How to Combine Midjourney, Nano Banana Pro, and Seedance 2: A Three-Stage Canvas Workflow for Stylized AI Video
Martini is the only multi-model canvas that ships Midjourney v7 as a first-class node — through an exclusive Youchuan API integration, not a risky Discord-scraping wrapper. That makes Midjourney → Nano Banana Pro → Seedance 2 a three-stage stylized video pipeline that simply does not exist anywhere else.
Key takeaways
- Each model owns one stage of the pipeline: Midjourney provides the aesthetic origin (signature look, mood, brand vibe), Nano Banana Pro converts that aesthetic into a consistent multi-image library with up to 14 references and identity locked across up to five characters, and Seedance 2 turns the chosen stills into cinematic motion. None of the three alone delivers a finished stylized video brief.
- This combination is exclusive to the Martini canvas because Martini is the only multi-model surface that ships Midjourney v7 as a first-class node. The integration runs through an exclusive Youchuan API path — not the Selenium-based Discord scrapers that risk Midjourney account bans — so you drop a Midjourney node, switch between V7 and Niji 7, set the Stylization / Variety / Weirdness sliders, pick from the 2x2 grid of 4 images Midjourney returns per call, and pipe the chosen frame straight into Nano Banana Pro downstream. No Discord, no PNG download, no upload step, no ban risk.
- Stage 2 is the part most readers underestimate. Sending a Midjourney image straight into Seedance 2 works, but you lose almost everything Midjourney is good for the moment the character has to recur or the scene needs to change. Running the Midjourney still through Nano Banana Pro first — passing it as one of up to 14 references — preserves the look and unlocks a multi-shot character library that holds identity across the rest of the project.
- For Seedance 2, the right variant depends on what stage 2 gave you. If Nano Banana Pro produced a multi-shot library where the same character recurs, drop the stills into Seedance 2 Omni nodes — Omni is the variant that genuinely respects image-side identity. If the brief is a single hero take with no recurrence, Seedance 2 Pro is the cleaner pick.
- The version tray on the canvas remembers every Midjourney 2x2 grid, every Nano Banana Pro variant, and every Seedance 2 take. Re-rendering one motion take does not force you to re-run any earlier stage — the references stay wired, the seeds stay pinned, the iteration history attaches to the canvas rather than scattering across three vendor dashboards. That persistence is what turns the three-stage pipeline from a thought experiment into a Tuesday-afternoon production loop.
Why combine three image and video models in the first place
Stylized AI video is a brief that breaks one-model pipelines. The brief has three independent requirements: an aesthetic that does not look like every other AI output, a character (or product, or world) that stays recognizable across shots, and motion that reads as cinematic rather than morph-y. No single model in 2026 delivers all three reliably. Midjourney owns aesthetic but not character recurrence and has no native motion. Nano Banana Pro owns character recurrence and reference fidelity but is not the strongest pure-aesthetic generator on the market. Seedance 2 owns motion but needs a strong reference still and is fairly faithful — it amplifies whatever look you give it, but it does not invent that look from scratch.
The argument for combining all three is almost arithmetic. Midjourney gives you the look. Nano Banana Pro takes that look and gives you a usable library — multi-image reference, identity-locked, ready to be referenced ten more times. Seedance 2 takes a single still from that library and gives you a usable cinematic take. The output of each stage is exactly what the next stage needs as input. That is what a pipeline is.
The reason teams default to one-model workflows is not that the multi-model workflow is bad — it is that the multi-model workflow has historically been a chore. Three browser tabs, three iteration histories, manual PNG shuttling between Discord and downstream tools, no shared color grade, no shared aspect ratio, and every change upstream forces a full re-run downstream. The Martini canvas removes those frictions, and because Midjourney v7 sits on it as a native node through an exclusive API integration, the pipeline collapses into something a producer can run on a Wednesday afternoon.
