Task Types

Task types define the input/output contract for jobs and agent nodes. Each model supports one or more task types — see the Models reference for which models support which tasks.

Overview

Task Type	Category	Sync	Description
openai/chat-completion	Text	Yes	Chat completion (LLM)
openai/chat-completion/vision	Text	Yes	Chat completion with image input
openai/chat-completion/ocr	OCR	Yes	OCR via vision model
openai/embeddings	Text	Yes	Text embeddings
openai/rerank	Text	Yes	Document reranking
openai/score	Text	Yes	Text pair similarity scoring
openai/audio-speech	Audio	Yes	Text-to-speech
openai/audio-transcription	Audio	Yes	Speech-to-text
openai/image-generation	Image	Yes	Image generation (OpenAI format)
fal/text-to-image	Image	Yes	Image generation (Fal format)
fal/image-edit	Image	Yes	Image editing / inpainting
fal/text-to-video	Video	No	Text-to-video generation
fal/image-to-video	Video	No	Image-to-video animation
fal/speech-to-video	Video	No	Speech-driven video (talking head)
fal/audio-transcription	Audio	Yes	Speech-to-text (Fal format)
fal/video-interpolate	Video	No	Frame interpolation (slow motion) — coming soon
fal/video-upscale	Video	No	Video super-resolution — coming soon
fal/ocr	OCR	Yes	Document OCR (Fal format)
openai/chat-completion/moderation	Moderation	Yes	Content moderation via chat model

Sync = can be dispatched synchronously (result returned inline). Async tasks return a request_id for polling.

Text Tasks

openai/chat-completion

Standard chat completion following the OpenAI API format.

Field	Type	Required	Description
messages	array	Yes	Array of {role, content} message objects
model	string	No	Model ID (set automatically when using a specific endpoint)
temperature	number	No	Sampling temperature (0-2)
top_p	number	No	Nucleus sampling
max_tokens	number	No	Maximum tokens to generate
stream	boolean	No	Enable streaming response
stop	string/array	No	Stop sequences
tools	array	No	Tool/function definitions

Output: choices[0].message.content (text)

Models: qwen3-0.6b-fp8, qwen3.5-4b, qwen3.5-9b, gpt-oss-20b, deepseek-ocr-v1, deepseek-ocr-v2

openai/chat-completion/vision

Chat completion with image input. Same parameters as openai/chat-completion, but messages can include image content:

{
  "messages": [{
    "role": "user",
    "content": [
      {"type": "text", "text": "What's in this image?"},
      {"type": "image_url", "image_url": {"url": "https://..."}}
    ]
  }]
}

Models: qwen3.5-4b, qwen3.5-9b

openai/chat-completion/ocr

OCR via a vision-capable model. Uses the same message format as vision, optimized for text extraction from images.

Models: deepseek-ocr-v1, deepseek-ocr-v2

openai/embeddings

Generate vector embeddings for text.

Field	Type	Required	Description
input	string/array	Yes	Text or array of texts to embed
encoding_format	string	No	float or base64
dimensions	number	No	Output dimensions

Output: data[0].embedding (float array)

Models: sglang-qwen3-0.6b-fp8-embed, sglang-gpt-oss-20b-embed

openai/rerank

Rerank documents by relevance to a query.

Field	Type	Required	Description
model	string	Yes	Model ID
query	string	Yes	Search query
documents	array	Yes	Array of strings or {text} objects
top_n	number	No	Number of top results to return
return_documents	boolean	No	Include document text in results

Output: Ranked documents with relevance scores

Models: qwen3-0.6b-fp8-score, gpt-oss-20b-score

openai/score

Compute similarity scores between text pairs.

Field	Type	Required	Description
model	string	Yes	Model ID
text_1	string/array	Yes	First text(s)
text_2	string/array	Yes	Second text(s)

Output: Similarity scores

Models: qwen3-0.6b-fp8-score, gpt-oss-20b-score

Audio Tasks

openai/audio-speech

Convert text to spoken audio.

Field	Type	Required	Description
input	string	Yes	Text to speak
voice	string	No	Voice ID (model-specific)
response_format	string	No	mp3, opus, aac, flac, wav, pcm
speed	number	No	Playback speed multiplier

Output: audio_url (audio file URL)

Models: qwen3-tts (9 voices), fox-tts (150+ voices)

openai/audio-transcription

Transcribe audio to text.

Field	Type	Required	Description
audio_url	string	Yes	URL to the audio file
language	string	No	Language code (ISO 639-1)
task	string	No	transcribe or translate

Also supports multipart file upload with a file field.

