PRYX ComfyUI Higgsfield Image Generate & Edit

Here's the honest pitch for this node: it's an API wrapper. Nothing runs on your GPU, your prompt and any source image leave the machine, and every render is metered. If that's already a dealbreaker, no amount of convenience fixes it - stay local.

What makes it worth reaching for anyway is the combination of two things local ComfyUI can't offer. First, models you cannot download at all: SOUL, Ideogram 4.0, Recraft V4.1, Grok Image 2.0, Marketing Studio, plus Qwen Image 3 and Z-Image Turbo for the open-weight side of the house. Second, a cost check before the request goes out - which is the part that makes a paid node tolerable in a graph you iterate on.

What it is

One node that fronts the catalog's 15 image-generation and image-editing endpoints. You pick the model from the dropdown; the node's widgets change to match what that model supports; you connect an IMAGE when you're using an edit model rather than a generate model.

The mechanism, briefly: your inputs are validated against the selected model's schema, local references are converted and pushed through Higgsfield's presigned upload flow if the provider needs public URLs, then the node calls the provider's POST /estimate/<endpoint> route. In estimate_only - the default - that's where it stops, and you read credits and usd. Switch request_mode to generate and the same path continues to the paid submit. The node talks to api.higgsfield.ai directly; PRYX doesn't proxy your requests or see your key.

The union-widget thing, which is the main gotcha

The node shows the union of every model's parameters, not the subset for your selection. So you'll see resolution choices 1k, 2k, 720p, 1080p, 4k when the model you picked only accepts two of them, and three differently-named prompt enhancers side by side: prompt_extend (which enable_thinking requires), prompt_extend_mode (direct, or agent for text-to-image only), and enhance_prompt. Vendor-specific fields that happen to mean similar things.

At request time, values that don't belong to the selected model are dropped and blanks are skipped. That's why "I set negative_prompt and nothing changed" is a common report - your model may have no such field. To see what it really accepts, expand the collapsed model-information panel.

The knobs you'll actually touch on a typical run: model, prompt (a connected input - wire a text or Config UI Prompt node into it), resolution and aspect_ratio (30 ratios, including auto), seed (-1 omits it entirely and lets the provider roll), batch_size (1 or 4 images per request), and for SOUL models style_id, which the pack fetches from Higgsfield's own style list and caches for 24 hours.

Vendor extras you'll meet as you switch models: image_weight (1–100, how hard the output hugs the input image), style_strength (0–1), rendering_speed (TURBO / DEFAULT / QUALITY), quality, moderation, colors, background_color, preset_id, and output_format.

Wiring

Connect a native Load Image to image for edit models, or your Reference Collector's references output for models that take ordered image references. Model Catalog's model_id can drive model if you want one place to swap models.

Outputs:

• images - native IMAGE. With batch_size 4 you get all four images as one batch. Empty in estimate_only.
• image_urls - JSON text of the provider's result URLs. For logging, not for wiring into other nodes.
• request_id - empty on an estimate-only run.
• credits and usd - provider estimates, -1 when the provider gives nothing. Estimates, not your balance.
• status - JSON diagnostics: status, elapsed time, URLs, reference mapping, provider status.

Install

Manager: search PRYX ComfyUI Higgsfield, install, restart. Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/PRYX-STUDIO/pryx-comfyui-higgsfield.git pryx-comfyui-higgsfield
python -m pip install -r pryx-comfyui-higgsfield/requirements.txt

Windows portable: ..\python_embeded\python.exe -m pip install -r .\pryx-comfyui-higgsfield\requirements.txt. Three dependencies (higgsfield-client, httpx, jsonschema), zero model files, ComfyUI 0.35.0+ and frontend 1.53.6+, Python 3.10–3.13.

Key: Settings → PRYX ComfyUI Higgsfield → Credentials, pasted as one key-id:key-secret value (it splits at the first colon). HF_KEY, or HF_API_KEY + HF_API_SECRET, for headless. It's stored locally in your ComfyUI user directory and never lands in the workflow.

Where people get burned

• Forgetting the mode. estimate_only gives you an empty images output and no error. That's the design, not a bug.
• Ignoring max_usd. 0 disables the limit; a positive value is a hard bound, and if the provider returns no USD estimate at all, a positive limit blocks generation instead of guessing. Set it.
• Trusting the estimate as a bill. It's the provider's best request-specific number. Account credits, model availability and content policy are all still provider-side, and a real generation can fail after a clean estimate.
• Assuming the consumer app's reputation is the API's. Higgsfield's subscription side has a well-documented billing-complaint trail, so be wary of the brand - but this pack only ever spends what a generate-mode request costs, and never buys a subscription at all.