H3 Visual Perception Producer:
The Visual Perception Producer creates Qwen-based visual observations from decoded images and constant-frame-rate video within an H3 Context workflow. Its purpose is to provide visual evidence that can help later context-processing stages interpret reference material, rather than generate or edit images. The documented visual profile is qwen38_27b_q4_visual_v2, served through a host-configured Ollama integration. The supported media path uses decoded CPU media, meaning image data is available in system memory for analysis. For reference-conditioning workflows, analysis can be restricted to the frames retained by conditioning, helping keep visual observations aligned with the reference material actually used. Processing depends on the host’s perception-service configuration, and transferring media to the local service requires your consent.
H3 Visual Perception Producer Input Parameters:
media
Required connection of type H3_MEDIA_PRODUCER_RESULT supplying the media producer result to be analyzed. The node description identifies decoded CPU images and constant-frame-rate video as supported visual material; constant-frame-rate video has frames spaced at regular time intervals. This is a structured H3 media connection, not a direct ComfyUI IMAGE connection or a filename field. The supplied media determines the visual material available for observation, so connect the appropriate upstream H3 media producer result. Numeric limits, a default value, and the complete supported-media validation rules are not provided in the supplied source.
route
Required COMBO selection controlling the service route used for visual perception. The node explicitly instructs you to choose ollama for the documented Qwen visual configuration. This selection must agree with the perception service configured by your host; selecting a route does not by itself install or configure that service. The full list of dropdown options and the default selection are not exposed in the supplied source, so only ollama can be documented as the recommended route.
profile_id
Required STRING identifying the perception profile. The documented value is qwen38_27b_q4_visual_v2, which selects the host-configured Qwen visual-observation profile. Use this identifier exactly rather than substituting a descriptive model name: the provided host code checks the profile identifier against its expected visual profile and can reject a mismatch with invalid_request. The source does not expose a default string, length limits, or additional supported visual profile identifiers.
device
Required COMBO selection for the execution device. The node description specifies a decoded CPU media path, but the supplied class does not show the available device choices, the default, or how this setting affects the external Ollama service. Choose the device setting supported by your host’s visual-perception configuration, and ensure the incoming media meets the decoded CPU requirement. Do not assume that this field directly controls GPU acceleration inside Ollama; that behavior is not established by the provided code.
cancel_requested
Required BOOLEAN indicating whether cancellation is being requested. Its possible values are true and false; use false when you want the node to proceed normally, and true to signal a cancellation request. The supplied source does not expose its default value or the underlying cancellation implementation, so the exact stopping point, cancellation response, and handling of work already sent to the service cannot be specified.
local_service_consent
Optional BOOLEAN recording consent for local-service transfer. The node description explicitly states that local transfer requires consent, so enable this setting when you authorize the workflow to send the relevant media to the configured local perception service. This is a permission setting rather than a visual-quality control. Its possible values are true and false, but the supplied source does not expose its default or the exact behavior when the optional input is omitted. Do not assume that leaving it disconnected grants permission.
request
Optional connection of type H3_CONTEXT_REQUEST used to align frame analysis with reference conditioning. The node description states that connecting it restricts sampling to the frames kept by reference conditioning, helping avoid observations based on frames that conditioning did not retain. The provided visual host code also validates the request’s exact type and graph structure, rejecting incompatible requests with invalid_request. Supply the corresponding H3 context request from your workflow rather than unrelated request data. No default, numeric limits, or detailed sampling behavior without this connection is shown.
H3 Visual Perception Producer Output Parameters:
result
Output of type H3_VISUAL_PRODUCER_RESULT containing the visual producer result for downstream H3 processing. Its role is to carry the outcome of visual perception through a dedicated, typed connection rather than return a generated image or a plain-text caption socket. Connect it to a downstream input that accepts this exact visual producer result type. The supplied snippets do not reveal the result’s internal fields, observation format, status representation, or error payload, so individual captions, confidence values, timestamps, or metadata fields cannot be documented as guaranteed contents. Treat the observations as model-produced interpretations of the analyzed material, not as independently verified facts.
H3 Visual Perception Producer Usage Tips:
- Use
ollamaforrouteandqwen38_27b_q4_visual_v2forprofile_id. These are the configuration choices explicitly documented for this visual node. - Confirm that your host has configured the required perception profile and Ollama integration before running the workflow. Selecting the profile in the node is not evidence that the service is installed or available.
- Supply decoded CPU image media or constant-frame-rate video through a compatible H3 media producer. Do not connect a standard
IMAGEsocket directly tomedia, because the declared socket types differ. - Connect the matching
requestwhen working with reference conditioning. This restricts sampling to retained reference frames and keeps the observations relevant to the conditioning selection. - Enable
local_service_consentonly after checking that you authorize transfer to the configured local service. A local destination still requires consent according to the node description. - Keep
cancel_requestedset tofalsefor normal runs. If a workflow unexpectedly stops or does not analyze media, check this setting before investigating service configuration. - Review observations against your reference material before relying on them for creative decisions. The node produces visual interpretations, and the supplied code does not establish guaranteed factual accuracy.
H3 Visual Perception Producer Common Errors and Solutions:
invalid_request
- Explanation: This is the explicit error string shown in the visual host code. It can occur when the profile identifier does not match the expected visual profile, when the request is not exactly the expected
RawContextRequesttype, or when validation of its graph structure fails. - Solution: Set
profile_idtoqwen38_27b_q4_visual_v2and connect the matchingH3_CONTEXT_REQUESTproduced by your H3 workflow. If the error persists, check that the connected H3 nodes come from a compatible installation and that the request has not been replaced with unrelated data.
Local-service transfer is not authorized
- Explanation: This is a troubleshooting description, not a confirmed error string. The node explicitly requires consent for local transfer, so missing authorization can prevent the documented service path from proceeding.
- Solution: If you approve the transfer, explicitly set
local_service_consenttotrue. If you do not approve it, do not run a configuration that requires sending the media to that service.
Visual perception service or profile is unavailable
- Explanation: This is a troubleshooting description; the supplied code does not provide the corresponding runtime message. The node relies on host-configured Qwen observations, so an unavailable Ollama service or an unconfigured visual profile can prevent successful execution.
- Solution: Verify that the host’s Ollama integration is available and that
qwen38_27b_q4_visual_v2is configured for it. Check the host or ComfyUI logs for the specific service failure rather than assuming the node will install the model automatically.
Incompatible media connection or unsupported media preparation
- Explanation: This is a troubleshooting description, not a quoted exception. The node requires
H3_MEDIA_PRODUCER_RESULTand documents decoded CPU images or constant-frame-rate video; other socket types or media preparations do not match that documented path. - Solution: Connect a compatible H3 media producer result and check that the upstream preparation produces decoded CPU media. For video, use a constant-frame-rate source or a preparation path that provides constant-frame-rate media.
Execution is cancelled or does not continue
- Explanation: This is a troubleshooting description; no exact cancellation message is supplied. A
truevalue oncancel_requestedsignals cancellation, although the supplied code does not show when execution stops or how cancellation is reported. - Solution: Set
cancel_requestedtofalsebefore starting a normal run. If cancellation is still reported, inspect any connected control source and the execution logs to determine where the cancellation request originated.