Why this combination is exclusive to the Martini canvas
Midjourney does not ship a public production API for the general market. Discord and the midjourney.com web interface are the official surfaces, and the third-party "Midjourney API" projects you can find on GitHub are almost all Selenium-based scrapers that automate Discord — a path Midjourney explicitly disallows in their terms of service and that gets accounts banned in waves. That is why every other multi-model canvas either omits Midjourney entirely or routes you to an external tab and asks you to bring back a PNG. The combination this guide describes simply cannot run end-to-end on those tools.
Martini is the exception because Martini ships an exclusive Midjourney v7 integration through the Youchuan API. The Midjourney node sits on the canvas like any other image generator — Nano Banana 2, Flux, GPT Image 2, Imagen 4. You pick V7 for photoreal and painterly work or Niji 7 for anime, turn the Stylization slider (0–1000) to shape how much aesthetic license the model takes, set Variety (0–100) to spread the 4-image grid further apart, and set Weirdness (0–3000) when you want outputs that drift away from the Midjourney average. Every call returns a 2x2 grid of 4 images, all of which live in the version tray and any one of which can be wired into a downstream node directly. No Discord login, no PNG export, no ban risk from a scraping wrapper.
The downstream half of the pipeline — Nano Banana Pro and Seedance 2 — is also unusual. Nano Banana Pro is Google's Gemini 3 Pro Image variant; Seedance 2 is ByteDance's second-generation video model. The number of canvas tools that ship sanctioned integrations to OpenAI, Google, ByteDance, and Midjourney in the same workspace is essentially one. That is the structural reason this guide exists here and almost nowhere else: every other surface is missing at least one stage of the pipeline, and the missing stage is usually Midjourney.
Stage 1 — Midjourney for the aesthetic origin
In the three-stage pipeline, Midjourney owns one job: produce a frame (or a small set of frames) that establishes the look the rest of the project will inherit. Treat this stage like a director of photography selecting a stock and a color grade — the goal is not a finished frame the audience will see, it is a reference frame the downstream models will read. Midjourney's strength is exactly this: a frame that feels stylistically distinct, with a sense of light, palette, and texture that does not read as default-AI.
Prompt Midjourney for the mood and the aesthetic, not for the specific shot. "Coastal cafe at golden hour, slatted wood interior, salt-bleached palette, soft window haze, painterly atmosphere, 35mm color stock look" produces a usable origin frame; "the spokesperson holding the product at the cafe counter while smiling" tries to make Midjourney do too much and the result is harder to reference downstream. Aesthetic in Midjourney, specifics in Nano Banana Pro.
Drop a Midjourney node on the canvas. Pick V7 for photoreal or painterly aesthetics; pick Niji 7 if the project leans anime, manga, or clean cel-shaded illustration. Tune the sliders before the first generation: Stylization at 50–150 for product photography and realistic portraits, 250–500 for editorial and concept art, 750–1000 for high-style poster work. Leave Variety at 20–40 if you want a more diverse 4-image grid without losing prompt adherence; push Weirdness to 200–500 for "interesting but still on-brief" exploration. Run the prompt. Midjourney returns a 2x2 grid of 4 images — pin the strongest as the canonical origin frame. The pinned frame is now an upstream node like any other image generator, and every downstream node on the canvas can reference it without re-fetching, re-prompting, or re-running Midjourney. For projects with an existing brand aesthetic, you can also reuse a frame from a previous Midjourney run on the same canvas, or wire Midjourney's Omni Reference (V7's identity-lock system) to a brand-approved character or product photo for an origin frame that already carries identity from the first generation.
Stage 2 — Nano Banana Pro as the consistency engine
Stage 2 is where the pipeline earns most of its value. Nano Banana Pro is Google's Gemini 3 Pro Image variant, sitting under the Nano Banana 2 model node on Martini. It accepts up to 14 reference images per generation, maintains character identity for up to five people in a single composition, and supports opt-in web search grounding for rendering real places, products, and public figures. For a pipeline whose job is to convert a stylistic origin frame into a working library, this combination of features is the right tool.
