MicroZoom INT4 on a 24 GB GPU - What Our RTX 3090 Ti Test Proved

60-second takeaway
MicroZoom's published setup targets an 80 GB A100 for training and a 40 GB A40 for inference. We made a bounded version execute on one 24 GB RTX 3090 Ti by quantizing the frozen FLUX transformer, T5 encoder, and ControlNet to INT4 while keeping LoRA adapters in BF16.
A six-step 1024 px run completed, saved every step, and peaked at 23,194 MiB of VRAM with only 919 MiB free. This proves short-run compatibility on our pinned fixture. It does not prove convergence, image quality, long-run stability, or safe shared-GPU operation.

Why we tested MicroZoom

MicroZoom is an object-specific generative pipeline from the University of Washington. It combines a full photograph with registered microscope close-ups, masks, texture label maps, prompts, and per-object LoRA training. Its multi-diffusion inference stage can then synthesize extreme-scale imagery.

That is more specialized than ordinary super resolution. MicroZoom is not a drop-in enhancer for a folder of existing images, and its generated microscopic detail should not be treated as recovered ground truth.

The official repository lists an 80 GB A100 or equivalent for training and a 40 GB A40 or equivalent for inference. Our practical question was narrower:

Can a frozen-weight INT4 port make a bounded MicroZoom inference and LoRA training path execute on a 24 GB consumer GPU?

The memory strategy

The upstream --use_8bit_adam option reduces optimizer-state memory. It does not quantize model weights. MicroZoom still has to accommodate several large components:

• the FLUX.1-dev diffusion transformer;
• the T5 text encoder;
• the Jasper FLUX.1-dev upscaler ControlNet;
• CLIP and the VAE;
• activations for the current crop; and
• trainable LoRA branches plus optimizer state.

Our experimental path used Optimum Quanto to quantize the three large frozen components to INT4. The texture-expanded input layer stayed in BF16 because it had to be resized before quantization. Rank-2 transformer and ControlNet LoRA adapters also stayed in BF16.

The training configuration was deliberately small: