aqualaguna/MOSS-Image-Edit-OpenVINO

MOSS-Image-Edit for OpenVINO (Intel Arc GPUs), experimental

OpenMOSS-Team/MOSS-Image-SFT with the community edit LoRA chinoll/Moss-Image-Edit (rank 64, project Chenkin-x/Moss-Image-Edit) merged into the weights and converted to OpenVINO IR. Tested on an Intel Arc Pro B60 24 GB with OpenVINO 2026.4. Staged and tested, never put into production.

Licences: base model Apache-2.0; the edit project's code is GPLv3; the LoRA repo states no weight licence. These merged weights inherit whatever terms apply upstream: check both before using them. Not affiliated with either project.

Needs the text-to-image repo

Only the edit DiT is here. The text encoder, VAE decoder and tokenizer are identical to aqualaguna/MOSS-Image-SFT-OpenVINO (te/, vae/, tokenizer/).

folder	content	size
dit_edit_fp16/	text_refine, edit_a, edit_b, edit_c (fp16), vae_encoder_f32, host weights	9.8 GB
dit_edit_int8/	same, NNCF per-channel INT8_ASYM, ffn_down kept fp16	6.4 GB
code/	pipeline + server with both text-to-image and edit samplers (edit_ops.py, edit_export.py)	<1 MB

The reference image enters as timestep-0 prefix tokens whose K/V are computed once per request.

What works

• Box-JSON layout edits localise: 88–99.5 % of the pixel change lands inside the given box; outside the box ~30 dB PSNR to the input.
• Text is only correct when the exact words are written in the prompt. Asked for e.g. an "infographic" without the text, it draws plausible-looking but garbled lettering.
• One reference image (the LoRA accepts up to 3; more were not measured). Capped at 1024×1024.

Speed and memory (Arc Pro B60, 1024², 30 steps, guidance 4)

	time	peak GPU memory
fp16	~105 s	21.6 GB
int8	~95 s	17.6 GB

Benchmark (2026-11-02, 12 edits fp16 + 6 int8 over 5 scenes: living room, kitchen, park, street, cafe, anime; box-JSON and plain-text prompts; 1024², 30 steps, guidance 4, 4096 reference tokens, all outputs finite):

	total per edit	s/step	DiT load	text encoder	ref prefill	VAE	peak GPU (peak_gpu_usm_device_mb)
fp16	103.5–106.6 s	2.96–3.02	6.1–6.9 s	5.7–6.4 s	1.4–1.7 s	0.4–0.9 s	20.9k (short text prompt) – 21.6k (box JSON)
int8 keepdown	92.7–95.6 s	2.80–2.84	6.4–7.3 s	cached*	1.4–1.5 s	0.4 s	17.6k

* the int8 runs reused the prompt embeddings from the fp16 runs; add ~6 s for a new prompt.

The edit DiT is unloaded after each call: OV scratch memory grows per distinct token count (reached 27.8 GB without that). The fp16 residual is scaled by 1/8 inside the graph to avoid fp16 overflow.

Use

Same requirements as the text-to-image repo. Put the folders as models/{te,tokenizer,vae,dit_sft_edit,dit_sft_edit_int8_keepdown} (rename dit_edit_fp16 → dit_sft_edit, dit_edit_int8 → dit_sft_edit_int8_keepdown).

import sys; sys.path.insert(0, "code")
import numpy as np
from moss_pipeline import MossImage
from PIL import Image

pipe = MossImage(models_dir="models", device="GPU", cache_dir="ov_cache", variant="edit-int8")
src = np.asarray(Image.open("in.png").convert("RGB").resize((1024, 1024)))
img, stats = pipe.generate_edit([src], "replace the sky with a sunset", width=1024, height=1024, steps=30, guidance=4.0, seed=1)
Image.fromarray(img).save("out.png")

code/server.py serves it over HTTP ("model": "edit" or "edit-int8" with an image field). Paths like /opt/mossimage/... are from the original box; adjust them.