Drop a Nano Banana 2 image node on the canvas and switch the variant to Nano Banana Pro. Wire the pinned Midjourney frame in as a reference — the connection is the same as wiring any other upstream image node on the canvas. If the project has a specific character, also wire in two or three character-reference images (front, three-quarter, expression study) — Nano Banana Pro can read them all together and will weight the Midjourney frame for aesthetic while weighting the character frames for identity. Prompt for the first scene the project actually needs: "the character standing at the cafe counter in the referenced style, soft window haze, holding a small ceramic cup, three-quarter view from camera right." The output is a frame that inherits Midjourney's look and Nano Banana Pro's identity discipline.
Repeat the Nano Banana Pro stage for each scene the project needs. Same Midjourney aesthetic reference, same character references, different scene prompts. The model holds the look and the identity across the library because the references are the same on every generation. For projects that need multiple characters in the same frame — a two-shot dialogue, a group composition — Nano Banana Pro's five-person identity lock is the variant that actually delivers, where most consistency models fall over at two.
When you need surgical edits to a Nano Banana Pro output (clothing change, prop swap, background fix) without disturbing the face, chain a Flux Kontext node downstream of the chosen take. This is the same pairing recommended in the Nano Banana 2 workflows handbook and it applies identically to the Pro variant. Kontext fan-out gives you multiple outfit or environmental variants from one Nano Banana Pro base, all of them inheriting the Midjourney aesthetic.
Stage 3 — Seedance 2 for motion
Stage 3 takes one or more of the stills from stage 2 and produces motion. The pick of Seedance 2 variant depends on what stage 2 produced and what the brief asks for. Seedance 2 Pro is the right node when the brief is a single hero take that will be color-graded downstream and live on a finished screen — the Pro variant tolerates longer prompts, gives the most reliable response to camera-direction language, and respects the stylistic grade of the input still. Seedance 2 Lite is the right node for iteration loops and motion tests where you are still settling the prompt; the visual gap to Pro narrows considerably on stylized stills with a strong reference. Seedance 2 Omni is the right node when the still depicts a character who must recur across multiple motion takes — Omni reads tagged references and preserves character identity across many parallel renders, which is exactly the pattern stage 2 produced.
Wire the chosen Nano Banana Pro still into a Seedance 2 node and write a tight one-shot motion prompt. The prompt grammar that works is unchanged from the Seedance 2 handbook: subject + action + camera + lens + lighting + atmosphere. The difference with a stylized stage-2 still is that you can usually drop most of the visual description from the prompt — the still carries it — and lean entirely on motion direction. "Subject begins still at the counter, lifts the cup, slow head-turn to camera left, slight smile, four-second take, locked focal length, anamorphic 35mm look" is a workable prompt because every other visual decision is already encoded in the upstream still.
For multi-shot sequences, duplicate the Seedance 2 node and re-wire each duplicate to a different Nano Banana Pro still from the library. Vary only the motion prompt across duplicates — different camera moves, different micro-actions, different beats. Render in parallel. The version tray holds every take. Drop an NLE export node downstream of all of them and the canvas assembles the takes in the order you wire them. A two-minute stylized character piece runs through this pattern with three to five Seedance 2 nodes, all rendering from different stills in the same Nano Banana Pro library, all of which trace back to the same Midjourney origin frame.
Workflow 1 — stylized character short
A typical end-to-end pipeline for a sixty-second stylized character short looks like this. Drop a Midjourney node on the canvas, switch to V7, set Stylization to 300 (editorial), prompt for the origin frame — for example, "painterly coastal cafe interior at golden hour, slatted wood, salt-bleached palette, soft window haze, 35mm color stock look" — and pin the strongest take from the 2x2 grid. Drop a Nano Banana Pro node next to it for the character library: prompt for the character description (no aesthetic, just the person), and produce four reference views (front, three-quarter, profile, expression). Pin all four.
