Continuous Function

Generative Models

Diffusion Image Generation: Noise, Guidance, and Latents

Follow a text-to-image diffusion pass through noisy latents, noise prediction, classifier-free guidance, sampler updates, and final decoding without mistaking the pretty preview for the mechanism.

status: reviewimportance: importantdifficulty 4/5math: graduateread: 24mlive demo
Back to Generative ModelsNext: Flow Matching & Rectified Flows

Concept Structure

Diffusion Image Generation: Noise, Guidance, and Latents

01Intuition

Start with the picture, metaphor, or geometric mechanism.

02Math

Make the objects explicit and connect them with notation.

03Code

Mirror the equations with runnable implementation details.

04Interactive Demo

Manipulate the mechanism and watch the idea respond.

2prerequisites
2next concepts
4related links

Learner Contract

What this page should let you do.

You are here becauseFollow a text-to-image diffusion pass through noisy latents, noise prediction, classifier-free guidance, sampler updates, and final decoding without mistaking the pretty preview for the mechanism.

This Generative Models concept is the current idea: keep the same invariant visible across Intuition, Math, Code, Interactive Demo.

By the end4/4 sections ready | runnable code expected | live demo

Explain the mechanism, trace the main notation, and test one prediction in the live demo.

Do this firstIntuition

Read the intuition before the notation; the math should name a mechanism you already felt.

Then go nextFlow Matching & Rectified Flows

Follow this edge after making one prediction here; the next page should reuse the result, not restart the route.

Test the linkManipulate one control and predict the visible change.Then continue to Flow Matching & Rectified Flows

Claim/source review status

Substantive review recorded

2/2 claims have bounded review metadata; still check caveats and source scope.Metadata-derived; review may be AI-assisted. Not a human certification.
Claims2/2 reviewed
Sources4 cited
Codeattached
Demolive
Reviewed2026-07-04
Updatedpage 2026-07-04

Learning item flow

4/4 sections readyAsk about thisResearch room
ConceptDiffusion Image Generation: Noise, Guidance, and LatentsGenerative ModelsEquationDiffusion Image Generation: Noise, Guidance, and Latents equation 1Exact equationCodeDiffusion Image Generation: Noise, Guidance, and Latents runnable code 1Exact runnable codeDemoDiffusion Image Generation: Noise, Guidance, and Latents interactive...Interactive demoClaimA practical diffusion image-generation pipeline repeatedly predicts n...Exact claim checkSourceDenoising Diffusion Probabilistic ModelsExact source
ConceptDiffusion Image Generation: Noise, Guidance, and LatentsGenerative Models
4 sources attachedLocal snapshot ready
concept:generative-models/diffusion-image-generation
CodenearbyPredictbefore revealRoomitem handoff
01IntuitionStart with the picture, metaphor, or geometric mechanism.02MathMake the objects explicit and connect them with notation.03CodeMirror the equations with runnable implementation details.04Interactive DemoManipulate the mechanism and watch the idea respond.
01

01

Intuition

Build the mental picture first so the rest of the page has something to attach to.

Section prompt

The easiest mistake with image diffusion is to stare only at the final picture.

The mechanism is a sequence:

  1. start from noise,
  2. ask the model what noise direction it sees at the current timestep,
  3. optionally mix an unconditional estimate with a prompt-conditioned estimate,
  4. let the sampler turn that estimate into the next, slightly cleaner state,
  5. repeat,
  6. decode the final latent into pixels.

That means the useful question is not "does the preview look nice?" It is:

which intermediate object moved the sample toward the prompt, and which step could distort it?

Classifier-free guidance is one of those places. Low guidance can ignore the prompt. Higher guidance can push harder toward the prompt, but very high guidance can also collapse detail or create artifacts. Latent diffusion adds another boundary: the repeated denoising may happen in a compressed latent space, while the visible image appears only after a decoder maps the latent back into image space.

The lab makes those hidden objects inspectable. You predict the next move, then reveal whether the answer used the noise estimate, guidance mix, sampler step, and latent decode in the right order.

02

02

Math

Translate the story into symbols, assumptions, and a derivation you can inspect.

