Model catalog

Every generation lane, the models in it, their tunable parameters, and whether they run today. Settings are the quality-tuned values from the platform's R&D whitepapers, not library defaults.

The /v1/models endpoint LIVE

The whole catalog below is also served as a live, machine-readable API, so your code can discover models, their capabilities, and their cost at runtime instead of hard-coding a list. It is public - no API key required.

GET https://flopco.in/v1/models

Query parameters

All optional. Absent means the whole catalog as JSON.

Response envelope

{
  "ok": true,
  "count": 55,
  "types": ["text", "image", "video", "audio"],
  "families": ["edit", "pixel", "vector", "generator"],
  "formats": ["json", "xml", "php"],
  "models": [ ...model objects... ]
}

A model object

{
  "id": "flux-kontext",
  "name": "FLUX.1 Kontext dev",
  "type": "image",
  "family": "edit",
  "familyLabel": "Instruction edit",
  "familyNote": "Edits an image you supply. Requires init_image; a text-only prompt cannot produce an image.",
  "capabilities": ["instruction-edit", "image-to-image"],
  "acceptsInitImage": true,
  "provider": "Black Forest Labs",
  "license": "FLUX.1 Non-Commercial",
  "cost": 9,
  "costUnit": "FLOP",
  "spec": "instruction edit, needs init_image",
  "status": "preview",
  "tags": ["edit", "gated"],
  "docs": "https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev"
}

Examples

# the whole catalog (JSON)
curl https://flopco.in/v1/models

# just the video models
curl "https://flopco.in/v1/models?type=video"

# text models that can see images
curl "https://flopco.in/v1/models?type=text&tag=vision"

# the audio lane as XML
curl "https://flopco.in/v1/models?type=audio&format=xml"

# PHP-serialized (unserialize() straight into an array)
curl "https://flopco.in/v1/models?format=php"

XML shape

<?xml version="1.0" encoding="UTF-8"?>
<response>
  <ok>true</ok>
  <count>2</count>
  <models>
    <item>
      <id>wan-2.2-i2v</id>
      <type>video</type>
      <capabilities><item>image-to-video</item></capabilities>
      ...
    </item>
  </models>
</response>

Text lane LIVE

Prompt to text through an ollama-compatible endpoint. Fulfilled inline by the worker. One config value (gen_text_endpoint) points it at the local VPS ollama or a RunPod serverless ollama with no code change.

ollama (default llama3.2:3b) LIVE

Image lane PLANNED

Prompt to image. The image lane holds four very different KINDS of model, and the /v1/models response tags each one with a family so a picker can group and warn instead of lumping them together. Pick the right family for the job:

All image models below are currently preview in the catalog (proven, worker wiring pending); none serve end-to-end through /v1/generate yet. Ids are authoritative - they match GET /v1/models?type=image exactly.

Generator - full-resolution text-to-image PLANNED

Real text-to-image. Self-hosted diffusion on RunPod GPUs plus a couple of hosted-API options. This is the lane for product shots, photoreal, painterly, and fantasy work at 1024px+.

Diffusion tuning (the RunPod generators), tuned values from the platform whitepapers:

Pixel art - low-res retro sprites (by design) PLANNED

Instruction edit - image-to-image (needs init_image) PLANNED

Vector / SVG - text-to-vector PLANNED

Prompt to a scalable SVG rather than a raster bitmap. First-party pipeline: prompt goes to a catalog image model, then a tracer (vtracer / potrace / pngtosvg) converts the raster to vector.

Transparent (RGBA) image output is a self-hosted target on a torch>=2.6 ComfyUI base; proprietary transparent generation runs via Azure gpt-image (never OpenAI direct). See the gpt-image provider page and the open models page.

Video lane PLANNED

Async GPU renders (submit to a serverless provider, store external_job_id, poll to adopt). No provider is configured in production yet, so video jobs stay queued. Model ids: wan-alpha, wan22-i2v.

Premium proprietary video (Runway, Veo) is a separate provider lane; see Runway.

Audio - TTS & voice clone PLANNED

Chatterbox Multilingual (ship pick, MIT) PLANNED

Commercial-safe zero-shot voice clone from a short user sample. Prototype flows are proven on F5; weights swap to Chatterbox before charging.

F5-TTS (prototype only, non-commercial weights) PLANNED

XTTS-v2 (benchmark only, non-commercial weights) PLANNED

Reference-sample spec (record.flopco.in, planned): 10-15 s of clean speech (sweet spot ~12 s) reading a fixed sentence, 44.1 / 48 kHz mono WAV; server normalizes, trims to <=12 s.

Audio - music & song-from-lyrics PLANNED

ACE-Step 1.5 (default engine) PLANNED

One engine: instrumental if no lyrics, sung if lyrics. Do NOT use the turbo checkpoints (quality-lossy); use base or sft.

YuE (premium A100 lyrics-to-song) PLANNED

Minutes-per-song, A100 80GB target. Premium/batch tier, not an interactive default.

Audio - text-to-audio / SFX PLANNED

Stable Audio Open 1.0 PLANNED

Stability Community License (free under $1M rev; must display "Powered by Stability AI" and register commercial use). A "Small" variant serves a fast SFX lane.

Output: 44.1 kHz stereo WAV, base64 in the provider payload. bf16 on A40/A100.

Sprite / pixel-art PLANNED

SDXL + nerijs/pixel-art-xl LoRA, then a kCentroid downscale + palette lock (the crisp-pixel secret). Sprite SHEETS keep one shared locked palette + grid across all frames for consistency.

Sheet/animation pipeline: one anchor sprite -> WAN 2.2 i2v -> BiRefNet bg-strip per frame -> kCentroid to the shared grid -> shared-palette quantize -> stitch to a horizontal strip + atlas JSON.

Talking-head / lip-sync PLANNED

Pipeline (all commercial-safe): SadTalker (512) -> MuseTalk v1.5 lip-resync -> GFPGAN v1.4 -> Real-ESRGAN x2 -> mp4. If a face VIDEO exists, MuseTalk direct is highest fidelity. Wav2Lip weights are non-commercial - do not ship.

LivePortrait (MIT) is an optional premium re-driver for smoother head motion; it needs a driving video (no native audio input).

Tool ops PLANNED

Non-model utilities exposed as API-callable jobs with the same media-in / media-out contract. Mostly ffmpeg + a couple of vision models.

Related video tools (chained i2v "video-long", ffmpeg chroma-key "video-chroma", alpha / solid-color stripping) follow the same media-in / media-out job contract.