Drop the working Nano Banana Pro nodes for each scene in the short. Wire the Midjourney origin frame plus the four character references into every node — that is up to five references per generation, well under the 14-reference cap. Prompt each node for the scene-specific composition: "character entering the cafe, soft window haze, painterly grade from the referenced frame," then "character at the counter, ceramic cup in hand, three-quarter view," then "character at the window seat looking outside, side light." Pin the strongest take from each.
Stage 3: for each pinned still, drop a Seedance 2 node. Use Seedance 2 Omni for the takes featuring the character (the canvas's recurrence-aware variant), and Seedance 2 Pro for any environmental take that does not depend on character identity. Write a single-shot motion prompt for each. Render. Wire the chosen takes into the NLE export node downstream. The final piece reads as one continuous aesthetic from frame one to frame eighteen-hundred — because every frame traces back to the same Midjourney origin reference, every character beat traces back to the same Nano Banana Pro library, and every motion beat shares the canvas color grade.
Workflow 2 — brand-aesthetic product video
For commerce work, the same three-stage pattern adapts to a product brief. Stage 1 in Midjourney produces a frame that establishes the brand-side aesthetic: lighting style, palette, surface treatment, the way the brand likes glass to refract or fabric to fold. This is one of Midjourney's strongest jobs because product brands often have a distinct visual language that default-AI tools fail to match. Drop a Midjourney v7 node on the canvas, lower Stylization to 100–200 to keep the look commercial rather than editorial, optionally wire the brand-approved product photo into Omni Reference for label identity, prompt the brand aesthetic, and pin the strongest frame from the 2x2 grid.
Stage 2 in Nano Banana Pro converts the aesthetic into a usable product library. Wire the Midjourney aesthetic frame in as a reference. Wire a clean product reference photo (the actual SKU on white, or the brand-approved hero shot) in as a second reference. Prompt for the placements the campaign needs: "the referenced product on a marble counter in the referenced aesthetic, soft window light, label centered." Generate three or four variants per placement and pin the strongest. Nano Banana Pro's strength here is that the product stays label-accurate (because the SKU reference dominates the product slot) while the surrounding environment inherits the Midjourney look (because the aesthetic reference dominates the environment slot).
Stage 3 in Seedance 2 produces the motion. Seedance 2 Pro is the default pick for product hero shots — the orbit, the push-in, the rack-focus reveal. Wire the chosen Nano Banana Pro still into a Seedance 2 Pro node and prompt for the camera move only. For variant sweeps (different angles, different motion intensities) duplicate the node and swap the prompt. The ad ends up with three or four motion variants per placement, each anchored in the same Nano Banana Pro still, each inheriting the same Midjourney aesthetic. Drop the chosen takes into the NLE export node and the ad is finished.
When to skip a stage
Not every project needs all three stages. The honest version of this guide acknowledges which combinations are overkill. If the brief does not call for Midjourney's specific aesthetic — the character is generic, the look is default-cinematic — skip Midjourney entirely. Drop the character library straight into Nano Banana Pro from scratch and run the rest of the pipeline. Stage 2 plus stage 3 is the canonical workflow for most character work and is documented in the Nano Banana 2 workflows handbook.
If the brief is a single hero take with no recurrence and no character — a product orbit, an environmental beat, a one-off social post — skip Nano Banana Pro. The Midjourney frame goes straight into a Seedance 2 node as a reference still. This is the right shape for projects that need Midjourney's look on one take and do not need the library or the recurrence economics. Two stages, two nodes, one canvas.
If the brief is the opposite — character recurrence matters but the aesthetic does not need to feel Midjourney-specific — skip Midjourney and start with Nano Banana Pro. Generate the character library with no aesthetic origin reference, then run stage 3 as documented. This is the standard pattern for AI-influencer content where the channel's look has already been set by the character itself rather than by an external aesthetic.
Run all three stages when the brief asks for all three things: a distinctive look, a character or product that recurs, and cinematic motion. That is the brief this guide is written for, and it is the brief that single-tool workflows handle worst.
