Image Lab Now Auto-Falls-Back To Higgsfield Soul When Gemini Blocks

May 18, 2026

When you ask Image Lab for something Gemini's safety filter doesn't like — anything from a famous-franchise poster to a stylised tribute to a movie scene — the request used to fail with a generic moderation error. You'd then have to rewrite the prompt and try again, often without knowing what specifically tripped the filter. Now Alfrada quietly re-routes the same request to Higgsfield Soul in the background, tells you it did so, and ships you the image. No manual retry, no jargon, no waiting for the agent to ask permission first.

What you can do

• Auto-fallback on every Gemini block. Whether Gemini explicitly blocks (RECITATION for IP overlap, SAFETY for content rules, PROHIBITED_CONTENT for the strictest tier) or just silently returns nothing, Image Lab catches it and re-runs the prompt on Higgsfield Soul automatically — Soul Standard for fresh creates, Soul Reference for edits with a source image.
• Transparent provider switch. The response includes a fallback_triggered_by block naming the original Gemini reason. Alfred surfaces this in plain English: "Gemini blocked this on RECITATION (IP overlap), so I rendered it on Higgsfield Soul instead." You always know which model produced the file you're looking at.
• Pre-emptive prompt rewriting for known-blocking combos. If your prompt pairs a recognisable real person + a named film/franchise + photoreal framing — a combo both providers reject — Alfred now spots it ahead of time and rewrites to a stylised version (anime, illustration, kanji poster) before submitting. You skip the 30-second moderation roundtrip.
• Chained refinement on Higgsfield. Once Soul produces an image, you can ask Alfred to refine it ("darker rim lighting", "more rain on the street", "tighter crop") and Image Lab feeds the previous render back in as the source. Build a hero shot in three or four hops without ever touching Gemini.
• Honest double-rejection handling. If Higgsfield Soul also refuses, Alfrada stops trying and tells you both providers blocked it — there's no third provider, and the agent will propose a prompt rewrite instead of cycling forever.

Where this shows up

• You ask: "Design a poster inspired by Akira's iconic red motorcycle scene." Gemini blocks for RECITATION (Akira is a named work). Before, you'd see a vague error. Now Alfred says "Gemini blocked this for IP overlap with the Akira franchise — Higgsfield Soul rendered it instead" and shows the poster.
• You upload your team photo and ask: "Restyle this in the look of a 70s noir film." Gemini refuses on RECITATION. Image Lab runs it on Soul Reference using your photo as the visual anchor; the noir restyle lands in out/. The cleanup of the temporary upload runs automatically.
• You ask: "Make a photorealistic image of [public figure] as Kaneda from Akira riding the red motorcycle." Both providers reject the photoreal-real-person + named-IP combo. Alfred catches the combo in advance, proposes "a stylised anime poster featuring [their visual traits described] as the rider", waits for your okay, then renders it cleanly on the first try.
• You ask: "Generate a stylised Akira-style biker poster, then refine it with darker lighting and more rain." Two image_lab calls in sequence: the first creates the poster on Soul Standard; the second uses the first poster as the reference and refines on Soul Reference. Both renders end up in out/.

Try it

• "Make a stylised anime poster of a lone biker in neo-Tokyo, kanji, red bike, 9:16." — and if Gemini blocks, Alfred swaps to Higgsfield without missing a beat.
• "Restyle this portrait in the look of a famous animated film series." — works for the edit-mode fallback too.
• "Generate a Studio Ghibli-style village scene with a rainy mood, then refine it with warmer evening light." — chained refinement, both hops on Higgsfield.
• "I want a photorealistic image of [public figure] dressed as Spider-Man." — Alfred will explain the combo gets blocked and offer a stylised alternative before sending anything.
• "Which model rendered the last image?" — Alfred reads the response and tells you Gemini or Higgsfield Soul, plain and simple.

Heads up

• Provider switch is automatic, not silent. The agent always tells you when Higgsfield Soul rendered the image instead of Gemini. If a render lands without that note, Gemini did it normally.
• Soul Reference is single-image only. It cannot composite "face from photo A onto scene B". If your edit needs that kind of multi-source compositing and Gemini refuses, Alfred will explain that the fallback can't recover the composite — you'll get a stylised re-imagining option, or you can pre-composite the two images in another tool first.
• Real person + named franchise + photoreal = both providers block. This is the one combo where rewriting is unavoidable. Alfred will spot it before submission and pitch you the stylised version (anime, kanji poster, painterly portrait, comic panel). Approve the rewrite or describe what you want differently.
• Chained refinements drift slightly per hop. Three or four hops is the sweet spot. If the look starts wandering, ask Alfred to anchor back to an earlier frame instead of compounding the drift.
• Double-rejection means stop. If Higgsfield Soul also blocks, Alfrada won't keep trying providers. It'll tell you both refusals and propose a rewrite — that's the honest signal that the prompt itself needs to change.