← Ten breakthroughs that brought AI hereA2 · LESSON 5 OF 7

A2, Lesson 5: Pictures From Noise, and Models You Can Download

Tags: Sourced / General / Our view / Illustrative. About 13 minutes. No math or code. Two stories in one lesson.

1. Story one: making images by cleaning up noise (2020 to 2021)

In 2020, researchers at UC Berkeley published "Denoising Diffusion Probabilistic Models." The abstract presents "high quality image synthesis results using diffusion probabilistic models," reporting, on a standard small-image test (CIFAR10), an Inception score of 9.46 and a state-of-the-art FID score of 3.17, and sample quality "similar to ProgressiveGAN" on larger 256 by 256 images. [Sourced: Ho, Jain and Abbeel, 2020, https://arxiv.org/abs/2006.11239 ; the authors' own results on named benchmarks] A year and a half later, a group of researchers at Heidelberg and Munich noted the cost problem: diffusion models "typically operate directly in pixel space," so training "often consumes hundreds of GPU days and inference is expensive due to sequential evaluations." Their fix was to run diffusion in the compact "latent space" of a pretrained autoencoder, and to add cross-attention so that text could steer the image. They report "significantly reducing computational requirements" with competitive quality. [Sourced: Rombach et al., "High-Resolution Image Synthesis with Latent Diffusion Models," https://arxiv.org/abs/2112.10752 ; authors' own claims] Plain version: start with static, then clean it up a little at a time, guided by the words you typed. [General]

2. Story two: Meta releases LLaMA (February 24, 2023)

Meta announced LLaMA on February 24, 2023, as "a state-of-the-art foundational large language model designed to help researchers," in sizes from 7B to 65B parameters. The paper says it trained "using publicly available datasets exclusively," that "LLaMA-13B outperforms GPT-3 (175B) on most benchmarks," and that LLaMA-65B was "competitive with the best models, Chinchilla-70B and PaLM-540B." Meta said it was releasing under "a noncommercial license focused on research use cases," with access granted case by case to researchers. [Sourced: Meta AI blog, Feb 24, 2023, https://ai.meta.com/blog/large-language-model-Llama-meta-ai/ ; LLaMA paper abstract, https://arxiv.org/abs/2302.13971 ; Meta's own claims about benchmark results] The Verge reported on March 8, 2023 that the model "leaked on 4chan a week after it was announced" (a downloadable copy was posted March 3), and that some people worried about misuse while others argued wider access would improve safety. [Sourced, secondary: The Verge, https://www.theverge.com/2023/3/8/23629362/meta-ai-language-model-llama-leak-online-misuse ; read in full]

On August 22, 2022, Stability AI announced the public release of Stable Diffusion, following an earlier launch announcement for researchers on August 10. The announcement says the model is released under a Creative ML OpenRAIL-M license that "allows for commercial and non-commercial usage" and focuses on "the ethical and legal use of the model," and that an AI-based safety classifier is included by default. [Sourced: Stability AI, "Stable Diffusion Public Release," https://stability.ai/news-updates/stable-diffusion-public-release ; company claims]

3. The one idea

Two shifts happened at about the same time. Generating images from text became practical on modest hardware, and the "weights" of capable language models became available to people outside the big labs. This widened who can build with AI, and who can misuse it. [Our view]

4. The kitchen-table version

Image generation: a sculptor starts with a rough block and chips away a little at a time, following your description. Open weights: instead of renting a chef, you are handed the recipe and can cook it in your own kitchen, but you need the kitchen, and the recipe may have conditions attached. [General]

5. What "open" can mean

  • Open weights: the trained numbers can be downloaded and run on your own machine.
  • Open source (in the strict sense): code, data and license all permit free reuse. Many "open" models do not meet this.
  • License terms: LLaMA's first release was for research use only. Later releases have had different terms; read the license for the version you want to use. [Sourced for LLaMA 1: Meta blog; later versions not checked here] For a business, the question is not "is it open?" but "what may I do with it, who is responsible if it goes wrong, and can I support it?" [Our view]

