The economics of generated video: why iteration is a budget line

Picture a read-through where every performer has to hear every other performer in the room. Two dozen people is manageable. Add another dozen and the number of pairs who must hear each other grows much faster than the headcount. Self-attention works like that room: every token attends to every other token, so the cost of the pass rises faster than linearly as the sequence gets longer. An image fills the room with tokens across width and height. A video adds time, and each new second brings a fresh crowd that everyone already present must also hear.

The real mechanism is spacetime tokenisation. The Sora report describes a transformer operating on spacetime patches of video and image latent codes. The scale is public: Movie Gen's largest video model is a 30-billion-parameter transformer trained with a maximum context length of 73,000 video tokens, corresponding to 16 seconds at 16 frames per second. Compression is the main defence. Video models encode into a latent space first, and LTX-Video reports a 1:192 compression ratio at 32 by 32 by 8 pixels per token, while noting that compression this aggressive inherently limits fine detail. That is the trade: tokens are the currency, and quality scales with them. The Diffusion Transformer work found that more input tokens means more compute and better results.

On top of the token count sit multipliers. Iterative sampling runs many denoising steps, and every step touches the whole clip, not one frame. Resolution, duration and frame rate each multiply the token count before any of that begins. Then the human multiplier: someone watches every generation and rejects most of them, and that selection time does not compress.

The production consequence is blunt. You do not pay per second of screen time. You pay per attempt.

What people get wrong

Each of these is a belief worth dropping, followed by what is actually happening.

• video costs about as much as an image times the frame count. Tokens span space and time and attention cost grows faster than linearly with sequence length, so doubling duration costs more than double.
• the model generates frame by frame, so a longer clip is just more of the same work. Sampling is iterative over the whole clip at once, and every denoising step processes the full spacetime latent.
• compute is the dominant cost of a generative sequence. Human selection is usually the larger line, because someone must watch and reject the majority of generations and that time does not scale down.
• generating at final resolution from the start saves a step. The standard economy is to prototype small and finish large, which the Sora report describes as an explicit benefit of training on native sizes.
• more compression is free efficiency. LTX-Video is explicit that its high compression ratio inherently limits the representation of fine detail, which is why the model puts effort back into the decoder.

Terms this introduces

Spacetime patch

The unit a video transformer processes, a small block of latent covering an area of frame across a span of frames.

Context length

The maximum number of tokens a model can attend over at once, which sets the longest clip it can generate in one pass.

Compression ratio

How much smaller the latent representation is than the pixels, traded directly against retained fine detail.

Denoising steps

The number of iterative passes over the whole latent clip, each one a full-cost forward pass.

Accept rate

The proportion of generations a human keeps, which is the number that actually determines cost per finished shot.