Nano Banana — Google’s Gemini Flash 2.5 — is already reshaping how creators approach image modification. In this episode of Denoised, hosts Addy and Joey break down the most practical use cases, prompt best practices, a Photoshop plugin that brings Nano Banana into an existing workflow, an impressive ComfyUI VFX tutorial that blends live action and AI without green screens, and the early ripple effects of AI in the music industry.
Quick takeaways
Nano Banana (Gemini Flash 2.5) excels at targeted image edits: changing props, camera angles, relighting, and maintaining likeness.
Annotative tools (like Freepik’s visual prompts) and semantic positive prompts improve control and precision.
Comfy UI workflows are becoming viable VFX pipelines for hybrid filmmakers—mask the actor, generate backgrounds, blend edges, and optionally feed camera-tracking data.
AI music is entering mainstream channels—from viral Suno tracks to fully synthetic Spotify bands and AI K‑pop idols—raising commercial and creative questions for musicians and studios.
Why Nano Banana matters to filmmakers and creators
Joey and Addy positioned Nano Banana not as an abstract novelty but as a practical tool that fills gaps in current image workflows. For many creators, the model’s strength is not in producing novel standalone art but in modifying existing assets with high fidelity—replicating likenesses, swapping props in hands, changing camera perspective, and touching only specific parts of a scene while leaving the rest untouched.
Desk test: a real-world example
To prove the point, a desk test swapped modern laptops for vintage clamshell iBook computers and requested aerial camera angle changes. The model accurately reconstructed occluded objects (like a mug barely visible behind a laptop), matched the room’s wood tone, and even preserved table grain in multiple panels. Those small, contextual details are what make edits feel believable in a film or commercial pipeline.
Practical use cases editors and VFX teams should know
Image-likeness replication: quick generation of digital doubles or consistent headshots for previs and casting tests.
Prop and product integration: drop a product into a subject’s hands with realistic reflections and shadows.
Changing camera angles and perspective: generate believable aerial or isometric shots from single photos.
Image restoration and archival work: restoring damaged black-and-white photos to reveal previously obscured detail.
Annotation and visual prompts
Annotation overlays—where a user marks precisely what to change—are particularly effective. Freepik released a visual prompts tool that makes annotated edits easier (note: it doesn't accept additional images yet). For anyone experimenting, Freepik and similar platforms are useful because they let you quickly swap between models for testing.
Prompting and workflow tips from experts
DeepMind's Philipp Schmid shared a short list of best practices that the hosts agreed were useful for everyday workflows:
Be hyper-specific. The more detail the prompt contains, the more control the model provides.
Provide context and intent. Describe the purpose of the image. For example, "create a logo for a high‑end minimalist skincare brand" is better than "create a logo."
Use semantic positive prompts. Rather than saying "no cars," describe what you want (e.g., "a deserted street with no signs of traffic").
Manage aspect ratios. When uploading multiple reference images, the model preserves the aspect ratio of the last image uploaded. Upload the image with the target aspect ratio last to avoid extra manual cropping.
Nano Banana vs. Flux Kontext and skin/detail work
Nano Banana’s improvements in skin detail—pores, sub-surface scattering, small wrinkles—make it noticeably closer to crossing the uncanny valley compared to previous models. Where Flux Kontext can sometimes alter unrelated elements, Nano Banana tends to confine edits to the requested region, which is critical for preserving actor continuity in shots.
Bringing Nano Banana into Photoshop
For editors who live in Photoshop, a community developer’s plugin from Rob de Winter exposes Nano Banana and Flux Kontext within Photoshop’s interface. That “last mile” integration lets artists work with layers, masks, and adjustment tools they already trust while calling modern generative models under the hood—saving round trips to web apps and simplifying client workflows.
Adobe, integrations, and the future of model choice
The hosts debated Adobe’s evolving approach: rather than being a single-model vendor, Adobe is moving toward a model-agnostic platform that can call third-party models. That offers a choice between “commercially safe” models and higher-quality but potentially more legally ambiguous models. For production teams, that means choosing the right model for the job and, when necessary, running jobs through customer-owned API keys to track usage and costs per client.
