Energy for generating one short video with a text-to-video model. Each entry records the claim as published: what it counts, who made it and who paid,
and any dispute. Figures with different boundaries are not comparable without reading what each counts.
Five-second CogVideoX video at 16 frames per second (MIT Technology Review)
944.4 Wh per 5-second video at 16 frames per second (published as: about 3.4 million joules; converted from J to Wh)
But three months later the company launched a larger, higher-quality model that produces five-second videos at 16 frames per second ... The new model uses more than 30 times more energy on each 5-second video: about 3.4 million joules, more than 700 times the energy required to generate a high-quality image.
One 5-second video at 16 frames per second from the newer, larger CogVideoX model
Scope
not stated
Counts
not stated
Leaves out
not stated
Model or version
CogVideoX (the larger model launched three months after the August version; exact version not stated)
Place
not stated
Period measured
not stated
Method
Measured by Sasha Luccioni (Hugging Face) with Code Carbon; the article does not say whether the doubling for non-GPU energy was applied.
Claimed by
Independent: MIT Technology Review (James O'Donnell and Casey Crownhart)
Paid for by
Tarbell Center for AI Journalism (grant)
Dispute
none
History
Added.
Older CogVideoX video at 8 frames per second (MIT Technology Review)
30.3 Wh per video at 8 frames per second (length not stated) (published as: about 109,000 joules; converted from J to Wh)
An older version of the model, released in August, made videos at just eight frames per second at a grainy resolution—more like a GIF than a video. Each one required about 109,000 joules to produce.
One video from the older CogVideoX model, at 8 frames per second and a grainy resolution
Scope
not stated
Counts
not stated
Leaves out
not stated
Model or version
CogVideoX (the older version released in August; exact version not stated)
Place
not stated
Period measured
not stated
Method
Measured by Sasha Luccioni (Hugging Face) with Code Carbon; the article does not say whether the doubling for non-GPU energy was applied.
Claimed by
Independent: MIT Technology Review (James O'Donnell and Casey Crownhart)
Paid for by
Tarbell Center for AI Journalism (grant)
Dispute
none
History
Added.
AnimateDiff video at default settings, Video Killed the Energy Budget
a mean of 0.14 Wh per video (16 frames, 512 × 512, 10 fps, 4 denoising steps)
We observe orders-of-magnitude disparities: AnimateDiff requires only 0.14 Wh in total, while WAN2.1-T2V-14B consumes over 415 Wh, a factor of nearly 3000×.
One video from AnimateDiff at its default Hugging Face settings (16 frames at 10 fps, 512 × 512, 4 steps), averaged over 50 prompts with five runs each
Scope
servers
Counts
GPU energy; CPU energy; RAM energy (estimated)
Leaves out
not stated
Model or version
AnimateDiff
Place
not stated
Period measured
not stated
Method
Run on one dedicated NVIDIA H100 SXM GPU with an 8-core AMD EPYC 7R13 CPU; GPU and CPU energy measured with CodeCarbon, RAM energy estimated with CodeCarbon's default heuristic; two warm-ups and five runs per prompt.
WAN2.1-T2V-14B video at default settings, Video Killed the Energy Budget
at least 415 Wh per video (81 frames, 720 × 1280, 15 fps, 50 denoising steps) (published as: over 415 Wh)
We observe orders-of-magnitude disparities: AnimateDiff requires only 0.14 Wh in total, while WAN2.1-T2V-14B consumes over 415 Wh, a factor of nearly 3000×.
One video from WAN2.1-T2V-14B at its default Hugging Face settings (81 frames at 15 fps, 720 × 1280, 50 steps), averaged over 50 prompts with five runs each
Scope
servers
Counts
GPU energy; CPU energy; RAM energy (estimated)
Leaves out
not stated
Model or version
WAN2.1-T2V-14B
Place
not stated
Period measured
not stated
Method
Run on one dedicated NVIDIA H100 SXM GPU with an 8-core AMD EPYC 7R13 CPU; GPU and CPU energy measured with CodeCarbon, RAM energy estimated with CodeCarbon's default heuristic; two warm-ups and five runs per prompt.