@ohmygame/fal
v0.4.0
Published
fal.ai's image, video, audio and 3D models, with your own API key.
Readme
fal.ai
Adds fal.ai to OhMyGame's Settings → Providers, with your own API key, for pictures, videos, audio and 3D models. It has no chat models.
- API key: From fal.ai → Dashboard → Keys.
| Kind | Model | What it takes |
| --- | --- | --- |
| Image | FLUX.1 [schnell] (fal-ai/flux/schnell) | A prompt, in 1:1, 4:3, 3:4, 16:9 or 9:16 |
| Background removal | BiRefNet (fal-ai/birefnet) | One still PNG, JPEG or WebP; preserves dimensions and colors |
| Video | Kling O3 (fal-ai/kling-video/o3/standard/image-to-video) | A first frame, an optional last frame, and 3, 5, 10 or 15 seconds |
| Speech | MiniMax Speech-02 HD (fal-ai/minimax/speech-02-hd) | Up to 5,000 characters, in one of MiniMax's voices, as MP3 |
| Sound effects, music | Stable Audio (fal-ai/stable-audio) | A description and up to 47 seconds, as WAV |
| Sound effects, music | Stable Audio 2.5 (stable-audio-25) | A description and 1–190 whole seconds, as WAV |
| 3D | TRELLIS (fal-ai/trellis) | One picture of the object, as GLB |
Generation models are on from the start. Background removal has its own processing picker and does not change the default image model. To use more of fal.ai's models, open the fal.ai provider in Settings → Providers and choose Load models under a kind: the plugin reads fal.ai's most used models of that kind and what each one takes (aspect ratios, lengths, voices, how many pictures), and you turn on the ones you want. A model that asks for something OhMyGame doesn't send, such as a mask or a sound clip, isn't listed.
Every request goes through fal.ai's queue, in
media/adapter.js. OhMyGame runs that script in a sandbox:
it reaches only queue.fal.run, api.fal.ai (to list models) and the hosts
fal.ai's files are on, and OhMyGame adds the key to its requests. A video,
audio file or 3D model a restart cut off is checked again on Retry, not
paid for twice. fal.ai bills each request to your account.
Needs OhMyGame with plugin API 2.27.0 or later. Version 0.4.0 remains Beta.
Published as @ohmygame/fal from ohmygame-plugins.
Stable Audio 2.5
Choose Stable Audio 2.5 on an audio node for music or sound effects. It uses the existing fal key and permissions, and the official text-to-audio queue. Set a whole number of seconds from 1 to 190. If no duration is supplied, the vendor's 190-second default is used. Output is WAV; this endpoint has no MP3, reference-audio editing, dedicated loop or instrumental-only control. Describe the desired sound in the prompt; a prompt is not a guarantee of seamless loops or absence of vocals.
The original 47-second Stable Audio preset and all existing selections remain.
The new preset uses a separate local ID so it cannot replace a previously
enabled, saved fal-ai/stable-audio-25/text-to-audio discovery definition or its
pending task. Newly discovered definitions for that endpoint correctly list WAV.
If you previously enabled it with the older MP3 display, finish its pending jobs
first, then use the new preset; do not change a pending job's model or format.
Merely loading models does not rewrite your enabled definitions.
New 2.5 jobs retain queue history (sync_mode: false). Stop stops local
waiting without withdrawing the remote job; Retry follows the same result.
Restarting the app or retrying a failed download also keeps its saved task.
A lost submission response is reported as uncertain and is never automatically
submitted again. If fal no longer has the task or output, check its history before
explicitly starting new work. This does not extend fal's retention period.
Tasks created by an older plugin retain their original cancellation behavior.
Local audio acceptance
- Upgrade and confirm the key, switch, old Stable Audio selection and hidden models remain. Select Stable Audio 2.5 explicitly and generate a short sound effect and music clip; check WAV playback and requested duration. A long clip may incur more cost; choose 190 seconds only when you want one.
- Stop a queued 2.5 job, then Retry; also quit/reopen and Retry. Confirm the same request in fal history and no second submission. Repeat after a failed output download, keeping the original node/model.
- Check rejection of MP3 and fractional/out-of-range duration before submission. Load audio models and confirm newly enabled 2.5 discovery reports WAV.
Tests use the real OhMyGame host, sandbox and task journal with simulated queue responses and a synthetic WAV. Real generation, sound quality and duration need local acceptance; the fixture's short silence is not a 190-second vendor result.
Background removal
On an image's canvas toolbar, choose Image tools → Remove background (中文:图片工具 → 去除背景), choose BiRefNet, then generate. Creation alone does not send a paid request. The existing AI image cutout remains a prompted redraw using an image generation model.
The plugin uses the official BiRefNet queue API,
General Use (Light), its default 1024×1024 operating resolution, PNG output,
sync_mode: false to retain request history and refine_foreground: false.
Operating resolution is internal to the model; the returned alpha must match the
source dimensions. OhMyGame applies that alpha to the normalized source RGB and
existing alpha, rejects resized/opaque/empty outputs and never substitutes a new
paid request after an uncertain failure. The source is limited to a still
PNG/JPEG/WebP up to 20 MiB and 64 million pixels; EXIF orientation and sRGB
normalization happen before submission. Edge quality still needs inspection on
your own artwork, especially hair, small sprites and existing transparency.
Closing the app or stopping locally keeps the queued task; Retry resumes it. A failed result download or import keeps its receipt too. Do not create a second node to recover an uncertain submission. Inputs are sent to fal.ai using the existing key and approved hosts; no additional permissions or callback setup are needed. An output on a host outside the existing allowlist fails explicitly; report the hostname instead of granting a wildcard.
Local acceptance
- Upgrade with the current app; confirm the existing key, provider switch, image default and model visibility are retained.
- Process a non-square JPG, a detailed PNG and a WebP; inspect foreground edges, dimensions, source colors and transparent PNG export. Try an already transparent source and confirm its existing alpha remains respected.
- Stop a queued cutout, then Retry; also quit/reopen during a queued job and Retry. Confirm in fal.ai request history that the same request is resumed.
- Disconnect before downloading the result, reconnect and Retry; no second submission should appear. Check explicit failure for invalid/opaque results.
- Confirm ordinary image generation and existing video/audio/3D selections still work. BiRefNet must not appear among image generation defaults.
