section 07 · status: live · 6 entries · updated 2026-08-29
Generative Media
Diffusion turned noise into photographs and prompts into film. How the trick works, how to direct it in practice, how to judge the output, where video and 3D stand, and the provenance problem the whole field now owes the world an answer to.
live · pure noise denoised into a sample, step by step
- Diffusion models3 minstart from static, subtract the noise
- Image generation in practice3 mindirecting a very literal artist
- Video, voice & music3 mincoherent for seconds, drifting by minutes
- 3D & world models2 minfrom generating pictures to generating places
- Judging generated media3 minthere is no accuracy score for a picture
- Provenance & detection3 minassume detection fails, prove the real
check yourselfAnswer before you open
Trying to recall something teaches it better than re-reading does. Have a go, then open the answer.
Your generated image is disappointing. Why is raising guidance usually the wrong first move?
Guidance controls literalness, not quality. Raising it makes the same idea louder — blown saturation, crushed contrast, less variety. Change the seed, the step count or the wording first; move guidance last and in small steps. Diffusion models →
A detector says an image is '98% AI-generated'. What is wrong with acting on that?
Detectors are trained on the artifacts of previous generators and drift toward a coin flip on new ones — and at any plausible false-positive rate over mostly-authentic traffic, most flags are innocent. Provenance and process, not a confidence score, should carry a consequential decision. Provenance & detection →
When would you use a captured 3D scan rather than a generated 3D asset?
When the answer must be measurable. Reconstruction tells you what is there; generation tells you what could be there. A dreamed hallway has no reliable dimensions — and may not be the same width twice. 3D & world models →
Your generated hero image looks great but the product has five buttons instead of four. What is the efficient fix?
Not a longer prompt. Counting, spatial relations and attribute binding are the stable weak spots, and restating them rarely helps. Generate wide and select, then fix the region directly — inpaint the failed area, keep the seed, change one clause at a time so each change is attributable. Image generation in practice →
A model tops the public preference arena. Why might it still be the wrong choice for your product?
Arenas measure a prompt mix that is not yours, raters reward eye-catching over correct, and a leaderboard that becomes a target gets a house style that wins comparisons and disappoints in use. Score your own prompt set against a checkable rubric, and count cost per accepted image and attempts per acceptance. Judging generated media →