Training Studio (/platform/studio) is where you invent multi-angle character views or train reusable looks from photos — image LoRAs for a person, product, or style, and video LoRAs for motion — then pin them in Chat. The top of the page is Invent a character (no photos). Below that, start a train job (Train a character / product / style / motion). Datasets, Jobs, Looks, and Voices are history under Your work. Clone a voice from Files Voices or Studio Voices (?recipe=voice) — not a fifth home job card. Training requires a paid plan. Legacy /assets redirects here.
Want the invent path? Fill name + description and click Generate character views — a new chat mints front / side / back plates into Cast. Want the guided train path? Train a character (/platform/studio?recipe=character, also reachable from chat Train character) is three beats — Prepare (name + photos; drop uploads immediately) → Train → Use character. Product, style, and motion use the same wizard with their own triggers and checklists. Progress is saved per organization and job kind.
1. Datasets
Datasets are history of sets you already started — not the place you collect photos. Start a job card (or ?fileIds= without recipe, which opens Person look). Open a set and Continue this look to bind that dataset into the recipe; the wizard does not create a second set.
- Collect (device upload, generation picker, URLs) lives on Prepare.
- Readiness still shows at a glance — item floor (8 images / 6 clips / 1 voice sample), caption coverage, pending moderation, blocked items. Rescan retries an infra scan failure.
- Every item is moderated before it becomes trainable — adult / photoreal / NSFW character photos are allowed. The scan hard-blocks only sexualized-minor suspicion (or an infra scan failure); age-alone does not block.
- Voice consent is still required for voice sets (Jobs / API) and for publishing a voice asset.
- Items are auto-captioned after the scan. Dataset storage counts against your plan’s asset budget (2 GB free / 20 GB starter / 100 GB pro+).
2. Training jobs
Pick a dataset, a job type, and a preset (character / style / product / motion / narration). Advanced settings (steps, rank, learning rate) are available within safe bounds.
| Type | Trains | Dataset size | Base price |
|---|---|---|---|
| Image LoRA | a subject or style for image generation | 8–400 images | 30 cr |
| Video LoRA | a look or motion for video generation | 6–200 clips | 90 cr |
| WAN dual-expert pair | a paired video adaptation for the WAN engine | 6–200 clips | 135 cr |
| Voice clone | a narration voice | 1–25 samples | 30 cr |
Prices cover the default step budget; extra steps price incrementally and are shown before you start. Failed jobs auto-refund.
While a job is running, the Jobs tab shows live progress (phase + percent) so you can see whether the trainer is still warming up or stepping.
3. After training
The hosted trainer currently emits one epoch. When the job finishes, Continue finalizes that checkpoint as your asset version — Studio does not show a sample-grid epoch picker on this path yet. Unpicked checkpoints expire after 14 days. Jobs can still Render samples without re-training.
4. Compare looks (not a Blind A/B tab)
Platform Studio does not run a blind pair bakeoff (no Vote a/b, no LUT picker, no locked-seed grid). Compare two trained looks by Use in Chat. ?tab=evaluate is an honesty panel plus a read-only list of past comparison records. Do not treat Evaluate as a fourth default tab.
5. Looks and publishing
Looks holds trained adapters and imported LoRAs — never voice clones (those live on Voices / Files Voices). Import from Civitai / Hugging Face by URL (imports carry a provenance attestation and a license check). Per-file upload ceilings: LoRA / LoRA pair ≤ 2 GB, LUT ≤ 10 MB, font ≤ 5 MB, voice sample ≤ 100 MB. A non-empty license note is required on import and on Discover publish.
Each asset carries a runnable / library-only / pending badge. Assets trained in the Studio are engine-resident and runnable immediately. Imported or uploaded weight files are honest about their status: they’re library-only until their base-model family is hostable on the platform’s engines (krea2, ltx, wan, hunyuan, and ElevenLabs voices — flux/sdxl files have no hosting engine today). The character look recipe is the supported path to an asset that runs right now.
Publishing may still warn if the asset never completed an API comparison record. Platform Studio does not run that bakeoff itself — compare looks in Chat.
From an asset you can:
- Bind it in Chat — Use character (and Use product/style) pins the Character chip on the krea2 Character leaf; Use motion pins Wan. Incompatible leaves (Seedream plain-images, Kling/Grok) tell you to switch engine rather than applying silently.
- Publish it to Discover: provenance goes to moderation; delisting later never breaks workflows that already bound it.
- Cast / style packs — save a named bundle of asset version pins (optional voice).
Uploads never train platform models — your datasets and assets are yours, scoped to your organization.
6. Over the API and MCP
The whole loop is programmable — see API & MCP and the developer reference:
POST $PLATFORM_API_URL/datasets
POST $PLATFORM_API_URL/datasets/{id}/items
POST $PLATFORM_API_URL/training-jobsKEEP worker /api/v1/datasets still exists for Studio Inngest; new integrations should use platform-api. See API & MCP and the developer reference.
When training finishes, Studio Continue finalizes the current checkpoint. MCP agents use the platform tool inventory (not worker OAuth /mcp, which returns 410).