The bottom line
The three-stage Midjourney → Nano Banana Pro → Seedance 2 pipeline is not a clever workaround for any tooling gap. It is the right division of labor for stylized AI video, and the Martini canvas — with its exclusive Midjourney v7 integration alongside sanctioned Google and ByteDance access — is the surface that makes it economical. Midjourney sets the look. Nano Banana Pro turns the look into a library that holds identity. Seedance 2 turns the library into motion. Each stage does what it is best at; none of the three is asked to do something it is wrong for.
The work that goes into setting up this pipeline once — finding the right Midjourney origin frame, generating the Nano Banana Pro library, picking the right Seedance 2 variants — pays back across every subsequent shot in the project. The version tray remembers the references, the prompt skeletons, and the takes; re-rendering one motion take does not unwind any of the earlier work. That persistence is the actual production unlock, more so than any individual model's strength.
For teams whose existing pipeline already includes Midjourney, this is the lowest-friction way to bring it into a real production workflow without giving up on Midjourney's aesthetic. For teams who do not yet use Midjourney, the same canvas pipeline runs without stage 1 — the architecture is the value, not any single model in it.
Workflow example
Sixty-second stylized character short on Martini using Midjourney, Nano Banana Pro, and Seedance 2: drop a Midjourney v7 node, set Stylization to 300, prompt for "painterly coastal cafe interior at golden hour, salt-bleached palette, soft window haze, 35mm color stock look," and pin the strongest frame from the 2x2 grid as the aesthetic origin. Drop a Nano Banana Pro node and generate a four-image character library — front, three-quarter, profile, expression — and pin all four. Drop three working Nano Banana Pro nodes, wire the pinned Midjourney frame plus the four character references into each, and prompt each for a different beat: character entering the cafe, character at the counter with a ceramic cup, character at the window seat. Pin the strongest take from each. Drop three Seedance 2 Omni nodes (one per still) and write a one-shot motion prompt for each — push-in on the entrance, slow head-turn at the counter, parallax drift across the window. Wire all three takes plus a logo end-card into the NLE export node downstream. Every stage runs on the same canvas through sanctioned API integrations: Youchuan for Midjourney, Google for Nano Banana Pro, ByteDance for Seedance 2. No tab-switching, no PNG download, no ban-risk scraper in the loop.
Recommended models
Recommended features
Related models and tools
Related how-to guides
Related comparisons
Related reading
Nano Banana 2 Workflows for Multi-Image Reference and Character Consistency
Multi-image reference and character consistency workflows on Martini using Nano Banana 2.
Seedance 2 Handbook: Variants, Best Workflows, and How to Use It on Martini
Hands-on guide to Seedance 2 — variants, strengths, and the production workflows it fits on Martini's canvas.
Awesome Seedance 2 Prompts: Curated Open-Source Prompt Library
A curated, open-source library of Seedance 2 prompts — copy-paste recipes for product spins, image-to-video motion, character shots, and cinematic camera moves. Maintained on GitHub, paired with the Martini canvas.
Frequently asked questions
- How does Martini run Midjourney natively — isn't Midjourney Discord-only?
- Midjourney does not ship a public production API for general access, which is why other multi-model canvases either omit it or ask you to bring back a PNG from Discord. Martini ships Midjourney v7 (and Niji 7) as a first-class canvas node through an exclusive Youchuan API integration — a sanctioned API path rather than the Selenium-based Discord scrapers that risk account bans. You drop a Midjourney node like any other image generator on the canvas, configure V7 / Niji 7, set the Stylization / Variety / Weirdness sliders, prompt, and the node returns a 2x2 grid of 4 images directly into the version tray. From there it wires into Nano Banana Pro, Flux Kontext, or Seedance 2 downstream just like any other upstream image node.
- What does Omni Reference on Midjourney add when I am already running Nano Banana Pro downstream?