ComfyUI VFX: Mickmumpitz’s live-action workflow
A standout example in the episode was a deep-dive into a Comfy UI workflow by Mickmumpitz that blends live-action footage and AI-generated backgrounds without green screens.
How the workflow works — the essentials
Film the actor in a practical environment. The actor must be masked out (a rough rotoscope is fine).
Provide a depth map (black/white mat) to indicate scene depth and occlusion.
Optionally import camera-tracking data as a simple text/JSON coordinate file or use Comfy’s tracking nodes.
Generate new background elements (text prompt and/or reference images) and let the model render the scene frame-by-frame.
Blend edges and refinements so the actor remains untouched while the environment changes.
That workflow makes an entire short — think Honey, I Shrunk the Kids / scale-change scenes — feasible on a single workstation using open-source models. The biggest production benefits are speed and flexibility: actors can be filmed in place and the environment can be altered later without elaborate pre-rigging or green-screen setups.
Camera tracking and data integration
Mick’s workflow accepts simple camera tracking exports (text/JSON). Comfy has nodes for point tracking so the camera motion can map to generated geometry. While traditional match-moving still offers the most robust result, this approach removes a large amount of pipeline friction—especially for indie filmmakers and small VFX teams.
What this means for filmmakers
Hybrid pipelines—shoot practical actors, mask them, and AI-generate environment elements—are now realistic for shorts, previs, and lower‑budget VFX sequences.
Local execution of open-source models (Comfy, etc.) empowers teams to iterate fast without recurring cloud costs.
For professional productions, integrating AI steps into established tools (Photoshop, Blender, After Effects) is where most time savings will occur.
AI in music: early wins and wide questions
The episode closes by surveying recent AI music headlines and their implications for creators and rights holders.
Notable stories discussed
Record deal for an AI-era creator: A British creator with no formal musical background used Suno to produce tracks that gained millions of streams and led to a record deal. The label framed the signing around a human-in-the-loop creative process—writing lyrics, iterating multiple versions, and adding a personal brand around the AI output.
Fully synthetic bands on streaming platforms: Indie projects that generate audio, imagery, and backstory purely with AI have accumulated high stream counts and prompted questions about transparency and royalties.
AI K‑pop idols: Companies like Higgsfield have launched AI idols—complete with generated music videos and avatars—spotlighting a commercial model where margins look different because there’s no live performer to pay.
Key business and creative implications
Labels and platforms are experimenting with AI-driven content; novelty can create short-term attention but doesn’t guarantee long-term fandom.
Copyright frameworks (human-authorship requirements) still matter—human editing, lyric-writing, or artistic curation helps secure rights in many jurisdictions.
For filmmakers, the music landscape suggests a bifurcation: AI-generated background or functional music (sleep, lobbies, playlists) versus human-authored works that rely on personal brand and cultural context.
Final thoughts for creators and studios
Across image generation, VFX, and music, the most consistent theme is practical augmentation: AI does the heavy lifting on repetitive or detail-heavy tasks, while human creators provide intent, taste, and narrative context. For filmmakers and producers, the immediate opportunities are clear—faster previs, more flexible VFX, and cheaper iteration—and the longer-term responsibilities are equally clear—defend authenticity, document human contributions for rights, and adapt pipelines to integrate these new tools.
As Joey and Addy put it, success will favor creators who combine technical fluency with a strong personal brand: the human connection will remain the currency people pay for, whether it’s a film, a live show, or a song.
Further reading and resources
Filmmakers and VFX artists should explore Freepik's visual prompts and the Photoshop Nano Banana plugin as immediate workflow enhancers.
Investigate ComfyUI tutorials (Mickmumpitz) for hands-on VFX workflows that blend live action and AI generation.
Follow industry conversations about AI and copyright to understand human-authorship thresholds for commercial use.