Section prompt
Equation 1ϵθ(xt,t,c).\epsilon_\theta(x_t, t, c).Equation 2ϵu=ϵθ(xt,t,∅),ϵc=ϵθ(xt,t,c).\epsilon_u = \epsilon_\theta(x_t, t, \emptyset), \qquad \epsilon_c = \epsilon_\theta(x_t, t, c).

At timestep tt, a denoising model often predicts a noise-like quantity:

ϵθ(xt,t,c).\epsilon_\theta(x_t, t, c).

For classifier-free guidance, keep both an unconditional estimate and a conditional estimate:

ϵu=ϵθ(xt,t,∅),ϵc=ϵθ(xt,t,c).\epsilon_u = \epsilon_\theta(x_t, t, \emptyset), \qquad \epsilon_c = \epsilon_\theta(x_t, t, c).

The guided estimate is:

ϵmix=ϵu+w(ϵc−ϵu),\epsilon_{\text{mix}} = \epsilon_u + w(\epsilon_c-\epsilon_u),

where ww is the guidance scale. Larger ww increases prompt pressure, but it is not free quality.

A DDPM-style reconstruction of the clean sample can be written:

x^0=xt−1−αˉtϵmixαˉt.\hat{x}_0 = \frac{x_t-\sqrt{1-\bar\alpha_t}\epsilon_{\text{mix}}} {\sqrt{\bar\alpha_t}}.

A sampler then combines x^0\hat{x}_0 and the current estimate to produce a previous timestep:

xt−1=step⁡(xt,ϵmix,t).x_{t-1} = \operatorname{step}(x_t, \epsilon_{\text{mix}}, t).

In latent diffusion, xtx_t is a latent rather than pixels. The final visible image is:

image⁡=D(z0),\operatorname{image} = D(z_0),

where DD is the decoder. If the latent still has too much noise, too few steps, or over-strong guidance, a nice-looking preview can still be mechanically suspect.

03

03

Code

Keep the implementation aligned with the notation so the algorithm is legible.

Section prompt
Runnable code 1import numpy as np x_t = np.array([0.84, -0.35, 0.18, 0.42]) eps_uncond = np.array([0.18, -0....python
import numpy as np

x_t = np.array([0.84, -0.35, 0.18, 0.42])
eps_uncond = np.array([0.18, -0.12, 0.09, 0.16])
eps_cond = np.array([0.23, -0.21, 0.03, 0.20])

w = 7.5
alpha_bar_t = 0.42
alpha_bar_prev = 0.48

eps_mix = eps_uncond + w * (eps_cond - eps_uncond)
x0_hat = (x_t - np.sqrt(1 - alpha_bar_t) * eps_mix) / np.sqrt(alpha_bar_t)

# A compact deterministic DDIM-like teaching step.
direction = np.sqrt(1 - alpha_bar_prev) * eps_mix
x_prev = np.sqrt(alpha_bar_prev) * x0_hat + direction

alignment_push = np.linalg.norm(eps_mix - eps_uncond)
step_size = np.linalg.norm(x_prev - x_t)

print("eps_mix", np.round(eps_mix, 3))
print("step_size", round(float(step_size), 3))
print("alignment_push", round(float(alignment_push), 3))

This is not a full text-to-image model. It exposes the contract: guidance changes the noise estimate, the sampler changes the latent, and decoding comes after the denoise path.

04

04

Interactive Demo

Use direct manipulation to connect the explanation to a moving system.

Section prompt

Use the Diffusion Image Generation Pipeline Lab as a prediction-first check:

  • Order route: choose the correct mechanism order from prompt to decoded image.
  • Guidance mix: decide what changes when the guidance scale increases.
  • Sampler step: decide what the sampler updates before the final decode.
  • Decode boundary: decide why the final preview is not enough evidence by itself.

Before reveal, the mixed estimate, step size, alignment push, artifact risk, and decode readiness stay locked. After reveal, inspect the verifier rather than trusting the preview alone.

Live Concept Demo

Explore Diffusion Image Generation: Noise, Guidance, and Latents

The stage is code-native and interactive. Use it to test the explanation against the mechanism.

difficulty 4/5graduatecode-aligned
Demo Prediction Checkpoint

Manipulate one control and predict the visible change.

Commit to what Diffusion Image Generation: Noise, Guidance, and Latents should make visible before reading the result.

After The First Pass

Turn the concept into an inspected object.

