Nano Banana Pro vs Soul Cast in Higgsfield: which image model produces better AI film characters?

Nano Banana Pro produces better AI film characters than Soul Cast. In a documented head-to-head test funded with thousands of credits, Soul Cast — Higgsfield Supercomputer's default character model — generated character sheets containing different characters than the reference images they were built from, while Nano Banana Pro delivered script-accurate characters. One caveat: Supercomputer won't retain your Nano Banana Pro selection between steps.

For narrative film characters, pick Nano Banana Pro over Soul Cast every time: in direct testing, Soul Cast produced inferior character and location generation for film use, while Nano Banana Pro held reference fidelity and script accuracy across a full production. A quick terminology note, since the wider community discusses "Soul 2.0" or "Soul ID": Soul Cast is the character-generation variant Higgsfield Supercomputer defaults to for fictional characters — that default is the model this comparison is about.

Where Nano Banana Pro wins. The decisive factor for film characters is whether the model preserves the character you gave it and the details your script demands. In the documented test, Nano Banana Pro character work stayed faithful to references, and when paired with strong script context it produced character sheets that included prop and wardrobe variations for later scenes — not just the initial look. It also handles image compositing well, so you can generate multiple variants of the same character holding different objects from one approved base image.

Where Soul Cast fails for film. Character sheets generated with Soul Cast contained different characters than the original reference images they were supposed to be built from — a disqualifying error when your film needs the same face in shot 1 and shot 40. Compounding this, Supercomputer locks character references and moves on before you can review or approve them, so a wrong Soul Cast face can propagate into video generation before you catch it.

The workflow caveat even when you choose correctly. Supercomputer does not retain model preferences between steps — if you select Nano Banana Pro for character creation, you must explicitly re-tell it at every subsequent stage or it reverts to defaults. It also rushes through setup before character creation is complete, so budget for redirecting it repeatedly.

Soul's honest strength. Some users who tested 10+ image models report Soul 2.0 stood out for its stylized aesthetic realism in standalone portraits, so it has a legitimate use case outside narrative work — though threads in the same community now describe the Soul model line as fading. For a film character that must survive dozens of shots, reference fidelity beats vibe, and that is Nano Banana Pro's territory.

Getting more out of Nano Banana Pro. The model is only half the result — the review loop around it is the other half. invideo is an agentic video creation tool with all the current image and video models available, including Nano Banana Pro, Recraft, and GPT-Image-2. In the same head-to-head test, the invideo agent built a context library from the script before generating, then analyzed every Nano Banana Pro output and automatically re-ran renders with prompt adjustments when something looked off — the tester reported zero character or location inconsistency issues and fewer generations and credits spent than the Supercomputer run of the same scene. If Nano Banana Pro is your answer for characters (it should be), running it through the invideo agent gives you retained model choice, reviewable character sheets, and automatic quality control instead of a re-select-every-step workflow.

Watch some of these to see what works for you:

It analyzed all of the generations provided by Nano Banana Pro, and if something looked off or rendered wrong, it would automatically rerun the render with a prompt adjustment. This is much more of like a real teammate.

— a filmmaker who spent thousands of credits testing both tools