MossAD-Demo — int8 ConvRot (ComfyUI)
This is not my model. It is an int8 ConvRot quantization of RicemanT/MossAD-Demo, a full finetune of MOSS-Image Base by RicemanT. All credit for the training goes to the original author; the license (Apache-2.0) is theirs. For generation settings, the artist trigger list and support for the author, see the original model card.
Source file: 1024-lr1e5/MossAD-Demo-1024-lr1e5.safetensors (sha256 009962c3…8effe,
repo revision 7806723), converted unchanged except for the quantization. The output is
5.0 GiB; the bf16 original is 10.2 GB. It is made for ComfyUI's native MOSS-Image support.
| file | format |
|---|---|
MossAD_int8_convrot_attn.safetensors |
DiT only: int8 ConvRot linears, bf16 everything else, int8 attention |
Its layout is the same as moss_image_base_int8_convrot_attn / moss_image_sft_int8_convrot_attn:
- the four linears of all 42 blocks (
proj.qkv,proj.o,proj.ffn_up_gate,proj.ffn_down; q/k/v fused) areint8_tensorwisewithconvrot: true. Each 256-column group is rotated by a normalized regular Hadamard matrix, then quantized to int8 per row with fp32 scales. Mean relative weight error is ≈0.9% per layer. - embedders, the text refiner, modulation, norms and the output head stay bf16.
- each block carries a
comfy_attention.configthat asks for comfy-kitchen int8 attention.
Usage
Put the file in models/diffusion_models/ and load it with Load Diffusion Model
(core ComfyUI, recent enough to have MOSS-Image and int8 ConvRot). Use it with the usual
MOSS-Image text encoder and the small FLUX.2 VAE decoder from the MOSS-Image repo.
The int8 kernels need an NVIDIA GPU with SM ≥ 7.5.
50 steps, CFG 3 (block 22 skipped in the negative branch), shift 5, seed 42, 1024×1024, RTX 3060 12 GB.
sha256 bc43ea1680f37bb62cf9a0a92934c951ebdf6162afc468ec74923b6e2c04b49c