Once the invariant is visible in the intuition, math, code, and demo, use these panels to inspect the mechanism visually, check source support, practice the idea, and attach a grounded research question.

Mechanism Storyboard

See the idea move before the page explains it

Follow a text-to-image diffusion pass through noisy latents, noise prediction, classifier-free guidance, sampler updates, and final decoding without mistaking the pretty preview for the mechanism.

Prediction open01 / Intuition
Prediction lens

Start with the picture, metaphor, or geometric mechanism.

Commit first

Before reading further, choose the kind of change Diffusion Image Generation: Noise, Guidance, and Latents should make visible.

Visual Inquiry

Make the image answer a mathematical question

Follow a text-to-image diffusion pass through noisy latents, noise prediction, classifier-free guidance, sampler updates, and final decoding without mistaking the pretty preview for the mechanism.

4/4 stages readyLive demo connected
Prediction

Which visible object should carry the first intuition?

Commit first

Pick the cue that should make Diffusion Image Generation: Noise, Guidance, and Latents easier to reason about before the page gives the answer.

Source Grounding

Canonical references for the mechanism on this page.

paper · 2020Denoising Diffusion Probabilistic ModelsHo, Jain, and Abbeel

Primary support for DDPM image synthesis through learned denoising/noise prediction and iterative reverse sampling.

Open source
paper · 2020Denoising Diffusion Implicit ModelsSong, Meng, and Ermon

Primary support for DDIM-style faster iterative sampling and step-count tradeoffs using the same training objective family.

Open source
paper · 2022Classifier-Free Diffusion GuidanceHo and Salimans

Primary support for combining conditional and unconditional estimates to steer conditional generation without a separate classifier.

Open source
paper · 2021High-Resolution Image Synthesis with Latent Diffusion ModelsRombach et al.

Primary support for running diffusion in a learned latent space, then decoding the latent back to image space.

Open source

Claim Review

Follow a text-to-image diffusion pass through noisy latents, noise prediction, classifier-free guidance, sampler updates, and final decoding without mistaking the pretty preview for the mechanism.

Status2 substantive reviews recorded

Claims without a substantive review badge still need exact source-support review.

Sources4 references

ho-2020-ddpm, song-2020-ddim, ho-2022-classifier-free-guidance, rombach-2021-latent-diffusion

Local checks4 local checks

Use equations, runnable code, and demos to check whether the source support is operational.

Substantively reviewedA practical diffusion image-generation pipeline repeatedly predicts noise or score-like direction at the current timestep, uses a sampler update to move toward a cleaner sample, and only then treats the decoded output as an image.Claim metadata: source checked

The references support learned denoising/noise prediction, iterative reverse sampling, and faster DDIM-style sampling tradeoffs.

Sources: Denoising Diffusion Probabilistic Models, Denoising Diffusion Implicit ModelsThe demo is a deterministic educational witness, not a trained image generator and not a claim about final image quality.A bounded review summary is present; still check caveats and exact reference scope.

Checked DDPM and DDIM for the bounded image-generation contract: learned denoising estimates drive iterative reverse updates, while sampler choice and number of steps change the reverse trajectory. GPT Pro publication critique remains pending because the local Oracle lane is unavailable on this workstation.

Reviewer: codex-local-primary-reference-audit; reviewed 2026-07-04
Equation 1Equation 2Runnable code 1Interactive demo
Substantively reviewedText-conditioned image generation can be taught as a guidance mix between conditional and unconditional estimates followed by sampler steps in latent space and a final decode to image space.Claim metadata: source checked

The references support classifier-free mixing of conditional/unconditional estimates and latent diffusion's compressed latent-space sampling plus decode route.

Sources: Classifier-Free Diffusion Guidance, High-Resolution Image Synthesis with Latent Diffusion ModelsA decoded preview is not proof that the mechanism was followed; the verifier checks ordering and intermediate quantities before trusting the output.A bounded review summary is present; still check caveats and exact reference scope.

Checked classifier-free guidance for the conditional/unconditional mixing contract and latent diffusion for the latent-space diffusion plus decode boundary. The local demo exposes guidance scale, over-guidance risk, latent-vs-pixel mode, and decode readiness.

