epithre-iris
Image generation and editing with a 3-LoRA style registry.
Capabilities
| Capability | Notes |
|---|---|
| Tier | Image |
| Max input | 2000 chars prompt |
| Max output | 960x960 (gen), 704x704 (edit) |
| Operations | Text-to-image, single-image edit, multi-reference compose |
| Style LoRAs | none (default), dark, anime |
| Reference images | Up to 5 for compositional editing |
When to use
- Marketing thumbnails, hero images
- Concept art / mood boards
- Product-in-context shots (with multi-reference)
- Style transfer
- Internal asset generation
When NOT to use
- Real people likeness (refused for public figures via safety filter)
- Exact logo reproduction (will get plausibly-wrong)
- Text-in-image (signs, captions): the model produces text-like shapes but rarely legible words
- Precise counting (5 cats reliably gets 3-7)
Pricing
- Rp750 per image, flat (generation or edit).
Latency
| Setting | Time |
|---|---|
| 4 steps @ 768x768 (preview) | ~12-19s |
| 20 steps @ 768x768 (production) | ~45-60s |
| 30 steps @ 960x960 (max quality) | ~90-120s |
| Anime LoRA | +5s |
| Multi-reference edit (3 refs) | +10-15s |
Parameters
prompt: 2000 char max. English works best, Indonesian also supported.size: WxH, max 960x960 gen / 704x704 edit. Rounds down to multiple of 16.num_steps: 1-50, default 4. Use 20-30 for production.seed: -1 random, integer for reproducibility.guidance_scale: 1.0 default. Higher = more literal.lora: none / dark / anime.lora_strength: 0-1.5, default 0.6.
Safety
- CSAM detector (regex + LLM classifier)
- Deepfake-of-real-person filter for named public figures
- NSFW keyword filter
- Layered safety in
_generate_image
Refused requests return HTTP 400 with code: "content_policy_violation".