Output: text (transcribed text)

Models: whisper-large-v3-turbo

fal/audio-transcription

Same as openai/audio-transcription but using the Fal request format.

Field	Type	Required	Description
audio_url	string	Yes	URL to the audio file
language	string	No	Language code
task	string	No	transcribe or translate

Output: text

Models: whisper-large-v3-turbo

Image Tasks

openai/image-generation

Generate images using the OpenAI-compatible format.

Field	Type	Required	Description
prompt	string	Yes	Image description
n	number	No	Number of images
size	string	No	Image dimensions
quality	string	No	standard or hd
response_format	string	No	url or b64_json

Output: data[0].url (image URL)

Models: qwen-image-2512

fal/text-to-image

Generate images using the Fal format. More parameters than the OpenAI format.

Field	Type	Required	Description
prompt	string	Yes	Image description
negative_prompt	string	No	What to avoid
image_size	string	No	square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9
num_inference_steps	number	No	Denoising steps
guidance_scale	number	No	Prompt adherence strength
num_images	number	No	Number of images (max 10)
seed	number	No	Reproducibility seed
loras	array	No	LoRA configs [{path, scale}] (scale 0-4)
enable_safety_checker	boolean	No	Content safety filter
output_format	string	No	Output image format

Output: images[0].url (image URL)

Models: nunchaku-flux1-schnell, sglang-diffusion-flux2-klein-4b, qwen-image-2512, sglang-diffusion-qwen-image-2512-fp8

fal/image-edit

Edit or inpaint images.

Field	Type	Required	Description
prompt	string	Yes	Edit instruction
image_url	string	Yes	Source image URL
mask_url	string	No	Mask for inpainting
strength	number	No	Edit strength (0-1)
num_inference_steps	number	No	Denoising steps
guidance_scale	number	No	Prompt adherence
num_images	number	No	Number of results (max 10)
seed	number	No	Reproducibility seed
loras	array	No	LoRA configs

Output: images[0].url (image URL)

Models: nunchaku-flux1-schnell, sglang-diffusion-flux2-klein-4b, qwen-image-edit-2511, sglang-diffusion-qwen-image-edit-2511-fp8

Video Tasks

All video tasks are async only — they return a request_id for polling.

fal/text-to-video

Generate video from a text prompt.

Field	Type	Required	Description
prompt	string	Yes	Video description
resolution	string	No	480p, 720p
aspect_ratio	string	No	16:9, 9:16, 1:1
num_inference_steps	number	No	Denoising steps
guidance_scale	number	No	Prompt adherence
num_frames	number	No	Number of frames
fps	number	No	Frames per second
seed	number	No	Reproducibility seed
output_format	string	No	Video format

Output: video.url (video URL)

Models: wan22-ti2v, sglang-diffusion-wan22-t2v-a14b-fp8, ltx2-distilled

fal/image-to-video

Animate a still image into video.

Field	Type	Required	Description
prompt	string	Yes	Motion description
image_url	string	Yes	Source image URL
resolution	string	No	480p, 720p
aspect_ratio	string	No	16:9, 9:16, 1:1
num_inference_steps	number	No	Denoising steps
guidance_scale	number	No	Prompt adherence
num_frames	number	No	Number of frames
fps	number	No	Frames per second
seed	number	No	Reproducibility seed

Output: video.url (video URL)

Models: wan22-ti2v, ltx2-distilled

fal/speech-to-video

Generate talking-head video driven by audio.

Field	Type	Required	Description
audio_url	string	Yes	Audio file URL
image_url	string	Yes	Face/character image URL
prompt	string	No	Additional scene description
num_inference_steps	number	No	Denoising steps
guidance_scale	number	No	Prompt adherence
seed	number	No	Reproducibility seed

Output: video.url (video URL)

Models: wan22-s2v

fal/video-interpolate (coming soon)

Increase video frame rate (slow motion effect). No model is currently enabled for this task.

fal/video-upscale (coming soon)

Upscale video resolution. This task is currently disabled while the model is being stabilized.

Agent Usage

Task types are used as the task field in agent DAG nodes. Each node specifies a task type and a model, and can wire outputs from upstream nodes into its inputs.

{
  "nodes": [
    {
      "id": "generate",
      "task": "fal/text-to-image",
      "model": "nunchaku-flux1-schnell",
      "payload": {
        "prompt": "{{input.prompt}}"
      }
    }
  ]
}

See the Agents guide for details on building multi-step pipelines.