WiktorMatuszek/MOSS-Image-SFT-Qwen3.5-FP8-WO

MOSS-Image-SFT — original Qwen3.5-9B FP8 weight-only encoder

Release status

Experimental encoder component — further testing required. The saved 248/248 converted Linear matrices are verified, but the 20-prompt embedding-equivalence acceptance gate remains not passed / incomplete. The impact on generated imagery and typography is not fully characterized.

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.

This is a true packed FP8 conversion of the original MOSS-SFT Qwen text encoder. It is an optional image-generation component, not a standalone LLM, LoRA or fine-tune.

Property	Saved and measured
Original source	OpenMOSS-Team/MOSS-Image-SFT at fc5de0d56fc910cbca6f6625a70f5f582b26deb1
Original BF16 encoder weights	15.633 GiB
Stored FP8 weight-only encoder	9.195 GiB
Coverage	248/248 selected Linear projections; all 759 source keys
Weight distortion vs BF16	MSE 1.248×10⁻⁷, SQNR 31.54 dB
Frozen 20-prompt embedding comparison	Median cosine 0.998061, minimum 0.984930
Validation outcome	Not passed / incomplete for equivalence to original conditioning

The original tokenizer, pipeline, DiT, scheduler and VAE remain required. Weight reduction does not certify faithful Polish text or faster generation.

Use it

From a compatible MOSS-SFT DiT repo, run:

python generate.py --encoder-repo WiktorMatuszek/MOSS-Image-SFT-Qwen3.5-FP8-WO --prompt-file example_layout.json --output image.png

Evidence and limitations

The complete source and saved weight receipt, per-Linear weight errors and 20 original-versus-packed prompt embeddings provide reproducible checks. Exact per-prompt CSV.

A separate four-pair, 50-step study combining FP8-WO Core DiT, this FP8 Qwen and FP16 VAE measured median SSIM 0.72178 / LPIPS 0.08329 vs original BF16, showing additional visual divergence versus DiT-only. See the original paired images and canonical archive; do not generalize to other transformers.

Attribution: conversion of the original encoder in OpenMOSS-Team/MOSS-Image-SFT, Apache-2.0 (LICENSE); independent and not official.