pryx-comfyui-higgsfield Introduction
pryx-comfyui-higgsfield brings Higgsfield image and video tools into ComfyUI. You can use familiar, connected ComfyUI nodes to create images, animate still pictures, make videos from prompts and references, edit footage, and extend clips.
The extension is designed to make a changing model catalog easier to work with. When you select a model, its node shows the controls that model supports, such as duration, resolution, or reference inputs. Model information and tooltips help you understand the available choices. Compatible model definitions can be discovered from Higgsfield’s API without waiting for a new extension release.
Generation may cost money. Generator nodes start in estimate_only mode. This lets you review Higgsfield’s estimated price before choosing to submit a generation request.
The ComfyUI backend communicates directly with api.higgsfield.ai. According to the extension documentation, the author does not operate a proxy or receive your API credentials or generated media.
How pryx-comfyui-higgsfield Works
Think of a ComfyUI workflow as a set of connected steps. A Higgsfield generator node takes your prompt, settings, and optional media, then prepares a request for the model you selected.
A typical workflow looks like this:
- Choose a task and model. For example, use Image to Video to animate a still image, or Video Edit to change an existing clip.
- Connect your creative inputs. Add a prompt, a source image or video where required, and any supported references.
- Review the model’s controls. The node displays relevant settings for the selected model. Its model-information panel explains supported inputs and limits.
- Check the estimate. Queue the workflow with
request_modeset toestimate_only. The node validates the request and asks Higgsfield for a price estimate, but does not submit a paid generation request. - Generate when ready. If the estimate is acceptable, change
request_modetogenerateand queue the workflow again. You can set a positivemax_usdvalue to block submission when the pre-submission USD estimate is too high.
For example, to animate a portrait, connect Load Image to an Image to Video node’s image input, describe the motion in the prompt, and run an estimate. After checking the price and settings, switch to generate. Once the result is ready, connect the node’s video output to a compatible ComfyUI video output or save node.
An estimate may still upload connected local media so Higgsfield can evaluate the complete request. It does not create the final image or video.
pryx-comfyui-higgsfield Features
Model-aware controls and catalog
Each generator has a validated model dropdown. Its controls change to reflect the selected model, so you do not have to guess which settings apply. Expand the model-information panel to check supported media, limits, output options, and reference guidance.
The optional Model Catalog node lets you choose a model once and connect its model_id to one or more generators. You can filter entries by provider, task capability, or status. This is useful when several parts of a workflow should use the same model.
Image generation and editing
Image Generate & Edit covers the image endpoints in the loaded catalog. Depending on the model, you can generate an image from a prompt or provide source images for an edit. Available controls vary by model.
Some image workflows have distinct reference roles:
- AI Influencer: Connect an identity photo to
image. Supply clothing or object references throughitem_imagesor a Reference Collector. The prompt is used as a brief, with a 4,000-character limit. Appearance choices depend on the selected tier; advanced overrides are generally best left blank unless you know the identifiers required. - Ads Studio: Accepts up to ten product or reference images and can generate one to eight images. Its token-based final charge may differ from the estimate.
Video generation, editing, and extension
- Text to Video creates a clip from a written prompt. A model may offer settings such as duration, aspect ratio, resolution, audio, or seed.
- Image to Video animates a starting
IMAGE. Models that support a defined ending frame also showend_image. - Reference to Video uses ordered references, which may include images, video, audio, documents, or web links when the model supports them.
- Video Edit changes a connected source video. Its available modes include documented Genjutsu Motion Transfer, Object Swap, Light, and Restyle options, alongside other catalog models.
- Video Extend continues an existing clip, using the duration and output options supported by its selected model.
For Genjutsu Light, Higgsfield determines duration from the source clip, rounds it up, and caps it at 15 seconds; there is no editable duration setting. For Genjutsu Restyle, you choose a preset and provide a video of at least four seconds. Clips longer than 30 seconds use their first 30 seconds, and you can add up to five character images.
Reference Collector and Reference Preview
Use Reference Collector to gather several connected images, videos, or audio inputs into one ordered references output. It supports up to 30 image slots, 10 video slots, and 10 audio slots; the selected generator checks its own, potentially lower, limits. Image batches retain their internal order. You can also provide an already public HTTPS document or webpage URL.
Collector sockets are read in numeric order, regardless of the order in which you drew the wires. Reference Preview shows the resulting order and prompt mapping before you run a generator. It makes no Higgsfield API request and uploads no media.
