WiktorMatuszek/MOSS-Image-SFT-VAE-FP16

MOSS-Image-SFT — original Flux2 VAE FP16

Release status

Experimental VAE component — further integration testing recommended. The 4-image same-latent precision comparison is a decoder test, not verification of every full MOSS-SFT pipeline or Polish typography.

Benchmark interpretation: The image gallery and measurements are reference examples for visual inspection, not a required reproducibility package or guarantee of quality, speed, or VRAM usage on another setup.

A reduced-precision copy of the original MOSS-SFT Flux2 VAE with the same architecture. This changes weight storage dtype only; it is not a new trained autoencoder.

Property	Saved and measured
Original source	OpenMOSS-Team/MOSS-Image-SFT at fc5de0d56fc910cbca6f6625a70f5f582b26deb1
Original FP32 VAE weights	237.962 MiB
Saved FP16 weights	118.994 MiB (49.99% smaller)
Converted tensors	250 floating tensors, original 251 state keys preserved
Weight MSE vs FP32	8.5461×10⁻¹¹
Real image comparison	4 original MOSS-SFT outputs, same re-encoded latents; median PSNR 63.64 dB, SSIM 0.999790, LPIPS 0.0000328

The original FP32 VAE first re-encoded each genuine BF16/SFT50 PNG (anime / bakery, seeds 42 and 2026). Each exact latent was then decoded with FP32, FP16 and BF16 VAE weights. This measures only decoder precision, not end-to-end diffusion fidelity or correct Polish text.

Real paired images

Same original latent, panels: input / FP32 / FP16 / BF16.

Full four-case VAE real-image audit and per-scene results.

Use in original MOSS-SFT

In a compatible MOSS-SFT DiT checkpoint, run:

python generate.py --vae-repo WiktorMatuszek/MOSS-Image-SFT-VAE-FP16 --prompt-file example_layout.json --output image.png

The unchanged tokenizer, text encoder, sampler and DiT are loaded as usual. A whole-pipeline combination with quantized Qwen / DiT requires separate verification; component-level visual similarity alone is insufficient.

Integrity: original/saved weight hashes, weight-domain measurements, exact original-image comparisons. Other hardware and model measurements are in the benchmark archive.

Attribution: original Flux2 VAE in OpenMOSS-Team/MOSS-Image-SFT, Apache-2.0 (LICENSE). Independent precision conversion, not an official model release.