Tiny AutoEncoder Previews
यह कंटेंट अभी तक आपकी भाषा में उपलब्ध नहीं है।
While a model denoises, InvokeAI shows a preview of the image after each step. By default the preview is a linear projection of the latents: one pixel per latent, upscaled, so it is blurry and its colours are approximate. A Tiny AutoEncoder (TAE) is a small decoder by madebyollin (MIT license) that turns the same latents into a preview close to the final image, in a few tens of milliseconds on an RTX 4090. It needs a GPU: on CPU-only installs it is not recommended (see Cost and fallback).
TAE previews change only what you see while a generation runs. The finished image is decoded by the model’s VAE as always and is identical with the setting on or off.
Enabling TAE previews
Section titled “Enabling TAE previews”-
Set the option in
invokeai.yaml, then restart InvokeAI:tiny_autoencoder_previews: trueIt is a server setting and is off by default. See All Settings.
-
In Model Manager → Starter Models, install the Tiny AutoEncoder for the families you generate with (search for preview decoder). Each one is 5 to 31 MB. Filtering by your model’s family also shows its decoder, even when the decoder is filed under another family, and a selected bundle names its missing decoder in its summary line.
| Starter model | Previews for |
|---|---|
| TAESD (SD 1.x/2.x preview decoder) | SD 1.x, SD 2.x |
| TAESDXL (SDXL preview decoder) | SDXL, SDXL Refiner |
| TAESD3 (SD3 preview decoder) | SD 3 |
| TAEF1 (FLUX.1 preview decoder) | FLUX.1, Z-Image |
| TAEF2 (FLUX.2 preview decoder) | FLUX.2, FLUX.3, ERNIE-Image, Ideogram 4, MOSS-Image |
| TAEQI2.1 (Qwen-Image-2.1 preview decoder) | Qwen-Image-2.1 |
| TAEW2.1 (Wan 2.1 VAE preview decoder) | Wan 2.2 T2V and I2V A14B, Qwen-Image, Qwen-Image Edit, Krea-2, Anima |
| TAEW2.2 (Wan 2.2 TI2V-5B preview decoder) | Wan 2.2 TI2V-5B |
| TAELTX2.3 (LTX-2.5 preview decoder) | LTX-2.5 |
A TAE serves every family that denoises in its latent space, so you need only one per row. To install a TAE file you already have, see Preview decoders for the files InvokeAI accepts.
FLUX.1, FLUX.2, FLUX.3, ERNIE-Image, MOSS-Image, Anima and LTX-2.5 preview the image the model predicts, so the preview is clean from the first step. Z-Image, SD 3, Qwen-Image-2.1, Qwen-Image, Krea-2, Wan 2.2 and Ideogram 4 preview the current noisy latent: early previews show sharp noise, which clears as denoising proceeds. SD 1.x/2.x and SDXL depend on the scheduler: DDIM, DDPM, Euler, Euler a, Heun, LMS, KDPM2, KDPM2 a and DPM++ SDE report their prediction and preview cleanly; DPM++ 2M, 2M SDE, 3M and 2S, UniPC, DEIS, PNDM, LCM, TCD and ER-SDE preview the noisy latent.
A video preview shows the middle frame of the clip. For Wan 2.2 the TAE decodes it together with the frame before it, which it needs as context for a sharp result; LTX-2.5 decodes the middle frame alone.
CogView4 keeps the linear preview. MiniMax H3 always previews with its own built-in decoder, whatever this setting says.
Cost and fallback
Section titled “Cost and fallback”Only the preview frames that are actually sent to the UI are decoded, between two denoising steps. Previews are capped on their long edge: at 512 pixels for TAEW2.1, whose decode would need about 2 GB of VRAM at 1024×1024, and at 1024 pixels for every other TAE. Larger generations are previewed at reduced resolution and shown at full size.
Measured on an RTX 4090 in fp16, decoding one latent at the cap:
| TAE | Cap | Time per preview | VRAM while it decodes |
|---|---|---|---|
| TAESD, TAESDXL, TAESD3, TAEF1, TAEF2 | 1024 pixels | about 25 ms | about 0.65 GB |
| TAEQI2.1 | 1024 pixels | about 25 ms | about 0.16 GB |
| TAEW2.1 | 512 pixels | about 10–20 ms | about 0.54 GB |
| TAEW2.2 | 1024 pixels | about 24 ms | about 0.54 GB |
| TAELTX2.3 | 1024 pixels | about 29 ms | about 0.27 GB |
A Wan 2.2 video preview, which decodes two latent frames, takes under about 32 ms. InvokeAI reserves up to 617 MiB of VRAM for it from 832×480 to 1280×720 and 1093 MiB at 1024×1024; on an RTX 4090 it used up to 599 and 1064 MiB. Without a GPU a preview takes about 2 seconds, so TAE previews are not recommended on CPU-only installs.
InvokeAI spaces TAE preview frames so that previews cost at most about a tenth of the denoising time. On fast models the preview therefore updates a few times per second rather than after every step. In practice the cost is lower: on an RTX 4090, denoising a 1024×1024 SDXL image in 30 steps took about 4.44 s with TAE previews against 4.18 s without, roughly 6% more.
InvokeAI falls back to the linear preview by itself, and the generation always continues:
- No matching TAE installed: linear previews. The server log says once which starter model to install, and under which family the starter list shows it.
- The TAE does not fit in VRAM next to the running model: linear previews for that generation, noted once in the server log. This is the case whenever the main model is only partially loaded into VRAM, as on low-VRAM cards.
- Loading or decoding the TAE fails: a warning with the error in the server log, once per kind of failure, then linear previews for the rest of that generation.