Direct inputs have specific jobs: image is typically a source or starting image, end_image an ending frame, video a source clip, and audio an audio reference. Use references for additional ordered material. Inputs that a selected model cannot use are hidden, and the request is checked again before upload.
Prompt references depend on the endpoint. For Wan 3.0 Reference to Video, documented tokens include Image 1, Image 2, Video 1, and Audio 1. You can label Collector items, for example image_1=person, and write {{ref:person}} in the prompt. The extension resolves that alias to the appropriate token. Do not assume the same alias syntax works with other endpoints; Reference Preview’s numbering may describe delivery order without implying supported prompt tokens.
SOUL ID character training
SOUL ID creates a reusable character from one to 100 connected images. It is a training workflow, not a regular image-generation request.
Its default validate_only action checks the name, model version, and image count without uploading images or contacting the API. The train action uploads images and submits a training request, which may cost money; regular inference estimate_only and max_usd controls do not apply. Save the returned custom_reference_id. Use check with that ID to read training progress instead of submitting another training request.
Choose v1, v2, or cinema to match the SOUL model that will later use the character. Once training is complete, connect the ID to a compatible model’s custom_reference_id input. Where available, custom_reference_strength changes how strongly the trained character influences the result.
Advanced requests
Advanced Request accepts JSON arguments for specialized workflows. It still uses a model from the validated catalog and cannot change the API host or endpoint. Its regular media inputs appear only when supported by the selected model. If you do not already know that model’s documented request fields, the standard generator nodes are easier to use.
Estimates, limits, and outputs
estimate_only validates your inputs and requests Higgsfield’s estimate without submitting paid generation. generate follows the same estimate-first process and submits generation only after the estimate succeeds.
A positive max_usd blocks generation if the returned USD estimate exceeds your limit or no numeric USD estimate is available. Setting it to 0 disables this local limit. It is not an account-level spending cap or a guarantee of the final bill. If an estimate fails, the extension stops rather than generating without a quote.
Generator outputs can include native ComfyUI images or video, result URLs, a local video file path, a request ID, estimated credits and usd, and diagnostic status text. Media outputs are empty during an estimate-only run. An unavailable numeric estimate is reported as -1, not as a free request.
auto_save controls whether completed media is stored in ComfyUI’s output or temporary directory. timeout limits how long the node waits locally. Progress messages describe the current phase and elapsed time, but this integration does not provide a reliable generation percentage or ETA.
pryx-comfyui-higgsfield Models
The extension’s model choices come from the Higgsfield catalog loaded by your ComfyUI session. The live dropdown is the best place to see what is selectable for your account and task. A model can appear in one task node but not another because its documented endpoints support different capabilities.
The bundled fallback catalog, revision 2026-10-21.2, contains 87 inference endpoints. That count includes separate modes and endpoints, not 87 distinct model families. The original documentation provides examples rather than a complete list of every model ID:
| Task | Bundled examples | Choose it when you want to… |
|---|---|---|
| Image generation | SOUL, Recraft, Ideogram, Z-Image | Create an image from a prompt. |
| Image editing | SOUL 2 Image to Image, AI Influencer, Ads Studio, Marketing Studio, Grok Image | Transform source images or use a specialized image workflow. |
| Text to video | MiniMax H3, Seedance, Kling, Wan, PixVerse | Create a clip from a written scene description. |
| Image to video | MiniMax H3, Seedance, Kling, Wan, PixVerse | Give a still image motion. |
| Reference to video | MiniMax H3, Seedance, Kling Omni, Wan | Guide a clip with supported reference media. |
| Video editing and motion control | Genjutsu, Seedance, Kling variants | Transform footage or guide motion with a supported mode. |
| Video extension | Seedance | Continue a source clip. |
Model settings can affect both the result and its estimate. For example, a longer video or a higher supported resolution may change the request-specific price. In the bundled catalog, Seedance 2.5 offers 480p, 720p, and 1080p across its five listed video modes. Supported Genjutsu Motion Transfer and Object Swap modes also offer 1080p; 720p is the default. Check the selected model’s controls rather than assuming every model offers the same choices.
SOUL ID training is separate from the 87 inference endpoints. The catalog does not include a guessed standalone upscaler endpoint: a tool on Higgsfield’s website is not necessarily available through its API.