Reviewer: codex-local-primary-reference-audit; reviewed 2026-07-04
Equation 2Runnable code 1Interactive demo

Source support candidates

paper 2020Denoising Diffusion Probabilistic Models

Primary support for DDPM image synthesis through learned denoising/noise prediction and iterative reverse sampling.

paper 2020Denoising Diffusion Implicit Models

Primary support for DDIM-style faster iterative sampling and step-count tradeoffs using the same training objective family.

paper 2022Classifier-Free Diffusion Guidance

Primary support for combining conditional and unconditional estimates to steer conditional generation without a separate classifier.

paper 2021High-Resolution Image Synthesis with Latent Diffusion Models

Primary support for running diffusion in a learned latent space, then decoding the latent back to image space.

Mechanism checks

Equation 1
ϵθ(xt,t,c).\epsilon_\theta(x_t, t, c).
Equation 2
ϵu=ϵθ(xt,t,∅),ϵc=ϵθ(xt,t,c).\epsilon_u = \epsilon_\theta(x_t, t, \emptyset), \qquad \epsilon_c = \epsilon_\theta(x_t, t, c).
Runnable code 1import numpy as np x_t = np.array([0.84, -0.35, 0.18, 0.42]) eps_uncond = np.array([0.18, -0....Demo stateLive mechanism probe

Practice Loop

Try the idea before it explains itself

Follow a text-to-image diffusion pass through noisy latents, noise prediction, classifier-free guidance, sampler updates, and final decoding without mistaking the pretty preview for the mechanism.

Readiness0/3 checks ready
Predict

Before touching the demo, predict one visible change that should happen in Diffusion Image Generation: Noise, Guidance, and Latents.

Hint 1

Reveal when your model needs a nudge.

Hint 2

Reveal when your model needs a nudge.

Hint 3

Reveal when your model needs a nudge.

Grounded research drawerClose
ConceptDiffusion Image Generation: Noise, Guidance, and LatentsGenerative Models
Runnable code comparisonDiffusion Image Generation: Noise, Guidance, and Latents runnable code 1x_t = np.array([0.84, -0.35, 0.18, 0.42])Prediction before revealDiffusion Image Generation: Noise, Guidance, and Latents interactive demoManipulate one control and predict the visible change.
Grounded room questionWhat is the smallest example that makes Diffusion Image Generation: Noise, Guidance, and Latents click without losing the math?Local snapshot ready

Research Room

Attach the question to a claim, equation, code, or demo

Pick the concept, equation, source, runnable code, claim, misconception, or demo state before asking for help. The handoff keeps that page item in context.
Next local actionNo local draft saved yet

Open the draft below to save one note and next action in this browser.

conceptGenerative Models

Diffusion Image Generation: Noise, Guidance, and Latents

Attached question

What is the smallest example that makes Diffusion Image Generation: Noise, Guidance, and Latents click without losing the math?

Local action draftNo local draft saved yetExpand only when ready to capture one local next action
Local action draft

This draft stays in this browser, attached to the selected learning item.

No local draft saved.
Evidence to inspect
  • References to inspect: attached references on this page.
  • Definition, prerequisite, and contrast concept links
  • The equation or runnable code that makes the concept operational
  • One demo state that shows the invariant instead of a slogan
What would resolve this
  • The learner can state the mechanism in their own words
  • The learner can name the prerequisite that would repair confusion
  • The learner can predict how the mechanism changes under one perturbation
Grounded AI handoff

I am working in Continuous Function's research reading room. Object: concept - Diffusion Image Generation: Noise, Guidance, and Latents Selected item key: recorded for copy. Context: Generative Models Page anchor: recorded for copy. Open question: What is the smallest example that makes Diffusion Image Generation: Noise, Guidance, and Latents click without losing the math? Evidence to inspect: - References to inspect: attached references on this page. - Definition, prerequisite, and contrast concept links - The equation or runnable code that makes the concept operational - One demo state that shows the invariant instead of a slogan What would resolve this: - The learner can state the mechanism in their own words - The learner can name the prerequisite that would repair confusion - The learner can predict how the mechanism changes under one perturbation Answer as a careful research tutor: stay source-grounded, separate verified evidence from assumptions, name the relevant math objects, and end with one next action.

View it in context
concept/concept-notebook/generative-models/diffusion-image-generation concept:generative-models/diffusion-image-generation