MOSS-Image-Edit-HQ

🚀 v3 is out — armV3 is live. Trained on ConceptEdit-12M (the dataset behind chinoll's LoRA) with quality + locality-purity filtering. Head-to-head vs chinoll/Moss-Image-Edit on 15 real portraits — full grids below, judge them yourself. Download: weights/armV3/armV3.safetensors (ai-toolkit format, loads natively in ComfyUI) · armV3-moss-edit-lora.safetensors (ComfyUI plugin format) · Try the demo Space · Blind eval grid below.

An open effort to build an instruction-based image-editing LoRA for OpenMOSS-Team/MOSS-Image-SFT that beats chinoll/Moss-Image-Edit on quality — every experiment, checkpoint, and side-by-side published, including the ones that lost.

V3 head-to-head vs chinoll (real portraits)

5 real Pexels portraits (never in training), matched settings (1024px long edge, 50 steps, CFG 3, seed 42). Columns: source → chinoll → armV3. The exact instruction is printed on every row — judge it yourself:

Batch 2 — 5 different portraits, different instructions (accessory, lighting, removal, background, outfit):

Batch 3 — 5 different portraits, the hardest surgical categories (eye color, freckles, eyeglasses) plus environment and style:

About these grids

Three batches, 15 real portraits total, never in training. Every row is source → chinoll → armV3 at identical settings (1024px long edge, 50 steps, CFG 3, seed 42), with the exact instruction printed on the row. We don't call winners — the grids are the evidence, judge them yourself. Full per-candidate images: akhaliq/wiki-eval-images.

Use in ComfyUI

Option A — native ComfyUI (recommended, no conversion)

ComfyUI merged native MOSS-Image Edit support (PR #17296, Oct 22) — it loads ai-toolkit-format edit LoRAs directly, which is exactly our checkpoint format:

Update ComfyUI past v0.42.1 (master/nightly — the PR merged after the tag).
Drop weights/armV3/armV3.safetensors into ComfyUI/models/loras/ (no conversion, no rename needed).
Standard LoRA loader after Load Diffusion Model (MOSS-Image single-file weights from Comfy-Org/MOSS-Image).
Encode with the experimental Text Encode MOSS-Image Edit node: connect VAE + reference images, set edit_format = ai-toolkit (that's our training layout — index_timestep_zero_isolated with KV cache; chinoll's LoRA uses the Moss-Image-Edit option instead).
References are downscaled to ≤384×384 automatically; 50 steps, CFG 3, Euler, SLG layer 22.

Option B — Chenkin-x plugin (moss-edit-lora format)

Use the armV3-moss-edit-lora.safetensors file (root of this repo) — it's in the moss-edit-lora format that the ComfyUI-Moss-Image-Edit nodes load, same as chinoll's adapter.

Install the plugin (ComfyUI Manager → search "Moss-Image-Edit", or git clone https://github.com/Chenkin-x/ComfyUI-Moss-Image-Edit into ComfyUI/custom_nodes/). Requires ComfyUI 0.42.3 or newer.
Base model: OpenMOSS-Team/MOSS-Image-SFT (+ the Qwen3.5 text encoder and Flux2 VAE the plugin's README lists).
Drop armV3-moss-edit-lora.safetensors into ComfyUI/models/model_patches/ (NOT models/loras/).
Download a ready workflow (LoRA filename and settings filled in):
- workflows/armV3_workflow_single.json — single reference (most edits)
- workflows/armV3_workflow_multi.json — up to 4 reference images

Recommended settings (what our evals and demo use): flow-matching Euler sampler, shift 5.0, 50 steps, CFG 3.0, reference at source resolution, output up to 1024px. For edits that land too weak, raise CFG to 3.5–4.

Troubleshooting

"Load Model Patch failed" → the LoRA is in the wrong folder. It goes in ComfyUI/models/model_patches/, not models/loras/.
Nodes won't load / "requires 0.42.3" → upgrade ComfyUI to 0.42.3+.
Edits too weak → raise steps to 50 and CFG to 3.5–4; make sure the ModelPatch strength is 1.0.

Status

Latest: armV3 is the recommended checkpoint. Try it in the demo Space.

Version	Recipe	Result
v1	warm-start + from-scratch, molbal multi-reference data, 1024px	❌ Lost the eval gate — chinoll preserves identity better. Kept for reproducibility.
v2 (armV2-lo)	warm-start, OmniEdit surgical data, lr 1e-5, 300 steps	✅ Matched chinoll, won atmosphere/style. Superseded by v3.
v3 (armV3)	continue from armV2-lo, ConceptEdit-12M surgical data (quality + purity filtered, 6,000 rows), lr 1e-5, 700 steps	✅ Published — wins small surgical edits (sunglasses/accessories/color), ties elsewhere.

Older eval grids (v1, v2): akhaliq/moss-edit-hq-eval.

Artifacts

weights/armV3/ — v3 LoRA checkpoints (495 MB each, peft format for ai-toolkit's MOSS-Image arch): final + step-250 + step-500, samples, config.
armV3-moss-edit-lora.safetensors — v3 in ComfyUI moss-edit-lora format.
workflows/ — ready ComfyUI workflows with the LoRA filename and settings filled in.
eval/real_grid_v3.jpg — the portrait head-to-head grid.
training/ — everything needed to reproduce: ai-toolkit arm configs, adapters/converters, training/eval orchestrators, dataset builders. Training runs on ostris/ai-toolkit @ 14f8f95 (moss_image arch, edit mode + KV cache, single GPU per arm).
Training data: akhaliq/moss-image-edit-conceptedit-packed (6,000 ConceptEdit-12M surgical rows, quality + purity filtered, TAR shards).
Live training metrics: trackio dashboard.

Notes for reproducing

chinoll's adapter uses split q/k/v LoRA keys; ai-toolkit's MOSS transformer fuses qkv. training/convert_chinoll_adapter.py converts with an exact rank-192 block-diag fusion (verified 336/336 tensor match) — warm-starting at rank 64 would silently truncate. The reverse (training/convert_to_moss_format.py) converts our checkpoints back to ComfyUI moss-edit-lora format.
v3 warm-starts from armV2-lo (INIT_LORA) rather than chinoll — building on v2's verified atmosphere/style gains while retraining on ConceptEdit data.
The v3 dataset filters surgical rows for local-edit purity (reject rows where source↔target differ on >65% of pixels) — OmniEdit-style global tonal shifts in training targets are what teach edit models to darken/re-tone whole images. env/style categories are exempt (global change is their point).
Gradient checkpointing is required at 1024px.
Keep the text encoder frozen (OpenMOSS guidance).