What’s New with pryx-comfyui-higgsfield
The supplied documentation describes catalog revision 2026-10-21.2, but does not provide a version-by-version release history. These are the changes and behaviors it identifies:
- Automatic model discovery: With a configured API key, the extension can obtain model modes and request schemas from Higgsfield. Compatible additions can become available without a new extension release.
- Safer catalog updates: A valid local cache and bundled snapshot remain available if discovery fails or returns incompatible information. An update found during a session is cached until you restart ComfyUI and reload the browser.
- Settings visibility: Automatic model updates shows check status, failures, and whether a restart is needed. Check models now is available when you need to bypass the normal cache interval.
- Expanded documented video choices: The bundled snapshot includes Seedance 2.5 resolutions across five video modes and
1080pchoices for supported Genjutsu Motion Transfer and Object Swap modes. Its Genjutsu endpoint paths were corrected in the documented snapshot.
The extension normally refreshes an aged model cache automatically; you do not need to check for models before every workflow. Newly introduced task types, media roles, output types, or API protocols may still require an extension update.
Troubleshooting pryx-comfyui-higgsfield
A model is missing from the dropdown.
- Confirm you are using the node for the model’s task. A model documented for Image to Video will not necessarily appear in Reference to Video.
- Open ComfyUI Settings → PRYX ComfyUI Higgsfield → Automatic model updates and check for errors or a pending restart.
- If an update is pending, restart ComfyUI and hard-reload the browser with
Ctrl+F5. - Use Check models now if you need to check beyond the current cache. A genuinely new API capability may still require an extension update.
The image or video output is empty.
- Check
request_mode. Empty generated media is expected inestimate_only. - Read the
credits,usd, andstatusoutputs before changing togenerate. - Set a suitable positive
max_usdlimit if you want the request blocked above a particular estimated USD amount.
A connected image, clip, or reference is rejected.
- Expand the selected model’s information panel to check accepted media types and counts.
- Make sure source media uses its intended direct input, such as
image,end_image, orvideo. - Run Reference Preview to confirm the order and mapping of additional references.
- Check model-specific requirements, such as source-video duration or a separate limit on item images.
My credentials validate, but generation fails.
- Run
estimate_onlywith the same model, settings, and inputs. - Inspect the node’s
statusoutput and the ComfyUI console for the reported error. - Check provider-side possibilities such as model access, account credits, content restrictions, or media requirements. Credential validation alone does not guarantee that every generation request will succeed.
The estimate returns an HTTP 500 error.
Do not bypass the failed estimate by trying to force generation. The node stops before submission. Retry later and check the sanitized error for the affected endpoint. The original documentation recorded HTTP 500 estimate responses for several Kling edit and motion-control modes during checks on October 21, 2026; those observations do not establish whether every valid request to those modes fails.
The prompt is difficult to edit in a small field.
Connect a text-producing node, such as a Config UI Prompt node, to the generator’s prompt input. It accepts a connected STRING.
Frequently asked questions
- Does an estimate create or charge for a generated image or video?
estimate_onlydoes not submit the paid generation request. Connected local media may still be uploaded to prepare the estimate. - Is
max_usda guaranteed spending cap? No. It checks the provider’s pre-submission USD estimate, not the final bill or your account’s total spending. - Why is a numeric estimate
-1? The provider did not return that numeric value. It means unavailable, not zero cost. - Do I need to host my reference images myself? No. Connected local media is prepared and uploaded through Higgsfield’s upload flow. An
external_url, however, must already be publicly reachable. - Can I use the same prompt-reference labels with every model? No. The documented alias-to-token behavior applies to Wan 3.0 Reference to Video; other endpoints may use different prompt rules.
- How do I check SOUL ID training progress? Keep the returned
custom_reference_idand run the SOUL ID node’scheckaction. Do not runtrainagain just to check progress.
Learn More about pryx-comfyui-higgsfield
- Higgsfield API model documentation (https://docs.higgsfield.ai/docs/models): Check a model’s documented capabilities and request options.
- Higgsfield video model documentation (https://docs.higgsfield.ai/docs/models/video-generation): Explore video tasks and model-specific behavior.
- ComfyUI project: Learn more about the node-based workspace used by this extension.
- pryx-comfyui-higgsfield issue tracker: Look for known extension issues or report a reproducible problem. Never include an API key in a report.
- Model update details: Read how catalog discovery, caching, and compatibility checks work.
- Security information: Review credential handling and request-safety details.
