ArtCraft takes a still into motion by keeping the image as the reference and letting you direct the shot: composite the frame in 2D, place it in a scene for depth, set pose and camera, then hand the step to a model and review the take. Here is that sequence, with the cost and account realities stated rather than glossed over.
Written against the build verified on 2026-10-13.
Image-to-video rewards preparation more than prompting. A still with a clear subject, readable depth and no ambiguous edges gives a model something to move. A cluttered frame gives it something to guess at. Before opening the app, decide the one thing that should change: a camera move, a subject action, a lighting shift. Shots that ask for two of those at once usually read as a mistake.
Bring the still into the compositor and fix what will not survive motion: patches, edges, an unwanted reflection. Work you skip here shows up multiplied in every frame of the result.
Placing the element in a 3D scene gives you camera and depth to direct. This is where a flat image becomes a shot, because you can decide where the viewer is standing rather than accepting the framing the generation gave you.
Character pose direction exists precisely so that you are not hoping a text prompt lands the composition. Set the pose you want, then check the silhouette at the resolution you will render.
Multi-model selection is the practical part of the engine: different models are good at different things, and the unit of work is the step rather than a single vendor relationship. Pick for the shot, not for brand loyalty.
Run the image-to-video step with the composite as reference. Short first: a two-second take that is right beats an eight-second take that is almost.
Play it once at full size before deciding. Frames you would never notice in a still become obvious in motion, especially around hands, hair and text in the background.
| Step | Runs where | Cost behaviour |
|---|---|---|
| 2D compositing and cleanup | Local | No per-run cost |
| Scene layout and camera | Local | No per-run cost |
| Generation of stills | Model provider | Charged per image by your provider |
| Image to video | Model provider | Charged per second or per clip, the most expensive step |
| Finishing edit | PhotoCraft, local | No per-run cost |
The cheap step is the one people skip: fixing the reference still before generation costs nothing locally, while re-running a video step to fix the same problem costs real money and takes real minutes.
Export the clip and bring it into PhotoCraft for the parts a generation model is reliably bad at: keeping a logo legible, holding a colour grade across the shot, masking out a warped background, and cutting the last frames where motion usually falls apart. This is where the two tools genuinely fit together, and it is the reason this portal documents both rather than one.
The current ArtCraft release ships macOS and Windows installers only. There is no Linux build in the published assets, which means a Linux-based workflow today uses PhotoCraft locally and reaches for ArtCraft on the machine that can run it.
The cheapest habit in this workflow is also the least popular: look at the first three frames before you look at the whole clip. Motion failures show up immediately as a silhouette that changes shape or an edge that swims, and catching that at frame three costs one retry instead of the seconds and credits a full render would have spent.
Keep a small reference sheet for recurring shots: the still you started from, the scene layout, the camera setting and the model you chose. When the client asks for a variation two weeks later, that sheet is the difference between a five-minute repeat and a fresh exploration at full price.
Write the shot list before the session rather than after it. One line per shot with the source still, the movement you want and the model you intend to use turns an open-ended exploration into a bounded task, and it makes the cost of an afternoon predictable enough to approve in advance instead of explaining afterwards.
That depends entirely on the provider you connect, because the application itself is free and open source. What you control is the number of paid retries, and cleaning the plate plus directing the camera locally first is the cheapest way to reduce them.
Usually because the reference is ambiguous: overlapping edges, unreadable depth or a subject with no clear silhouette. Fix the still in the composite step and the motion step then has something concrete to follow.
There is no Linux desktop asset in the current release. The official project also exposes a browser entry, but its feature and hardware behavior is a separate upstream surface; the usual Linux workflow pairs PhotoCraft locally with ArtCraft on a machine that can host the desktop release.
Short. A two-second take that is right beats an eight-second take that is almost, because failures in a longer generation accumulate and you end up cutting the result back to its first second anyway.
Choose per shot rather than per brand. Some models hold identity better and some move more plausibly, and since the engine treats the model as a parameter of the step, testing two on the same reference is a workflow rather than indecision.
