This iconic action scene turned out to be incredibly difficult to recreate — even when you have the original sequence right in front of you as a reference. At first, it seems simple: the shots already exist, the camera movement is there, the timing is known. You just need to repeat it. In reality, it’s the opposite. A very interesting experiment — and definitely one of the more challenging scenes I’ve tried to reimagine with AI.

I created the entire action sequence first as a 3D previs using a Clay Render — focusing purely on choreography, camera movement, timing, composition, and scene continuity. No textures. No polished lighting. No distractions. Just the core mechanics of the scene. For action sequences, I find this approach especially useful: if the scene already feels cinematic in a simple clay render, the final visuals only make it stronger. Previs → Clay Render → Final Visual Sedance 2.0.

I created this video using Seedance 2.0, The goal was to test how far I could push AI video while keeping a strong cinematic language — controlled camera movement, composition, atmosphere, lighting, and consistency between shots. Most of the work was about experimenting with prompts, reference frames, camera behavior, and multiple generations until the shots started to feel like parts of the same film. Created with Seedance 2.0. Would be interesting to hear what you think about the result.

Sedance 2.5 1080p I tried to create one proper cinematic shot using AI — and in the end, nothing really worked out 😅 I went through tons of generations, different settings, references, and approaches. I even used a Clay Render to give the AI a more accurate understanding of the scene geometry, composition, and camera movement. In the end, I’ve already spent around $200 experimenting, and I still haven’t managed to get the shot I actually wanted. I used the prompt from the attached TXT file together with 4 key visual frames as references for the generation. I ran a lot of tests and iterations, trying different approaches, settings, and workflows. After all that experimentation, I still consider the final result very disappointing and far below my expectations, especially considering the amount of time and money spent.