledger, not a World in World output.
Choose the released workflow before preparing a clip
The pinned official README identifies free-camera re-cinematography, bullet time and video editing as released warping-based paths. It lists other research capabilities separately on its roadmap. Do not assume that a paper result, a hosted interaction or a roadmap item is part of the released command-line workflow.
The documented own-video route prepares frames and embeddings, estimates depth, creates a camera trajectory and generates the new view. The source clip, depth, source camera and target camera are different inputs. A camera path alone does not supply the geometry of areas the source video never revealed.
The authors report testing on an 80 GB A100 with approximately 72 GB peak memory. Those are upstream setup figures rather than our benchmark or a measured universal minimum. Their release uses a 14B LingBot World base model; our existing LingBot input guide explains the foundation’s separate variants and camera conventions.
Inspect the original frozen-event camera comparison
Watch this clip and view its details
An orange ball moves through an original scene with a blue arch and a near marker. During the freeze, the left view keeps the source camera fixed. The right view translates the native camera out and back while the event remains held. Both views come from known 3D geometry, so they demonstrate the intended shot rather than a model’s reconstruction quality.
The comparison contains 93 authored frames and lasts 5.8125 seconds at 16 fps. It uses the first 93 entries of our explicit frozen-source ledger as a native planning reference. This does not establish the exact temporal behavior of a generated clip.
Open the native re-filming watch page or download the original study archive. The archive contains source and native target videos, proposed camera arrays and measurements. It is an inspection study rather than a prepared WorldinWorld model case.
Start from a source clip with a clear event
Watch this clip and view its details
Our unmodified native source contains 85 frames at 832 by 464. It lasts 5.3125 seconds at 16 fps. The camera remains fixed while the ball moves, making source motion and target camera motion easy to distinguish. The blue arch and near marker provide depth cues without relying on an AI-generated scene.
For your own pilot, choose an original or permitted clip with a readable subject and a modest requested camera move. Watch the complete source before choosing the freeze frame. A large unseen camera angle asks the model to invent more of the scene than a small shift around a well-observed subject.
Use the source watch page to review the original action. The source MP4 is an authored input, while the native target MP4 is a reference re-render from the known scene. Neither is an inference result.
Check padding trimming and the generated-frame count
The pinned preparation code first makes its deduplication, optional resampling and frame-selection decisions. Its length helper then chooses a prepared count of the form 16n + 5. It pads by repeating the last frame when the next such count is at most eight frames away; otherwise it trims to the previous supported count.
The free planner starts after those preprocessing decisions. For example, 77 selected frames pad to 85 with eight repeats. By contrast, 76 selected frames trim to 69, dropping seven, because padding to 85 would require nine. The planner does not decode a video, apply the duplicate detector or resample timestamps.
The source documents a generated count of prepared frames minus eight. Our 85-frame source therefore corresponds to 77 planned generated frames before any inserted freeze. That is arithmetic from the pinned path, not a measured output. The planner flags counts below thirteen as too short for its supported preparation check.
Count inserted freeze frames and the original occurrence
The pinned bullet-time helper inserts repeated copies of a selected frame and then retains the original occurrence in the remainder of the sequence. For a freeze at source frame 40 with 16 inserted frames, our ledger repeats frame 40 seventeen times in total.
The inserted window is frames 40 through 55 inclusive. The original source frame 40 then occurs again at prepared frame 56. The base prepared count of 85 becomes 101, and the planned generated count and trajectory length become 93. This distinction helps avoid making a held moment one frame shorter than intended.
Save the planner JSON with the chosen source frame, inserted count and exact index map. The default freeze review ledger describes this example. It is not a configuration file that can be substituted for the runtime’s prepared frames, latents or embeddings.
Match the camera coordinate and calibration conventions
The pinned trajectory source expects OpenCV camera-to-world poses relative to the source camera at frame zero, with the first pose equal to identity. It saves one pose per planned generated frame and repeats the final pose for a short dwell. Our proposed trajectory contains 93 matrices and returns to the initial camera after the freeze-window move.
The configuration stores intrinsics at an 832-by-480 reference even though working frames are 832 by 464. Its trajectory utility scales fy and cy by 464 divided by 480 for projection at the working resolution. Treating the stored row as already calibrated for the shorter image would change vertical projection.
Our Blender camera uses the same reference values with matching lens shift and pixel-aspect settings. The native script compares a known point’s camera-matrix projection against Blender’s own projection for each frame. The independent inspector checks that agreement, finite arrays, rigid rotations and first-pose identity. It does not validate the upstream warping tensors, estimated depth or generated scene.
Prepare and review an actual runtime case separately
- Inspect the current environment and model requirements for the exact released route you intend to run.
- Prepare your source frames and embeddings, recording deduplication, resampling and selection settings.
- Estimate and inspect depth for your own source. A plausible-looking RGB frame does not prove a usable depth map.
- Generate the intended camera trajectory with the correct count, first-pose identity and intrinsics convention.
- Retain the returned video and the evidence, mask and metadata files produced by the runtime. Inspect both normal playback and the regions exposed by the new view.
The release distinguishes a full subject prompt from a scene-only prompt for later chunks. Follow the actual example configuration for the route you choose rather than treating every prompt as interchangeable. We have not executed that pipeline with the supplied study.
Review identity, contact, background geometry, newly exposed surfaces and the transition out of the freeze. A smooth camera path does not prove that unseen parts of a subject are correct. Keep the test tied to a narrow shot goal and compare a modest move before requesting a much wider re-shoot.
Reproduce the native study and inspect the pinned sources
The original Blender builder, source project, frozen camera project and measurement record expose the authored scene. The independent inspector uses NumPy and local ffprobe without importing WorldinWorld.
The companion comparison asset builder and frame ledger source reproduce the comparison PNG sequence using Bun and Sharp after the native study is complete. The encoding note records the separate FFmpeg steps.
The pinned audit records nine small documents, including the base model card at revision 5c33dd40b213598c418fd25bff30fdbd23fd38a7. Its read-only audit script fetches documents only. It downloads no weights and executes no upstream runtime code.
The repository states CC BY-NC-SA 4.0 and says third-party weights keep their own terms. Review the actual components for your intended use. This source availability is not an unrestricted commercial-use grant.
For original character and cutout motion, use the puppet and cutout animation guide. For hosted media creation and connected story coverage, inspect supported inputs in Clout Image and Video or plan scenes in AI Cinema. This does not claim that Clout hosts WorldinWorld.