- Omni Reference is Midjourney v7's built-in identity-lock system — it accepts one reference image (character, product, vehicle, or logo) and carries that identity into the generated frame while preserving Midjourney's aesthetic. It is V7-only, not Niji 7. The complementarity with Nano Banana Pro is real: use Omni Reference on the Midjourney node when you want the origin frame itself to already carry a brand-approved identity (a product label, a spokesperson likeness), then let Nano Banana Pro handle the harder multi-character or multi-reference scenes downstream where its 14-reference + five-person identity-lock starts to matter. For single-character single-frame work, Omni Reference alone can sometimes carry the project without needing the Nano Banana Pro stage at all.
- Why send the Midjourney image through Nano Banana Pro instead of straight into Seedance 2?
- Because Seedance 2 will faithfully animate whatever still it gets, but it will not give you a multi-shot library off one still. If the project is a single take with no recurrence, Midjourney directly into Seedance 2 is fine and you can skip stage 2. The moment the character has to recur across shots or the scene has to change while keeping the aesthetic, you need Nano Banana Pro in the middle — it reads the Midjourney frame as an aesthetic reference and the character images as identity references, and produces a library that holds both. Without that stage, every Seedance 2 take would need its own one-off Midjourney prompt, and the aesthetic would drift.
- What does Nano Banana Pro add that the base Nano Banana 2 does not?
- Three things matter for this pipeline. First, Nano Banana Pro accepts up to 14 reference images per generation (vs the base model's 10) — useful when you want to pass a Midjourney aesthetic plus several character views plus a product reference in the same prompt. Second, Pro maintains character identity for up to five people in one composition, which the base model does not do reliably. Third, Pro supports opt-in web search grounding when you need real places or products to render correctly. For pure single-character work without product specifics, the base Nano Banana 2 may be enough; for the full three-stage pipeline as written, Pro is the variant that earns the cost.
- Which Seedance 2 variant works best with Nano Banana Pro stills?
- Seedance 2 Omni when the character has to recur across multiple motion takes — Omni is the variant that genuinely respects image-side identity and preserves it across parallel renders. Seedance 2 Pro for hero environmental shots and any take that will be color-graded and live on a finished screen. Seedance 2 Lite for prototyping the motion prompt before committing to Pro renders. A typical multi-shot sequence will use all three variants on the same canvas, all reading from different stills in the same Nano Banana Pro library.
- Doesn't running three models cost three times as much as picking one?
- Not really, once you account for the full iteration loop. The Midjourney cost is modest and the per-call economics are good — every Midjourney generation returns a 2x2 grid of 4 images, so one call gives you four candidate origin frames at once. The Nano Banana Pro cost scales with the size of the library you actually need — typically three to six generations for a short. The Seedance 2 cost dominates the total. Running this pipeline costs the same Seedance 2 spend as running Seedance 2 alone, plus a modest stage-1 and stage-2 overhead that pays back the moment any shot needs to be re-rendered with a tweaked motion. On a single-model workflow, every Seedance 2 re-render forces a manual re-pipeline; on the canvas, it does not.
- Can I skip Midjourney if I do not need its specific aesthetic?
- Yes — that is exactly the pattern documented in the existing Nano Banana 2 workflows guide. Drop the character library straight into Nano Banana Pro from scratch, then run Seedance 2 downstream. Two stages instead of three, same architecture. Use the full three-stage pipeline when the brief explicitly asks for Midjourney's look; use the two-stage variant when the aesthetic can come from Nano Banana Pro alone or from a different reference image source. The canvas does not care which.
- How does this differ from combining Sora 2 and Kling on the canvas?
- The Sora 2 plus Kling workflow is about combining two video models for different beats of one sequence — Sora 2 for the environmental world, Kling for the character motion, both rendering from the same image. The Midjourney plus Nano Banana Pro plus Seedance 2 workflow is a vertical pipeline — Midjourney for the look, Nano Banana Pro for the library, Seedance 2 for the motion, each stage feeding the next. Many production projects use both shapes on the same canvas: a vertical pipeline to produce the stills, then a horizontal fan-out across multiple video models for the motion takes.
Ready to try it on the canvas?
Open Martini and fan your prompt across every frontier model in one workflow.