Noise / Denoising
What is Noise / Denoising?
In AI image generation, noise is a starting image of pure random static, and denoising is the process of the AI gradually turning that static into a coherent picture. In post-production, denoising refers to removing unwanted grain or sensor noise from footage to make it cleaner.
At a glance
- Also known as
- Diffusion processScore matchingNoise scheduling (for the training process)
- Used for
- AI image and video generationNoise removal in post-productionUpscaling and restorationControlling generation randomness via seeds
- Common tools
- Stable diffusionMidjourneySoraTopaz video AINeat videoDaVinci resolve
- Related terms
- Diffusion modelLatent spaceCFG scaleStepsSeedSampling method
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
How it compares
In AI generation, noise is intentional and generative: it is the structured starting point that the model shapes into an image. In post-production, noise is an unwanted artefact introduced by camera sensors, film grain, or compression, which denoising tools attempt to remove. Both fields use similar mathematical models of noise, but with opposite intent.
Think of it like…
Imagine a sculptor who starts with a block of marble (pure noise) and gradually chips away to reveal the figure hidden within. Each pass of the chisel is a denoising step, and the sculptor's vision (the text prompt) guides which material to remove and which to keep. The final sculpture emerges not by building up from nothing, but by progressively removing what doesn't belong.
Pro tip
In diffusion-based AI generation, reducing the number of denoising steps speeds up output significantly, but at the cost of coherence: experiment with a mid-range step count (20–30 for many samplers) as a starting point, and only increase it if you're seeing structural incoherence. For post-production denoising, always apply noise reduction before any sharpening or detail enhancement to avoid amplifying noise patterns.
Types and variations
- In diffusion-based generation, noise schedules vary: different models use different rates of noise addition and removal during training and inference, affecting output characteristics.
- Gaussian noise (normally distributed random values) is standard, but some research explores structured noise forms.
- In post-production denoising, approaches include temporal denoising (comparing noise patterns across frames over time), spatial denoising (analysing noise within a single frame), and AI-based denoising (using trained networks to distinguish noise from genuine detail).
- Each approach suits different types of noise and footage, and combining them yields the best results on heavily noised material.
Ready to make your first scene in Morphic?
Try MorphicCommon use cases
- Noise and denoising are central to two major creative AI use cases.
- First, in AI generation: every time a prompt is submitted to Stable Diffusion, Midjourney, Sora, or a similar tool, the model begins with a noise tensor and iteratively denoises it to produce the output: adjusting the number of denoising steps and the guidance scale are primary creative controls.
- Second, in post-production: AI denoising is used to rescue underexposed or high-ISO footage, clean up archival film scans, reduce compression artefacts in streaming deliverables, and make visible detail that noise would otherwise obscure.
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
FAQs
A sampling method (also called a sampler or scheduler) is the algorithm used to navigate the denoising process from noise to image. Different samplers ( such as DDIM, Euler, DPM++, and PLMS ) make different trade-offs between speed and quality. Some produce similar results in fewer steps; others are better suited to certain types of subject matter.
A seed is a number used to initialise the random noise at the start of the generation process. Using the same seed with the same prompt and settings reproduces the same output, making seeds essential for iterating on a composition or style without losing a result you want to build on.
Beyond a certain threshold ( typically 30–50 steps for most models ) additional steps yield diminishing returns and can even introduce over-smoothing or slight shifts in composition. Some modern samplers are specifically designed to reach high-quality results in 8–12 steps, making step count less critical than sampler choice.
Traditional noise reduction uses mathematical filters to blur or average pixel values, often destroying fine detail in the process. AI denoising uses a trained neural network that has learned to distinguish genuine texture and detail from noise patterns, enabling it to remove grain while preserving sharpness far more effectively.
Yes: both in generation and post-production. In generation, inpainting allows denoising to be applied only within a masked region. In post-production tools like DaVinci Resolve and Topaz Video AI, spatial masks can limit denoising to specific areas such as shadowed backgrounds, preserving intentional grain or texture in other regions.
CFG (Classifier-Free Guidance) scale controls how strongly the text prompt steers the denoising process. A higher CFG value makes the model follow the prompt more rigidly, producing results more closely aligned with the description but sometimes at the cost of naturalness. A lower value gives the model more freedom, which can produce more aesthetically pleasing but less prompt-accurate results.
