Gemma 4 Captioner (v0.1)

About this version

About this app (See raw metadata.json)

Applies a Gemma 4 vision-language model, run with llama.cpp from a GGUF file, to video frames selected by input TimeFrame annotations for prompt-driven captioning / scene description (experimental llama.cpp backend). Captioning is composite: the frames sampled from a TF are fed to the model in a single prompt and yield one caption. When maxImagesPerPrompt is set, a TF with more frames than the cap is split into several prompts and yields one caption per prompt, all aligned to that TF. A model runtime parameter selects the Gemma 4 size and quantization (default ggml-org/gemma-4-E4B-it-GGUF:Q8_0); see metadata.py for the GPU memory each needs. The app runs its own llama-server process and needs the llama-server executable (included in the container image).

Inputs

(Note: “*” as a property value means that the property is required but can be any value.)

Configurable Parameters

(Note: Multivalued means the parameter can have one or more values.)

  • tfLabels: optional, defaults to []

    • Type: string
    • Multivalued: True

    Label(s) of input TimeFrame annotations to caption. By default ([]), all TimeFrames are processed regardless of label. When set, TimeFrames without a label are not processed.

  • tfLabelIgnore: optional, defaults to []

    • Type: string
    • Multivalued: True

    Label(s) of input TimeFrame annotations to leave out, applied after tfLabels. TimeFrames without a label are never dropped by this list.

  • prompt: optional, defaults to These frames are sampled in order from one segment of a video. Describe what is shown and quote any visible text exactly. Describe only what the frames show; do not guess the date or era.

    • Type: string
    • Multivalued: True

    User prompt(s) sent to the model. A single value runs as a one-shot generation. A multi-value list is interpreted as a multi-turn static prompt; see promptMode for how turns are assembled.

  • systemPrompt: optional, defaults to Answer in one plain-text paragraph. Do not use Markdown, headings, lists, or a separate summary.

    • Type: string
    • Multivalued: False

    Optional system-role text prepended to the conversation. Empty by default.

  • promptMode: optional, defaults to turn-taking

    • Type: string
    • Multivalued: False
    • Choices: user-only, turn-taking

    How to interpret a multi-value prompt list. Has no effect when prompt has a single value. For semantics of each choice and worked examples, see https://clams.ai/clams-python/app-baseclasses.html#promptable-multiturn

  • maxNewTokens: optional, defaults to 512

    • Type: integer
    • Multivalued: False

    Maximum number of new tokens generated per inference call. Forwarded to the backend’s generate-equivalent. Larger values grow the KV cache linearly and increase GPU memory usage; reduce if VRAM is constrained.

  • useReasoning: optional, defaults to false

    • Type: boolean
    • Multivalued: False
    • Choices: false, true

    Request the model’s reasoning (“thinking”) mode. Off by default. Honored only by apps whose backing model has a distinct reasoning mode; apps without one ignore it. When honored and enabled, the reasoning trace is split from the answer and stored in the modelReasoningTrace property of the output TextDocument (kept out of the document text). Reasoning is markedly slower and far more token-hungry: the whole trace is generated before the answer and drawn from the same budget capped by maxNewTokens, so raise maxNewTokens substantially (thousands of tokens, not hundreds) when enabling this, or the trace may consume the entire budget and the answer be truncated or empty. Small reasoning models (as a rule of thumb, roughly 4B parameters and under) are especially prone to non-terminating “thinking loops” that exhaust the budget without producing an answer; validate termination per model before relying on it.

  • temperature: optional, defaults to 0

    • Type: number
    • Multivalued: False

    Sampling temperature. The default 0.0 selects deterministic / greedy decoding for maximum reproducibility; override for sampled generation.

  • topP: optional, defaults to 1

    • Type: number
    • Multivalued: False

    Nucleus-sampling cumulative probability cutoff. Only meaningful when temperature is greater than 0.

  • topK: optional, defaults to 50

    • Type: integer
    • Multivalued: False

    Top-K sampling cutoff. Only meaningful when temperature is greater than 0.

  • maxImagesPerPrompt: optional, defaults to 0

    • Type: integer
    • Multivalued: False

    Maximum number of images the app bundles into a single prompt (one generation, one output TextDocument). 0 means no limit, so all images for a TimeFrame go into one prompt and yield one TextDocument. With a positive N, the app splits a TimeFrame whose frames exceed N into consecutive groups of at most N; each group produces its own TextDocument aligned to the same TimeFrame and grounded to its own source TimePoints via origins. This bounds GPU memory for long dynamic scenes. The value is operator-set, so the number of output TextDocuments is deterministic and does not vary by GPU. Composes with the number of prompts the backend runs together (parallelPrompts in apps that offer it).

