ausboss
cecffcc592
The camera stopped at every key: each stretch between two keys eased in and out on its own. Each value now follows a monotone cubic through the keys, so it keeps its speed through a key, never runs past one, and only eases at the first and last key. The moved picture shook. The depth model guesses the lens again for every frame, the depth-scale smoothing kept the frame-to-frame wobble instead of removing it, and still things got a slightly different depth each frame. Every frame now uses the clip's median lens, each frame's median depth follows the smoothed line, and the depth of spots that hold still is averaged over nearby frames. With the camera held at one angle the guide now moves 0.02 px a frame between frames, against 0.28 before and 0.04 in the source clip.
4.1 KiB
Camera Move
Moves the camera of your clip. You steer it on the node's picture, the way you turn a model in a 3D viewer, and set keys on the time bar. What the old camera never saw comes out pink, and the CrossView-Warp LoRA for MiniMax H3 paints it in.
How to use it
- Load your clip, scale it to the size you will render, and run it through Load MoGe Model and MoGe Inference (core ComfyUI nodes).
- Connect MoGe Inference's geometry to Camera Move.
- Press Show my clip on the node. It runs the depth once and shows your first frame.
- Steer:
- Drag the picture to turn the camera round.
- Pinch or scroll to pull back or push in. Two fingers, or Shift-drag, pan.
- Tap a spot to circle round it.
- Top right: drag the axis ball to turn, or tap an axis to look from that side. Drag the magnifier to zoom and the hand to pan. The camera button puts you back on your own shot.
- Keys: click the time bar to put a key there, using the camera at that moment. Or drag the time bar, then move the camera, and a key appears there. The diamond adds or removes a key; the arrows either side jump between keys. Drag a key along the bar to change its time. ⏮ goes back to the start, ▶ plays, and the loop button plays it round and round.
- Run the workflow.
guidegoes to the H3 guide at frame 0.
The keys move the camera from where your clip's own camera was. If your clip's camera already moves (walks down an aisle, for example), that move stays, and your keys add to it. After the last key the camera holds that turn, but it still travels along with your clip.
The line under the picture turns amber when a key goes past where the LoRA was trained: 45° of turn, 30° up, 20° down. Past that it may leave pink.
Mouse and keys (like Blender)
Middle-drag turns, Shift + middle-drag pans, Ctrl + middle-drag zooms, and the wheel zooms. On the numpad: 4 6 8 2 turn, Ctrl with them pans, + and − zoom, 1 is the front, 3 the right side (Ctrl: left), 7 the top (Ctrl: bottom), 0 your own shot. Space plays and pauses and Home goes back to the start, while the pointer is over the node.
The wheel over the picture zooms the camera, not the graph. Everywhere else on the node it zooms the graph as usual.
Inputs and outputs
moge_geometry |
Depth from MoGe Inference, one per frame. |
camera |
The keys, written by the node's picture. You never type here. |
guide |
Your clip from the moved camera, pink where nothing was seen. |
holes |
White where the guide is pink. |
Technical details
- Each key holds a moment
t(0 to 1 over the clip), a turnazand a riseelin degrees,back(how far the camera pulls back, in distances to the pivot; below 0 pushes in), and a panpx,py(same unit). The camera passes through each key without slowing down, and only eases at the first and last key. Each value follows a monotone cubic through the keys (Fritsch-Butland slopes), so it never runs past a key: a value that turns back or holds at a key stops there. - The camera turns round the pivot: a spot you tap (saved as a place on one
frame,
pivot: {u, v, t}) or, by default, the nearest fifth of what sits in the middle of the first frame, never further than 3 m. R = Ry(-az) · Rx(-el),C = pivot − R·pivot + R·[px·d, py·d, −back·d]withd = |pivot|, in OpenCV axes (x right, y down, z forward).- MoGe's depth scale and lens wobble from frame to frame, which shook the moved picture. Every frame is scaled so its median depth follows a smoothed line (15 frames wide) through all of them, and every frame uses the clip's median lens.
- Each pixel is drawn as 4 points, nearest wins, and gaps mostly surrounded by picture (5 of 8 neighbours, two passes) are filled. Everything else stays magenta (1, 0, 1).
- The node is an output node so its picture can run it on its own. Each run writes up to 61 small frames (416 px on the long side) with their depth to ComfyUI's temp folder; the depth is a PNG with a log scale in red (high byte) and green (low byte), 0 for no depth.