6. Worked example: steps and size

All numbers [Illustrative], computed by script. Image steps. Suppose each cleanup step takes 0.05 seconds. A 1,000-step process takes 50 seconds per image; 250 steps, 12.5 seconds; 50 steps, 2.5 seconds; 20 steps, 1 second. Newer methods try to use fewer steps, but there is a quality tradeoff. The paper above names "sequential evaluations" as the cost. Compressing first. A 512 by 512 color image has 512 x 512 x 3 = 786,432 numbers. Working on a compressed 64 x 64 x 4 version has 16,384: 48 times fewer. That is the idea of latent diffusion: do the heavy work in the small space. (The sizes here are an example, not the paper's exact settings.) Run it yourself or rent it? Suppose a rented graphics machine costs $1.50 an hour, $1,095 a month. If a hosted service charges $2 per million tokens, the machine pays for itself only above about 547.5 million tokens a month. Below that, renting is cheaper, before counting your time to set up and maintain it. [Our view]

7. What it costs and what it breaks

  • Rights: image models trained on scraped pictures face lawsuits and rules (see C3 Lesson 6).
  • Fakes: easy image creation makes convincing fakes easy too; check images before trusting them.
  • Leaks and licenses: "research only" did not stop the LLaMA leak the press reported; a license is a promise, not a lock.
  • Support: if you run a model yourself, updates, safety fixes and bugs are yours.
  • Quality varies: benchmark wins by the maker are claims.

8. Monday checklist

  1. If you use generated images, record where they came from and what the tool's terms say.
  2. Before choosing an open model, read its license for your use.
  3. Ask whether a hosted service would be cheaper than running your own at your volume.
  4. Check any "outperforms" claim against your own test.
  5. Add a note to your image-use policy about disclosing AI images.

9. Try it yourself (no code, 20 minutes)

Take a blurry photo on your phone and think of how you would restore it step by step. That is the diffusion idea. Then write down one job in your business where you would prefer to run a model yourself (privacy?) and one where you would rent.

10. Self-check

  1. What is the core idea of diffusion image models? Start from noise and clean it up in steps, guided by text.
  2. What did latent diffusion change? It did the work in a compressed space, cutting compute.
  3. Under what license did Meta first release LLaMA? A noncommercial research license, access case by case.
  4. What did The Verge report a week later? That the model leaked online.
  5. When does running your own model beat renting in the example? Above about 547.5 million tokens a month.

11. Where this is going

Both trends continue: more capable models you can download, and image, audio and video generation. Rules on rights and fakes are still being settled. [Our view]

Sources

Review flags

  • RESOLVED (Oct 7, 2026): affiliations match the author lines; DDPM first posted Jun 19, 2020 and latent diffusion Dec 20, 2021 (arXiv submission pages); LLaMA paper Feb 27, 2023, Meta blog Feb 24, 2023.
  • RESOLVED: The Verge Mar 8, 2023 read in full. It says the leak was posted March 3; Meta's Joelle Pineau confirmed that "some have tried to circumvent the approval process." The lesson does not claim more.
  • RESOLVED: Stable Diffusion researcher launch Aug 10, 2022 and public release Aug 22, 2022 are both on Stability AI's own pages (read). License and safety-classifier statements are company claims.
  • All benchmark claims are the authors' or Meta's own (abstracts and the Meta blog read, not the methods). The lesson labels them.
  • LLaMA's later licensing is not covered; the lesson tells the reader to read the license for the version used.
  • Step times, sizes, rental price and token price are invented.

Claim ledger (A2 Lesson 5)

#ClaimSourceVerifiedStatusRe-check
1DDPM: CIFAR10 Inception 9.46, FID 3.17; LSUN 256 similar to ProgressiveGANarXiv 2006.112392026-10-06Read abstractnone
2Latent diffusion: pixel-space cost "hundreds of GPU days"; latent space and cross-attention; reduced computearXiv 2112.107522026-10-06Read abstractnone
3LLaMA Feb 24, 2023; 7B-65B; public data only; 13B beats GPT-3 on most benchmarks; noncommercial licenseMeta blog; arXiv 2302.139712026-10-06Read; company claimsnone
4LLaMA leaked on 4chan a week laterThe Verge (Mar 8, 2023)2026-10-06Secondary, summary onlyRead in full
5Step, size and rent-vs-run arithmeticOur computation2026-10-06Illustrative, script-checkednone
6DDPM first posted Jun 19, 2020; latent diffusion Dec 20, 2021; LLaMA paper Feb 27, 2023arXiv pages2026-10-07Readnone
7Stable Diffusion researcher launch Aug 10, 2022; public release Aug 22, 2022Stability AI pages2026-10-07Read; company claimsnone