Polymesh compresses the modeling → retopology → UV → texturing pipeline into a single browser session. This page is the pitch: the problem, the team, the plan, a working prototype — and an API you can call right now.
Every studio we talked to drowns in the same bottleneck: concept art and blockouts pile up faster than 3D artists can turn them into clean, textured, engine-ready meshes.
Small and senior — the entire stack, from diffusion pipelines to the edge serving this page, was designed and shipped by this team. (Names and profile links below are placeholders — swap in the real ones.)
Stack in use across the product:
ComfyUI · Trellis2 · Qwen-Image-Edit · FLUX.2 klein · Python stdlib APIs · SQLite · Cloudflare (Tunnel, DNS, R2) · Caddy · Inter/JetBrains Mono design system
This is not a deck-ware project: the full pipeline — six-view diffusion, shape generation, texturing, retopology — runs live on this domain today and produces the GLBs in your Asset Library. What you click here is the prototype.
We build custom ComfyUI node chains, wire LoRA distillation (Lightning 4-step) and multi-angle conditioning ourselves, and swap model families mid-pipeline without touching the product layer. That's what keeps quality moving every week.
Dedicated RTX 5090 worker, hardened private networking, job queue with real status, automatic recovery. We control cost and latency end-to-end — no per-call reseller margins baked into the price.
Outputs are defined by what engines accept: poly budgets, UV sets, PBR maps, GLB/OBJ delivery. The product decisions in the studio (view controls, white-mesh preview, V1/V2 variants) come directly from art-pipeline experience.
Four gates. The first four are behind us and verifiable in the product; the fifth is where the next months go.
Validate the bottleneck with studios: concept→mesh turnaround is the top cost of pre-production art. Choose the wedge: single-image to game-ready textured asset.
End-to-end chain on one GPU: background removal → six-view synthesis (Qwen-Image-Edit + multi-angles LoRA) → Trellis2 shape+texture → decimate/UV → GLB out. Proved quality bar on real inputs.
Studio product: four generation modes, accounts with credit metering, private Asset Library (Cloudflare R2), API endpoints, admin panel. First external users generating meshes.
polymesh.ai on Cloudflare edge with a dedicated RTX 5090 worker, self-healing tunnel + watchdog. Billing runs in test mode; free Spark credits for everyone who signs up.
Live Stripe billing and Google sign-in, autoscaled multi-GPU queue with webhooks, PBR material controls, hosted REST SDK + docs portal, rig-ready exports (Mixamo-compatible).
Five stages, each already standing in production — this is the architecture the demo runs on, stage by stage.
Text→Image and prompt/reference Image Edit run FLUX.2 klein 9B. For 3D: RMBG cutout → six-view synthesis with Qwen-Image-Edit 2511 (multi-angles LoRA, Lightning 4-step distillation) → Trellis2 4B structured-latent shape + texture stages → diffusion paint passes.
Supervised services: the ComfyUI engine, an api-wrapper with a job queue (submit → poll → base64/URL outputs), and a read-only output server for the large 3D artifacts. Everything binds to loopback; the box is unreachable from the public internet except through the tunnel.
A hardened edge API owns sessions (hashed, expiring tokens), the credit ledger, per-account private asset storage (images and GLBs archived to R2), generation cost rules, history and an admin panel with usage events.
Cloudflare Tunnel publishes the site over HTTPS with no open ports; Caddy routes statically and reverse-proxies the API, the image engine and the 3D output server under one domain. A watchdog revives every component every minute — the stack self-heals.
Four modes (Text→Image, Image Edit, 3D Single View, 3D Multi View), live stage progress, six-view + white-mesh + textured-V1/V2 previews in a WebGL viewport, a recoverable asset library (interrupted generations are rebuilt from server-side output listings), profile, billing UI.
Image-conditional mesh diffusion (SLAT), multi-view consistency with pose LoRAs, texture-space diffusion painting, learned background removal, automatic retopology to a target budget, prompt-aware T-pose reshaping for character inputs.
Three captures, one under another — full Studio walkthrough, pipeline run, latest demo. Each plays on its own; nothing to switch.
Waiting for the final capture — it will play in this box. Nothing to install, nothing to refresh twice.
Videos are added manually from this page — no server access needed. Upload an MP4/WebM/MOV with a title and description; it lands at the top of the playlist above and the older clips stay switchable. Up to 400 MB per file (the public CDN edge passes roughly ≤ 100 MB — compress anything longer first). Requires being signed in.
Four modes, live worker, private library: polymesh.ai/studio.html. New accounts start with 200 free Spark credits, no card.
Ask the assistant anything about the product — formats, timings, pricing, the API, what runs under the hood. It answers from Polymesh's live knowledge base, with example questions preloaded below the input.
Try questions like:
Everything the studio does is a public HTTP call against https://polymesh.ai/api — queue a workflow, poll the job, fetch the result. Personal API keys are minted in Studio → Profile; sessions authenticate with a bearer token.
| Route | Purpose |
|---|---|
POST/api/auth/register · /login | account + bearer session |
POST/api/generate | queue a ComfyUI API-format workflow → {id} |
GET/api/result/{id} | job status + outputs (base64 / urls) |
POST/api/upload/image | multipart image upload → {name} |
GET/api/output-list/{folder}/ | discover 3D artifacts as they land |
GET/api/output/{path} | download GLB / OBJ / PNG files |
GET/api/state · /api/assets | credits, history, private library (CRUD) |
GET/api/health | stack + worker status, no auth |
# 1 · session token curl -X POST https://polymesh.ai/api/auth/login \ -H "Content-Type: application/json" \ -d '{"username":"[email protected]","password":"…"}' # → {"token":"…","user":{…}} # 2 · queue: ComfyUI api-format workflow curl -X POST https://polymesh.ai/api/generate \ -H "Content-Type: application/json" \ -d '{"input":{"workflow_json":{…},"return_outputs_as_base64":true}}' # → {"id":"a91f…"} (queued → running → completed) # 3 · collect the result curl https://polymesh.ai/api/result/a91f… # → {"status":"completed", # "output":[{"type":"base64","mimetype":"image/png",…}]}
Bearer session tokens authenticate the account layer. Long-lived pmk_ API keys (Studio → Profile) are designed for server-to-server use and the decimate service contract. Job queue depth is limited per worker; heavy 3D jobs are serialized for VRAM safety.