Spectrallic/MOSS-Image-SFT-8step-DMD-int8-convrot

MOSS-Image-SFT 8-step DMD — INT8 ConvRot

OpenMOSS-Team/MOSS-Image-SFT with the lodestones 8-step DMD LoRA merged in, quantized to INT8 ConvRot. It is a drop-in replacement for moss_image_sft_int8_convrot.safetensors from Comfy-Org/MOSS-Image: the tensor names, dtypes, shapes and quantization format are the same. The only difference is that the distillation is baked in, so you don't need a LoRA loader.

Layout: like the Comfy-Org file, the attention projections are fused into layers.N.proj.qkv, and each quantized linear has its own .comfy_quant tensor. Some other MOSS-Image INT8 ConvRot releases keep proj.q, proj.k and proj.v as separate tensors instead. Loaders that expect the Comfy-Org layout, such as Nox, load this file as-is.

Examples

Rendered with Nox using this checkpoint at CFG 1 with an empty negative prompt, using the Euler flow sampler on the release sigma schedule (shift 5). The text encoder is qwen3.5_9b_int8_convrot, from Comfy-Org/MOSS-Image.

Steps	CFG	Size	Seed
10	1.0	1056×1408	4873513220996790616

An ancient floating city hangs above a vast sea of cumulus clouds. Its intricate, hand-drawn architecture is rendered in a soft pastel color palette, with a waterfall cascading over the city's edge into the clouds below. Shafts of golden sunlight break through the cloud cover, illuminating the ultra-detailed digital painting.

Steps	CFG	Size	Seed
8	1.0	1536×1536	11841263543751983967

A realistic closeup shot of an futuristic floating metropolis hangs above a vast sea of cumulus clouds, it has detailed complex architectural buildings spanning the floating island.

with a waterfall cascading over the city's edge into the clouds below. Thousands of small lights illuminate the city

Shafts of sunset break through the cloud cover, illuminating the ultra-detailed digital painting.

How it was made

1. Source: the official BF16 SFT transformer (OpenMOSS-Team/MOSS-Image-SFT, revision f972bd4b) and lora.safetensors from lodestones/MOSS-Image-SFT-8-steps-dmd (revision f95df83b; update 2600, EMA, not lora_latest). The LoRA is rank 128 / alpha 128, merged at strength 1.0 into every block's proj.{q,k,v,o,ffn_up_gate,ffn_down}.
2. Merge in FP32, and keep it FP32: compute W' = W + (alpha/rank) · up @ down in FP32, then concatenate proj.{q,k,v} into proj.qkv (Comfy-Org layout).
3. Quantize: using convert_to_quant v1.3.4:

   ctq -i merged_f32.safetensors -o out.safetensors --comfy_quant --save-quant-metadata \
       --int8 --scaling-mode row --convrot --convrot-group-size 256 --simple --optimizer original \
       --low-memory --exclude-layers '^(?!layers\.\d+\.proj\.)'
   

   Only the 168 block linears (layers.*.proj.{qkv,o,ffn_up_gate,ffn_down}) become INT8: row-wise absmax scales after a group-256 regular Hadamard rotation. Everything else stays BF16, as in the Comfy-Org file. The comfy_attention.config markers are copied from the Comfy-Org checkpoint.

Credits & license

• Base model: OpenMOSS-Team/MOSS-Image-SFT (Apache-2.0)
• Distillation LoRA: lodestones/MOSS-Image-SFT-8-steps-dmd (Apache-2.0)
• Reference quantization layout: Comfy-Org/MOSS-Image (Apache-2.0)
• Quantizer: silveroxides/convert_to_quant

This derivative is released under Apache-2.0. The distillation LoRA is a work in progress upstream, so check the LoRA repo for newer updates.