Sido/ImageEdit-MOSS-DMD16-WebGPU

MOSS W4 with unmerged FP32 DMD LoRA for browser inference

An experimental browser conversion prepared from these pinned upstream assets:

• smile-forge/moss-image-onnx at 1b6c0496cfcbfa53426245bd084d01d85ca0326e (Apache-2.0).
• lodestones/MOSS-Image-SFT-8-steps-dmd at f95df83b5e90ae91ea412434b3f9b7f1c8c6314f (Apache-2.0).
• OpenMOSS-Team/MOSS-Image-SFT at f972bd4b4b9e3852133cc5da954d87bc32cd8fb9 (Apache-2.0).

Conversion and inference profile

The image transformer was converted from the Smile Forge W8 ONNX export to experimental W4 RTN, block size 32. The DMD lora_latest.safetensors adapter is applied unmerged in FP32 at strength 0.7 over the unchanged W4 base files. The Qwen3.5-9B text encoder uses FP16 weights/activations, FP32 projection accumulation and browser attention state; it encodes the user's prompt at run time, with a 128-token limit. No precomputed caption embeddings are included. The decoder is the original FP16 ONNX export. Original external filenames are retained so each ONNX graph resolves the adjacent binary files.

The *-fullhd.onnx derivatives split image attention queries into 16 chunks and decoder queries into 32 chunks with independent dynamic balanced Slice nodes. This avoids high-output-count Split kernels exceeding WebGPU storage buffer limits. Every query retains all original keys/values. The first two text-only MHA refiners remain unsplit. No denoiser/decoder weights or LoRA matrices are modified by this packaging step. Modified ONNX graph metadata records this transformation. The legacy fullhd names are retained; chunks support varying sequence lengths at least the chunk count, including small and uneven final chunks.

The intended app profile is 16 release shift-5 Euler steps, CFG disabled, guidance 1. This is not the upstream data-space DMD generator/sampler. The minimum host memory setting is 32 GiB, with 64 GiB recommended. Runtime target is onnxruntime-web 1.30.0 with WebGPU and shader-f16. Device compatibility, quality and all image sizes require real GPU checks; ONNX validation and CPU attention equivalence tests alone do not establish end-to-end support.

Files, verification and publication

publication-plan.json records the public runtime inventory, byte counts, SHA-256 values, source pins and conversion provenance. text/text.json is the unaltered conversion manifest; its unused noise.f32 fixture record does not require shipping that fixture. Runtime files omit every fixed noise/caption fixture. README, LICENSE and NOTICE are public attribution files and do not need to be cached by the app. The offline packager never uploads or authenticates. After upload, supply the real repository and immutable uploaded commit to scripts/prepare-moss-assets.py --repository OWNER/REPO --revision COMMIT to generate the app's manifest-draft.json; no remote repository is invented.

License and attribution

Upstream model cards declare Apache-2.0 for the base model, ONNX conversion and DMD adapter. See LICENSE for the complete Apache License 2.0 and NOTICE for attribution and modification notices. This package is an independent conversion; it does not imply endorsement by OpenMOSS, Qwen, FLUX, Smile Forge or lodestones.