TR4F/crackio-retouch

Crackio Retouch eyes

A LoRA on Liquid AI's LFM3-VL-2B that maps a picture for a photo editor. It lists the separate things in it (people, animals, objects, the sky, the floor) with a box each, and, for one thing cut out close, its parts (hair, face, each piece of clothing; a car's wheels, lights and doors). Crackio's Retouch uses it on your own GPU so you can click any single thing in a picture and change just that. The boxes stream in as they're written: on an RTX 5090 the first one shows in under a second, and 23 things take about 8 s.

Files

File	What it is
lfm3-vl-2b-lora-retouch.safetensors	The LoRA as Crackio's ComfyUI bridge loads it, from ComfyUI/models/lfm3/
peft/	The same LoRA as a PEFT adapter (rank 32, alpha 32, on the language model's attention and MLP)

Use it

In Crackio, setting up Retouch in Settings → Tools with Crackio's eyes puts LFM3-VL-2B and this LoRA on your GPU, and Retouch's Finds things with (Auto or Your GPU) uses it. Crackio's bridge serves it at POST /crackio/lfm3 (or /crackio/lfm3/stream) with "lora": "retouch".

With PEFT, load LiquidAI/LFM3-VL-2B and apply peft/ on top. It was trained on Crackio's two prompts (in src/lib/server/detect.ts), on pictures up to 1536 px on the long side, and answers in JSON with boxes as [x0, y0, x1, y1] in 0–1000 of the picture:

• things: {"things": [{"label": "pigeon", "detail": "grey pigeon on the left", "box": [...]}]}, the most prominent first, up to 24;
• parts of one thing, on a close crop: {"whole": [...], "parts": [{"label": "beak", "detail": "orange beak", "box": [...]}]}, 3 to 10.

How it was trained

Version 1: one epoch (435 steps) on 3,473 examples, 1,473 whole pictures (things) and 2,000 close crops (parts).

Labels. Every label and box was written by Qwen3-VL-8B-Instruct (Apache-2.0), asked with Crackio's own prompts. No closed model's answers were used as labels.

Pictures. The 1,473 pictures it practised on are from the Crackio author's own library, pictures they made themselves in Crackio. The 2,000 close crops were cut from them.

Results. These were measured on 60 held-out pictures and 60 held-out part crops. Each answer was scored against Qwen3-VL-8B's boxes after SAM 3.1 tightened them; a box counts as found when it overlaps one of those at IoU ≥ 0.5.

	This LoRA	LFM3-VL-2B alone
Things: answers that parse	60/60	52/60
Things: found per picture (gold 14.8)	14.4	7.6
Things: recall	.53	.33
Things: precision	.55	.73
Things: mean IoU of the matches	.87
Parts: recall	.68	.20
Parts: precision	.69	.50

The untuned model finds about half as many things, though more of the ones it finds are right.

Known limits.

• It copies its teacher's habits, such as listing clothes as separate things.
• It finds at most 24 things per picture.
• It draws boxes, not outlines: Crackio cuts the outline with SAM 3.1 once you click a thing.

Licence

LFM3-VL-2B and this adapter, a Derivative Work of it, are under the LFM Open License v1.0 (see LICENSE). Commercial use is licensed only for companies with under US$10M annual revenue. See NOTICE for attribution and what was changed.