Character consistency in AI video feels mostly solved.
9.4K followers
19,779 views
52 Comments
Comments from the original post on X (Twitter)

@madpencil_
Here are the two refs I used for this test. 1. The set: one over head top down 90° nadir shot of the pond with a small compass in the corner. This was the "floor plan". 2. The character: a simple turnaround sheet (front, back, side, 3/4 and a face close-up). That's it. No 3D scene, no extra angles of the location. Just these two images and directions in the prompt. Feel free to grab them and try it yourself.
37 likes

@madpencil_
That's the experiment so far. Next, I want to test the same idea indoors: bedrooms, kitchens, small rooms where furniture loves to teleport between cuts. A top-down floor plan might help even more there. I'll keep sharing what works and what breaks along the way. If you try it, tag me. I'd really love to see your results.
14 likes

@martianprompts
@magnific I tried something similar for my short film too. Adding text overlays over the image helped me a lot. Also, you N S W E direction display is clever idea!
9 likes

@madpencil_
@magnific Thanks bro, annotation has worked for a while now from early veo times.
3 likes

@ImgMotion
Ok, I have a question. I have been creating automated scenes with a group of characters in a fictious town. I have many interior environments and I am finding it hard to get consistency when filming. I have tried Gaussian splats and H3 / QWEN 2.1 multi angle captures, but still the room changes. Also what Model and Workflow are you using with your reference images?
3 likes

@madpencil_
@magnific I used just one image of the location and one character sheet rest is just normal prompt, that's the workflow. I think keeping it simple doesn't confuses the model.
1 like

@AiImagePro
@magnific I built something similar. a virtual set generator . Take photos of important aspects of your location. Also take a top down or find a plan. Import the plan. Set its north and import the images set the virtual cameras on the plan and you have a lovely virtual set.
2 likes

@azed_ai
@magnific This is honestly so beautiful
2 likes

@madpencil_
@magnific Thanks Amira.

@kristofcreative
I literally did almost the exact thing for training an image gen model. Instead of adding a compass (great idea, BTW), I described the same prominent window in each photo as "West facing". Keeping with the "map" idea, perhaps try adding a legend with a few height and distance markers so it will better understand scale. That was going to be my next step but haven't tried it yet. Excellent work!
2 likes

@madpencil_
@magnific Thank you, I'll try out your inputs too we need hive mind to solve this.
1 like

@JakeMercer1997
@magnific Looks great! I’ve been trying the same sort of thing with Claude setting up the scene in Blender, then taking different camera angle blocked shots. Your way might be easier!
2 likes

@imbrowniegarcia
@magnific Brilliant
2 likes

@Moriarty_tech
@magnific Compass in the corner is a nice trick. Does 'looking south' really hold, or does the model mostly lean on the top-down ref and ignore the letters?
2 likes

@madpencil_
@magnific You have to guide the directions with prompting too just letting it know where is your camera facing angles etc, check my other post after this one, I tested with an Indoor setup.
1 like

@Cappinski
@magnific Exteriors are a tough nut to crack. I typically go for a gods eye wide angled down and throw labels if anything's unclear. For the most part, it goes well! Interiors I just use isometric layouts and if needed be describe the reverse if the camera is swapping.
2 likes

@ChristianimaNet
@magnific I create the environment shot, then create a 360 video of it. I extract the stills of the angle I want and provide that reference plus the original reference. Another trick I use is to create the shot then generate 20 different camera angles. I get some amazing shots this way.
2 likes

@gb9886
@magnific 很好
1 like

@Cole_A_Higgins
@magnific Hell yeah
1 like

@JinLam7
@magnific thanks
1 like




















