WAS/Krea2_MOSS_graft_lora

Krea 2 MOSS Graft LoRA

TenStrip's MOSSkr2 graft (Krea 2 attention and MLP weights grafted into MOSS-Image SFT) as LoRA and patch files. They load onto any MOSS-Image SFT file, int8 included, in place of the 10.2 GB MOSSkr2 checkpoint. Each file is MOSSkr2 minus Comfy-Org's moss_image_sft_bf16.

Files

File	Size	Share of the graft kept
krea2_moss_graft_diff_exact_bf16.safetensors	6.15 GB	all of it
krea2_moss_graft_lora_r256.safetensors	0.76 GB	23%
krea2_moss_graft_lora_r128.safetensors	0.38 GB	14%
krea2_moss_graft_lora_r64.safetensors	0.19 GB	8%
krea2_moss_graft_lora_r32.safetensors	0.09 GB	5%

Share kept is the part of the weight change a file reproduces, averaged over the 168 changed tensors. The exact file on MOSS-Image SFT bf16 gives MOSSkr2's images. The LoRAs give a different result with more fine texture; see the comparison.

Use

ComfyUI with native MOSS-Image (core commit e087d8f6c or later):

1. Load Diffusion Model: a MOSS-Image SFT file, moss_image_sft_bf16 or moss_image_sft_int8_convrot.
2. LoraLoaderModelOnly: one file from this repo, strength 1.0.

Both fused-qkv (Comfy-Org) and split (OpenMOSS) checkpoints load these files. They were extracted against SFT and have not been tested on MOSS-Image Base.

Comparison

All images use one prompt and one seed at 1664x1248: euler, normal, 25 steps, CFG 3, SkipLayerGuidanceDiTSimple on single layer 22, seed 42, qwen3.5_9b_int8_convrot text encoder, fk9_vae. The base is moss_image_sft_int8_convrot unless the label says bf16.

A wide cinematic photograph of an elderly fisherman mending a net on a weathered wooden pier at dawn. Behind him a busy harbour stretches back: moored fishing boats with peeling paint, stacked crates and coiled ropes, gulls over the water, a hillside town of terracotta roofs and narrow stairways, laundry lines strung between houses, and a white lighthouse on the far headland. Low warm sun, light mist on the water, fine detail throughout the background.

Model	Toward MOSSkr2	Fine detail
SFT int8	0	47.6
MOSSkr2 bf16	1	54.7
SFT bf16 + exact	0.91	54.8
SFT int8 + exact	0.72	57.3
SFT int8 + LoRA r256	0.32	58.7
SFT int8 + LoRA r128	0.23	61.4
SFT int8 + LoRA r64	0.17	65.4
SFT int8 + LoRA r32	0.04	61.6

• Toward MOSSkr2 is 1 - MSE(image, MOSSkr2) / MSE(SFT int8, MOSSkr2) at quarter scale: 0 is plain SFT int8's image, 1 is MOSSkr2's.
• Fine detail is the mean absolute Laplacian of luma at full size, times 1000.
• SFT bf16 + exact matches MOSSkr2 at 29.7 dB PSNR (quarter scale). On int8, the exact file keeps MOSSkr2's composition. Every LoRA rank gives a third composition with more fine texture than MOSSkr2, and no grain at 1:1.
• One seed and one prompt. Any change to the weights re-rolls the layout, so a different composition is not a quality difference.

Full-size grid: images/comparison_grid_full.png. 1:1 crops: town, boats.

What the files contain

MOSSkr2 changes four projections in all 42 layers. Everything else is bit-identical to MOSS-Image SFT.

Tensor	Changed
proj.q	all 24 heads
proj.k, proj.v	heads 0 to 11
proj.ffn_up_gate	rows 8192 to 16383
proj.o, proj.ffn_down, norms, embedders, modulation	unchanged

• MOSSkr2 also scales each Q and K head by 0.48 to 0.62. MOSS-Image applies a per-head RMSNorm to Q and K, which cancels that scale, so the files leave it out.
• LoRAs: truncated SVD of each tensor's difference, no clamping, bf16, no alpha (scale 1).
• Exact: the full difference as .diff keys, bf16.

Licence

These files are a Derivative of Krea 2 under the Krea 2 Community License Agreement (LICENSE.pdf). Downloading, using or redistributing them binds the user to that agreement, including its Acceptable Use Policy, content-filtering requirement and commercial revenue threshold. The MOSS-Image portions are Apache-2.0 (LICENSE-MOSS-Image). See NOTICE.

Not affiliated with or endorsed by Krea, the OpenMOSS Team or TenStrip.

Credits

• Graft: TenStrip, MOSSkr2
• MOSS-Image: OpenMOSS Team
• Krea 2: Krea