Cycles Texture Cache: What a Memory Saving Actually Changes

Two fictional 8 GiB scenes show why the same texture reduction changes total memory differently. Separate cache preparation, render cost and the image check.

Violet translucent scene frame beside separate amber texture tiles and a rounded charcoal form.

Two scenes can each use 8 GiB of memory and respond very differently to a texture cache. In one, most of that memory belongs to images. In the other, geometry and other data dominate. Even an impressive reduction in texture memory can therefore leave a production problem largely unchanged.

The useful question is not simply whether caching saves memory. It is which allocation becomes smaller, what work produces that saving, and whether the resulting image still meets the shot’s requirements. Those questions give a small team a better starting point than carrying a percentage from someone else’s scene into its own schedule.

The mechanism: request the detail the render needs

Blender’s 1 May 2026 development account describes a Cycles system that loads selected image tiles and resolutions from texture-cache files. Rather than keep every source image at full resolution in memory, the renderer can request the portions needed by its calculations. The article explicitly connects the size of the saving to the scene’s balance of texture and geometry memory.

Picture a fictional interior shot. A wall has a detailed image texture, but in the current view it occupies a small part of the frame. A camera move brings a labelled object into close-up. The useful level of detail is now different for that object. This is a conceptual example, not a render we measured: it explains why a scene’s asset list alone cannot predict which texture data a shot will need.

The implementation described in the developer account gathers missing texture requests, loads the required tiles in batches and runs the affected kernels again. That is real work. Loading less data does not mean the renderer has no additional work to do, nor does it establish that every view will render faster.

What the documented controls actually change

The Blender 5.2 LTS manual, consulted on 13 November 2026, describes tiled, mipmapped .tx files and a trade-off involving memory, rendering performance and disk space. Its Auto Generate control creates or updates cache files when rendering; Generate prepares them for images used by materials, while Clear removes generated cache files. The documented default location is a blender_tx/ folder beside the image, with a file-path preference for changing it.

That distinction matters when a collaborator asks what to deliver. The source image, the generated cache and the memory used during a render are three different things. A smaller resident allocation does not make the source asset unnecessary. For the actual build and storage arrangement, establish where generated files live and which process is responsible for preparing them. Do not assume that a second machine inherits a cache merely because it receives the same scene file.

The same texture improvement, two different totals

Consider an invented arithmetic exercise. We divide each scene’s memory into two simple, additive parts: textures and everything else. Suppose the texture portion falls to one quarter of its original size. This is not a prediction of Cycles behaviour; real allocations and peaks can be more complicated.

Fictional memory components, in GiB — not Blender benchmark results
Component Texture-heavy example Other-data-heavy example
Everything except textures 2 6
Textures before 6 2
Total before 8 8
Textures after the assumed reduction 1.5 0.5
Total after 3.5 6.5
Reduction in this simplified total 56.25% 18.75%

Both examples reduce their texture component by 75%. Yet one total falls by more than half, while the other falls by less than one fifth. The denominator changes the claim. Reporting only the texture percentage conceals the distinction that may matter most to the person trying to finish a shot.

Nor does the table establish that either scene will fit a particular device. It omits real allocation behaviour and other resource limits. Its purpose is narrower: before investigating a memory problem, identify which component is responsible instead of assuming the most visible collection of image files is the largest allocation.

Separate the first run from a prepared run

A useful local comparison can distinguish three conditions: the reference configuration, a cache-enabled run that must prepare the needed files, and a cache-enabled run with those files already prepared. Keep the scene revision, frame, camera, device and output settings the same. Record preparation and render time separately if the question requires that distinction.

For example, a producer asking how long a newly received scene takes to deliver needs to account for preparation. An artist repeatedly reviewing the same shot may instead care about a prepared workflow. Both questions are legitimate. A result becomes misleading when a prepared timing is presented as the complete first-run cost.

Use a named memory measurement as well. Peak device memory, host memory and generated file size are not interchangeable readings. Save the method and the run conditions with the result; if a measurement cannot distinguish the components, report that limit rather than estimating a texture-only figure from the total.

Follow the camera before accepting the saving

For the invented interior shot, compare the wider view and the close-up because they ask different things of the textures. Inspect details that matter to delivery, such as the legibility of the label and the appearance of the wall. A lower output resolution would also change the request for detail, so changing it alongside the cache setting would no longer isolate the same comparison.

A still frame leaves the rest of an animation untested. Choose additional frames for a reason — a changed view, a different visible asset or a known difficult part of the shot — and state what remains unchecked. This keeps the exercise manageable without allowing one successful frame to stand in for an entire sequence.

The outcome should identify the shot conditions, preparation cost, chosen memory metric and image check. It may reveal a useful saving, a tolerable trade-off or a different bottleneck. Any of those is more actionable than a percentage detached from the scene that produced it.

Mechanism and controls follow the linked developer account and versioned manual. The two memory examples and interior shot are invented teaching examples. This publication has not reproduced the developer benchmarks or measured a production scene.