  • imagesPerPromptMode: optional, defaults to max

    • Type: string
    • Multivalued: False
    • Choices: max, balanced

    How the images of a TimeFrame that exceeds maxImagesPerPrompt are spread over its prompts. Has no effect when maxImagesPerPrompt is 0. Both choices produce the same number of prompts. “max” fills each prompt to the cap and puts the remainder in the last one (33 images at a cap of 32 give 32 and 1). “balanced” spreads the images evenly, so prompt sizes differ by at most one (33 images at a cap of 32 give 17 and 16).

  • model: optional, defaults to ggml-org/gemma-4-E4B-it-GGUF:Q8_0

    • Type: string
    • Multivalued: False
    • Choices: ggml-org/gemma-4-E2B-it-GGUF:Q8_0, ggml-org/gemma-4-E2B-it-GGUF:Q4_0, ggml-org/gemma-4-E2B-it-GGUF:BF16, ggml-org/gemma-4-E4B-it-GGUF:Q8_0, ggml-org/gemma-4-E4B-it-GGUF:Q4_0, ggml-org/gemma-4-E4B-it-GGUF:BF16, ggml-org/gemma-4-12B-it-GGUF:Q8_0, ggml-org/gemma-4-12B-it-GGUF:Q4_0, ggml-org/gemma-4-12B-it-GGUF:BF16, ggml-org/gemma-4-26B-A4B-it-GGUF:Q8_0, ggml-org/gemma-4-26B-A4B-it-GGUF:Q4_0, ggml-org/gemma-4-26B-A4B-it-GGUF:BF16, ggml-org/gemma-4-31B-it-GGUF:Q8_0, ggml-org/gemma-4-31B-it-GGUF:Q4_0, ggml-org/gemma-4-31B-it-GGUF:BF16

    Model to use for this request. Must be one of the model ids declared in this app’s analyzer_versions; the SDK pins the corresponding commit hash at load time. A model id is a Hugging Face repository, optionally followed by :<quantization> (e.g. org/repo-GGUF:Q8_0). When the app ships a single model, this parameter defaults to it and can be omitted.

  • pretty: optional, defaults to false

    • Type: boolean
    • Multivalued: False
    • Choices: false, true

    The JSON body of the HTTP response will be re-formatted with 2-space indentation

  • runningTime: optional, defaults to true

    • Type: boolean
    • Multivalued: False
    • Choices: false, true

    The running time of the app will be recorded in the view metadata

  • hwFetch: optional, defaults to false

    • Type: boolean
    • Multivalued: False
    • Choices: false, true

    The hardware information (architecture, GPU and vRAM) will be recorded in the view metadata

  • tfSamplingMode: optional, defaults to representatives

    • Type: string
    • Multivalued: False
    • Choices: representatives, single, all

    Sampling mode for TimeFrame annotations. Has no effect when the app does not process TimeFrames. “representatives” uses all representative timepoints if present, otherwise skips the TimeFrame. “single” uses the middle representative if present, otherwise extracts an image from the midpoint of the start/end interval (midpoint is calculated by floor division of the sum of start and end). “all” uses all target timepoints if present, otherwise extracts all images from the time interval.

Outputs

(Note: “*” as a property value means that the property is required but can be any value.)

(Note: Not all output annotations are always generated.)

  • http://clams.ai/vocabulary/type/TextDocument/v2
    • origins = “*”
    • origination = “derived”

    Caption text generated by the Gemma 4 model for each prompt: one per processed TimeFrame, or several when maxImagesPerPrompt splits a TimeFrame. The origins property points to the TimePoint(s) anchoring the image(s) that caption was generated from.

  • http://clams.ai/vocabulary/type/Alignment/v1 (of any properties)

    Alignment between each parent TimeFrame and the TextDocument(s) derived from it.

  • http://clams.ai/vocabulary/type/TimePoint/v5
    • timeUnit = “milliseconds”
    • timePoint = “*”

    Optional output. Newly-created TimePoint annotations for images sampled from a TimeFrame interval without an existing backing TimePoint (see tfSamplingMode).