Downloads · 30 days
35
32% of all-time downloads
mcp-tool-shop/portlight-ships-anatomy-lora
portlight-ships-anatomy-lora is a text-to-image model from mcp-tool-shop. Use it when you need an image from a text prompt. It is set up for diffusers. The card lists the license as mit.
A Qwen-Image LoRA that knows ship anatomy — fourteen distinct hull types, each rig traceable to a documented historical vessel.
Downloads · 30 days
35
32% of all-time downloads
All-time downloads
108
Public
Repo size
295 MB
Likes
0
Public
Click a slice to open those files.
.safetensors295 MB · 100%
From the Hugging Face model README
A Qwen-Image LoRA that knows ship anatomy — fourteen distinct hull types, each rig traceable to a documented historical vessel.
Trigger token: plsvessel
| Base | Qwen/Qwen-Image (bf16, unquantised) |
| Rank / alpha | 16 / 16 |
| LR / optimiser | 1e-4, adamw8bit, flowmatch |
| Steps | 6000 (recommended checkpoint) |
| Dataset | mcp-tool-shop/portlight-ships-anatomy — 336 pairs, pristine only |
| Trained on | 1× RTX PRO 6000 Blackwell 96 GB, ~43 GB VRAM resident |
Base Qwen-Image renders a xebec, a brig, a brigantine and a junk as four versions of the same generic square-rigger. Lateens, battened lugsails, and the rig distinctions between two-masted vessels simply are not in the model.
But it renders a wrecked ship convincingly with no help at all.
So the base model has damage and lacks anatomy — and that asymmetry is the whole design. This adapter teaches only anatomy, from 336 pristine renders, and borrows the damage the base already owns.
That works. There is not one damaged image in the training set, and the adapter still infers
damage correctly — verified with SigLIP2 (a different model family from the generator):
wrecked 0.982 vs intact 0.255 on trained output, 0.069 / 0.118 on a pristine render.
Rotation is the hard part. Diffusion models cannot generate ship headings — a ship is a near-symmetric object with no directional cue, so "port bow" and "starboard quarter" both return the same broadside. Three prior attempts failed this way.
So the training images are not generated per-heading. Each hull is one generated plate → image-to-3D (Tripo v3.1) → textured mesh → Blender turntable at 8 headings × 3 elevations. Identity holds across headings because it is the same mesh.
Captions are derived from canon (rig_plan names every mast and sail), never
hand-written — caption drift is how a LoRA learns fuzzy concepts.
carrack · fluyt · cog · galleon · frigate · brig · barque · brigantine · schooner · sloop · caravel · xebec · galley · junk
Stated plainly, because the failures are as reusable as the successes.
from diffusers import DiffusionPipeline
import torch
pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights("mcp-tool-shop/portlight-ships-anatomy-lora")
image = pipe(
prompt=(
"plsvessel, a xebec, 3-masted sailing ship, lateen-rigged foremast, mainmast and "
"mizzenmast, 4 sails set, pristine undamaged condition, port broadside view, "
"game asset render on a plain background"
),
width=1024, height=1024,
num_inference_steps=25, true_cfg_scale=3.0,
).images[0]
Caption shape that matches training:
plsvessel, a <class>, <N>-masted sailing ship, <derived rig phrase>, <N> sails set,
<signature features>, pristine undamaged condition, <heading>, <elevation>,
game asset render on a plain background
Damage and patina are not in any training caption — they come from the base model, so
phrase them naturally (heavily damaged, mainmast snapped and fallen, canvas torn to tatters).
Full pipeline, canon schema and hard-won settings: mcp-tool-shop-org/portlight-ships
MIT